Deloitte的六大技術(shù)趨勢(shì)

在這個(gè)技術(shù)變革加速的時(shí)代，人工智能（AI）正以前所未有的速度改變企業(yè)的核心運(yùn)營(yíng)模式。此份報(bào)告圍繞空間計(jì)算、AI未來(lái)趨勢(shì)、智能硬件、IT升級(jí)、量子計(jì)算、智能核心六大主題展開(kāi)深入探討，無(wú)論是企業(yè)決策者還是技術(shù)管理者，都可以從中獲取戰(zhàn)略性洞察，為未來(lái)的技術(shù)升級(jí)和數(shù)字化轉(zhuǎn)型做好準(zhǔn)備。

Spatial Computing Takes Center Stage

What is the future of spatial computing?

With real-time simulations as just the start, new, exciting use cases can reshape industries ranging from health care to entertainment.

Kelly Raskovich, Bill Briggs, Mike Bechtel, and Ed Burns

Today’s ways of working demand deep expertise in narrow skill sets. Being informed about projects often requires significant specialized training and understanding of context, which can burden workers and keep information siloed.

This has historically been true especially for any workflow involving a physical component. Specialized tasks demanded narrow training in a variety of unique systems, which made it hard to work across disciplines.

One example is computer-aided design (CAD) software. An experienced designer or engineer can view a CAD file and glean much information about the project.

But those outside of the design and engineering realm—whether they’re in marketing, finance, supply chain, project management, or any other role that needs to be up to speed on the details of the work—will likely struggle to understand the file, which keeps essential technical details buried.

Spatial computing is one approach that can aid this type of collaboration. As discussed in Tech Trends 2024, spatial computing offers new ways to contextualize business data, engage customers and workers, and interact with digital systems.

It more seamlessly blends the physical and digital, creating an immersive technology ecosystem for humans to more naturally interact with the world.

For example, a visual interaction layer that pulls together contextual data from business software can allow supply chain workers to identify parts that need to be ordered and enable marketers to grasp a product’s overall aesthetics to help them build campaigns.

Employees across the organization can make meaning of and, in turn, make decisions with detailed information about a project in ways anyone can understand.

If eye-catching virtual reality (VR) headsets are the first thing that come to mind when you think about spatial computing, you’re not alone.

But spatial computing is about more than providing a visual experience via a pair of goggles.

It also involves blending standard business sensor data with the Internet of Things, drone, light detection and ranging (LIDAR), image, video, and other three-dimensional data types to create digital representations of business operations that mirror the real world.

These models can be rendered across a range of interaction media, whether a traditional two-dimensional screen, lightweight augmented reality glasses, or full-on immersive VR environments.

Spatial computing senses real-world, physical components; uses bridging technology to connect physical and digital inputs; and overlays digital outputs onto a blended interface (figure 1).

Spatial computing’s current applications are as diverse as they are transformative.

Real-time simulations have emerged as the technology’s primary use case.

Looking ahead, advancements will continue to drive new and exciting use cases, reshaping industries such as health care, manufacturing, logistics, and entertainment—which is why the market is projected to grow at a rate of 18.2% between 2022 and 2033.

The journey from the present to the future of human-computer interaction promises to fundamentally alter how we perceive and interact with the digital and physical worlds.

一、空間計(jì)算成為焦點(diǎn)

空間計(jì)算未來(lái)會(huì)是什么樣子？ 從實(shí)時(shí)模擬的簡(jiǎn)單應(yīng)用開(kāi)始，這項(xiàng)技術(shù)正逐漸改變從醫(yī)療到娛樂(lè)等多個(gè)行業(yè)。它不僅是一項(xiàng)新興技術(shù)，更是一種可能重新定義我們生活和工作的工具。

挑戰(zhàn)：信息孤島和協(xié)作難題

現(xiàn)在的工作方式要求員工在非常專業(yè)的領(lǐng)域具備深入的技能。如果不了解具體背景，想快速上手項(xiàng)目就很難。這在涉及到實(shí)際物理操作的工作中表現(xiàn)得尤其明顯。比如設(shè)計(jì)師或工程師能快速?gòu)腃AD（計(jì)算機(jī)輔助設(shè)計(jì)）文件中看出項(xiàng)目的關(guān)鍵細(xì)節(jié)，但如果是非專業(yè)領(lǐng)域的人，比如營(yíng)銷、財(cái)務(wù)、供應(yīng)鏈或項(xiàng)目管理人員，就很難理解這些文件里的內(nèi)容，這導(dǎo)致了信息被孤立，團(tuán)隊(duì)協(xié)作受限。

空間計(jì)算如何改變這一現(xiàn)狀

空間計(jì)算能讓團(tuán)隊(duì)協(xié)作變得更加簡(jiǎn)單?？臻g計(jì)算可以用新的方式把業(yè)務(wù)數(shù)據(jù)可視化，讓客戶、員工更容易理解和互動(dòng)。它能把物理和數(shù)字結(jié)合起來(lái)，打造一個(gè)沉浸式的技術(shù)環(huán)境，讓人與世界的交互更加自然。

舉個(gè)例子：空間計(jì)算可以通過(guò)一個(gè)可視化交互界面，直接從業(yè)務(wù)軟件中提取相關(guān)數(shù)據(jù)。比如，供應(yīng)鏈工作人員可以快速找到需要訂購(gòu)的零件，而營(yíng)銷人員能更直觀地理解產(chǎn)品的外觀，從而更高效地制定推廣方案。通過(guò)這種方式，不同部門(mén)的人都能輕松獲取項(xiàng)目信息，快速做出決策，而不用因?yàn)閷I(yè)壁壘而卡殼。

不僅僅是炫酷的VR頭顯

說(shuō)到空間計(jì)算，很多人會(huì)想到那些看上去很酷的VR頭顯。但空間計(jì)算并不僅僅是“戴個(gè)眼鏡看畫(huà)面”這么簡(jiǎn)單。它是將傳感器數(shù)據(jù)、物聯(lián)網(wǎng)、無(wú)人機(jī)、激光雷達(dá)（LIDAR）等技術(shù)整合起來(lái)，打造出能真實(shí)還原業(yè)務(wù)操作的數(shù)字模型。無(wú)論是傳統(tǒng)的二維屏幕、輕便的增強(qiáng)現(xiàn)實(shí)（AR）眼鏡，還是完全沉浸式的VR環(huán)境，都可以用來(lái)呈現(xiàn)這些模型。

空間計(jì)算的核心是感知真實(shí)的物理環(huán)境，使用技術(shù)將物理和數(shù)字連接起來(lái)，并將數(shù)字信息疊加到一個(gè)融合的界面上（見(jiàn)圖1）。

多元化應(yīng)用，正在改變各行各業(yè)

現(xiàn)在，空間計(jì)算的應(yīng)用已經(jīng)覆蓋了多個(gè)領(lǐng)域，實(shí)時(shí)模擬是目前的主要應(yīng)用場(chǎng)景之一。未來(lái)，隨著技術(shù)的進(jìn)步，這項(xiàng)技術(shù)將推動(dòng)更多創(chuàng)新，重新定義醫(yī)療、制造、物流和娛樂(lè)等行業(yè)。這也是為什么從2022年到2033年，空間計(jì)算市場(chǎng)預(yù)計(jì)每年將增長(zhǎng)18.2%。

從現(xiàn)在到未來(lái)，空間計(jì)算將徹底改變我們與數(shù)字和物理世界互動(dòng)的方式，讓工作和生活變得更高效、更有趣，同時(shí)也帶來(lái)更多新機(jī)會(huì)！

Now: Filled to the Rim with Sims

At its heart, spatial computing brings the digital world closer to lived reality. Many business processes have a physical component, particularly in asset-heavy industries, but, too often, information about those processes is abstracted, and the essence (and insight) is lost.

Businesses can learn much about their operations from well-organized, structured business data, but adding physical data can help them understand those operations more deeply. That’s where spatial computing comes in.

“This idea of being served the right information at the right time with the right view is the promise of spatial computing,” says David Randle, global head of go-to-market for spatial computing at Amazon Web Services (AWS). “We believe spatial computing enables more natural understanding and awareness of physical and virtual worlds.”

Advanced Simulations: A Primary Application

One of the primary applications unlocked by spatial computing is advanced simulations. Think digital twins, but rather than virtual representations that monitor physical assets, these simulations allow organizations to test different scenarios to see how various conditions will impact their operations.

Imagine:

? A manufacturing company where designers, engineers, and supply chain teams can seamlessly work from a single 3D model to craft, build, and procure all the parts they need.

? Doctors who can view true-to-life simulations of their patients’ bodies through augmented reality displays.

? An oil and gas company that can layer detailed engineering models on top of 2D maps.

The possibilities are as vast as our physical world is varied.

The Portuguese soccer club Benfica’s sports data science team, for example, uses cameras and computer vision to track players throughout matches and develop full-scale 3D models of every move its players make.

The cameras collect 2,000 data points from each player, and AI helps identify specific players, the direction they were facing, and critical factors that fed into their decision-making.

The data essentially creates a digital twin of each player, allowing the team to run simulations of how plays would have worked if a player was in a different position. X’s and O’s on a chalkboard are now three-dimensional models that coaches can experiment with.

“There’s been a huge evolution in AI pushing these models forward, and now we can use them in decision-making,” says Joao Copeto, chief information and technology officer at Sport Lisboa e Benfica.

This isn’t only about wins and losses—it’s also about dollars and cents. Benfica has turned player development into a profitable business by leveraging data and AI.

Over the past 10 years, the team has generated some of the highest player-transfer deals in Europe. Similar approaches could also pay dividends in warehouse operations, supply chain and logistics, or any other resource planning process.

Simulations in Medicine

Advanced simulations are also showing up in medical settings.

For instance:

? Virtual patient scenarios can be simulated as a training supplement for nurses or doctors in a more dynamic, self-paced environment than textbooks would allow.

? Fraser Health Authority in Canada has pioneered the use of simulation models to improve care by creating a system-wide digital twin.

This public health authority in British Columbia generated powerful visualizations of patient movement through different care settings and simulations to determine the impact of deploying different care models on patient access.

Although the work is ongoing, Fraser expects improvement in appropriate, need-based access to care through increased patient awareness of available services.

New: Data is the Differentiator

Enterprise IT teams will likely need to overcome significant hurdles to develop altogether-new spatial computing applications. They likely haven’t faced these hurdles when implementing more conventional software-based projects.

For one thing, data isn’t always interoperable between systems, which limits the ability to blend data from different sources.

Furthermore, the spaghetti diagrams mapping out the path that data travels in most organizations are circuitous at best, and building the data pipelines to get the correct spatial data into visual systems is a thorny engineering challenge.

Ensuring that data is of high quality and faithfully mirrors real-world conditions may be one of the most significant barriers to using spatial computing effectively.

Rethinking Spatial Data Management

David Randle of AWS notes that spatial data has not historically been well managed at most organizations, even though it represents some of a business’s most valuable information.

“This information, because it’s quite new and diverse, has few standards around it and much of it sits in silos, some of it’s in the cloud, most of it’s not,” says Randle. “This data landscape encompassing physical and digital assets is extremely scattered and not well managed. Our customers’ first problem is managing their spatial data.”

Taking a more systematic approach to ingesting, organizing, and storing this data, in turn, makes it more available to modern AI tools, and that’s where the real learnings begin.

Data Pipelines: The Fuel for Business

We’ve often heard that data is the new oil, but for an American oil and gas company, the metaphor is becoming reality thanks to significant effort in replumbing some of its data pipelines.

The energy company uses drones to conduct 3D scans of equipment in the field and its facilities, and then applies computer vision to the data to ensure its assets operate within predefined tolerances.

It’s also creating high-fidelity digital twins of assets based on data pulled from engineering, operational, and enterprise resource planning systems.

The critical piece? Data integration.

The energy giant built a spatial storage layer, using application program interfaces to connect to disparate data sources and file types, including machine, drone, business, and image and video data.

Few organizations today have invested in this type of systematic approach to ingesting and storing spatial data. Still, it’s a key factor driving spatial computing capabilities and an essential first step for delivering impactful use cases.

模擬技術(shù)大展身手

空間計(jì)算的核心，是讓數(shù)字世界更貼近現(xiàn)實(shí)生活。很多業(yè)務(wù)流程都跟物理世界相關(guān)，特別是在那些“重資產(chǎn)”行業(yè)中，但問(wèn)題是，這些流程的信息常常被抽象化，關(guān)鍵的細(xì)節(jié)和洞察就這么丟掉了。

企業(yè)當(dāng)然能從有條理、結(jié)構(gòu)化的數(shù)據(jù)中學(xué)到不少東西，但如果再加入物理數(shù)據(jù)，他們對(duì)業(yè)務(wù)的理解會(huì)更加深入。這就是空間計(jì)算大顯身手的地方。

“空間計(jì)算的優(yōu)勢(shì)在于：在正確的時(shí)間，以正確的方式，提供正確的信息?！?a href="http://www.wenjunhu.com/tags/亞馬遜/" target="_blank">亞馬遜云服務(wù)（AWS）全球空間計(jì)算負(fù)責(zé)人David Randle說(shuō)道，“我們相信，空間計(jì)算能幫助人們更自然地理解和感知真實(shí)世界和虛擬世界?！?/p>

高級(jí)模擬：空間計(jì)算的拿手好戲

空間計(jì)算解鎖了一個(gè)非常重要的應(yīng)用——高級(jí)模擬。這不只是傳統(tǒng)的“數(shù)字孿生”概念，而是更進(jìn)一步。除了虛擬化地監(jiān)控物理資產(chǎn)，它還能讓企業(yè)測(cè)試各種情境，看看不同條件會(huì)如何影響業(yè)務(wù)運(yùn)營(yíng)。

舉幾個(gè)例子：

1、一個(gè)制造企業(yè)，設(shè)計(jì)師、工程師和供應(yīng)鏈團(tuán)隊(duì)可以一起通過(guò)同一個(gè)3D模型，完成設(shè)計(jì)、制造和采購(gòu)所有零部件的任務(wù)。

2、醫(yī)生可以通過(guò)增強(qiáng)現(xiàn)實(shí)設(shè)備，查看幾乎真實(shí)還原的患者身體模型，更直觀地了解病情。

3、石油和天然氣公司可以把精細(xì)的工程模型直接疊加在二維地圖上，輕松規(guī)劃業(yè)務(wù)。

這些應(yīng)用場(chǎng)景，可能性幾乎和物理世界一樣多樣。

再來(lái)看看一個(gè)特別的例子：葡萄牙足球俱樂(lè)部本菲卡的體育數(shù)據(jù)科學(xué)團(tuán)隊(duì)，利用攝像機(jī)和計(jì)算機(jī)視覺(jué)技術(shù)，實(shí)時(shí)跟蹤球員的比賽動(dòng)作，并為每名球員生成完整的3D模型。

這些攝像機(jī)能從每個(gè)球員身上收集2000多個(gè)數(shù)據(jù)點(diǎn)，而AI會(huì)幫忙識(shí)別球員的身份、他面對(duì)的方向，以及影響他決策的關(guān)鍵因素。通過(guò)這些數(shù)據(jù)，俱樂(lè)部實(shí)際上為每個(gè)球員創(chuàng)建了一個(gè)數(shù)字孿生，可以用來(lái)模擬“如果某個(gè)球員位置不同，比賽戰(zhàn)術(shù)會(huì)如何變化”。那些過(guò)去畫(huà)在戰(zhàn)術(shù)板上的“X”和“O”，現(xiàn)在成了教練可以隨意調(diào)整的3D模型。

“AI在推動(dòng)這些模型方面真的有了巨大的進(jìn)步，現(xiàn)在我們能用它們來(lái)幫助做出更好的決策。”Joao Copeto，本菲卡俱樂(lè)部的首席信息和技術(shù)官這樣說(shuō)道。

Multimodal AI Creates the Context

In the past, businesses couldn’t merge spatial and business data into one visualization, but that too is changing. As discussed in “What’s next for AI?” multimodal AI—AI tools that can process virtually any data type as a prompt and return outputs in multiple formats—is already adept at processing virtually any input, whether text, image, audio, spatial, or structured data types.

This capability will allow AI to serve as a bridge between different data sources, and interpret and add context between spatial and business data. AI can reach into disparate data systems and extract relevant insights.

This isn’t to say multimodal AI eliminates all barriers. Organizations still need to manage and govern their data effectively. The old saying “garbage in, garbage out” has never been more prescient. Training AI tools on disorganized and unrepresentative data is a recipe for disaster, as AI has the power to scale errors far beyond what we’ve seen with other types of software.

Enterprises should focus on implementing open data standards and working with vendors to standardize data types.

But once they’ve addressed these concerns, IT teams can open new doors to exciting applications. “You can shape this technology in new and creative ways,” says Johan Eerenstein, executive vice president of workforce enablement at Paramount.

Next: AI Is the New UI

Many of the aforementioned challenges in spatial computing are related to integration. Enterprises struggle to pull disparate data sources into a visualization platform and render that data in a way that provides value to the user in their day-to-day work. But soon, AI stands to lower those hurdles.

As mentioned above, multimodal AI can take a variety of inputs and make sense of them in one platform, but that could be only the beginning. As AI is integrated into more applications and interaction layers, it allows services to act in concert. As mentioned in “What’s next for AI?” this is already giving way to agentic systems that are context-aware and capable of executing functions proactively based on user preferences.

These autonomous agents could soon support the roles of supply chain manager, software developer, financial analyst, and more.

What will separate tomorrow’s agents from today’s bots will be their ability to plan ahead and anticipate what the user needs without even having to ask. Based on user preferences and historical actions, they will know how to serve the right content or take the right action at the right time.

When AI agents and spatial computing converge, users won’t have to think about whether their data comes from a spatial system, such as LIDAR or cameras (with the important caveat that AI systems are trained on high-quality, well-managed, interoperable data in the first place), or account for the capabilities of specific applications.

With intelligent agents, AI becomes the interface, and all that’s necessary is to express a preference rather than explicitly program or prompt an application.

Imagine a bot that automatically alerts financial analysts to changing market conditions or one that crafts daily reports for the C-suite about changes in the business environment or team morale.

All the many devices we interact with today, be they phone, tablet, computer, or smart speaker, will feel downright cumbersome in a future where all we have to do is gesture toward a preference and let context-aware, AI-powered systems execute our command. Eventually, once these systems have learned our preferences, we may not even need to gesture at all.

The Full Impact

The full impact of agentic AI systems on spatial computing may be many years out, but businesses can still work toward reaping the benefits of spatial computing. Building the data pipelines may be one of the heaviest lifts, but once built, they open up myriad use cases.

Autonomous asset inspection, smoother supply chains, true-to-life simulations, and immersive virtual environments are just a few ways leading enterprises are making their operations more spatially aware.

As AI continues to intersect with spatial systems, we’ll see the emergence of revolutionary new digital frontiers, the contours of which we’re only beginning to map out.

多模態(tài)AI創(chuàng)造上下文

過(guò)去，企業(yè)無(wú)法將空間數(shù)據(jù)和業(yè)務(wù)數(shù)據(jù)合并到一個(gè)可視化界面中，但這種局面正在改變。多模態(tài)AI——能夠處理幾乎任何類型數(shù)據(jù)輸入，并以多種格式輸出的AI工具——已經(jīng)非常擅長(zhǎng)處理各種輸入數(shù)據(jù)，包括文本、圖像、音頻、空間數(shù)據(jù)和結(jié)構(gòu)化數(shù)據(jù)。

這種能力將使AI成為連接不同數(shù)據(jù)源的橋梁，幫助解釋并建立空間數(shù)據(jù)與業(yè)務(wù)數(shù)據(jù)之間的上下文關(guān)聯(lián)。AI可以深入分散的數(shù)據(jù)系統(tǒng)，提取相關(guān)洞察并為決策提供支持。

這并不意味著多模態(tài)AI可以消除所有障礙。企業(yè)仍需要有效管理和治理數(shù)據(jù)。俗話說(shuō)“輸入垃圾，輸出垃圾”（Garbage in, garbage out），在AI時(shí)代，這句話比以往更加貼切。如果用混亂或不具代表性的數(shù)據(jù)訓(xùn)練AI工具，其錯(cuò)誤會(huì)被放大到前所未見(jiàn)的程度。因此，企業(yè)應(yīng)優(yōu)先實(shí)施開(kāi)放的數(shù)據(jù)標(biāo)準(zhǔn)，并與供應(yīng)商合作實(shí)現(xiàn)數(shù)據(jù)類型的標(biāo)準(zhǔn)化。

一旦這些問(wèn)題解決，IT團(tuán)隊(duì)就可以探索令人興奮的新應(yīng)用領(lǐng)域。“你可以用新穎且富有創(chuàng)造力的方式塑造這項(xiàng)技術(shù)。”派拉蒙公司（Paramount）負(fù)責(zé)員工賦能的執(zhí)行副總裁Johan Eerenstein說(shuō)道。

AI是全新的用戶界面

空間計(jì)算的許多挑戰(zhàn)都與數(shù)據(jù)集成相關(guān)。企業(yè)通常難以將分散的數(shù)據(jù)源整合到一個(gè)可視化平臺(tái)中，并以對(duì)日常工作有實(shí)際價(jià)值的方式呈現(xiàn)數(shù)據(jù)。但AI的加入將很快降低這些障礙。

正如前文提到的，多模態(tài)AI能夠處理各種輸入數(shù)據(jù)，并在一個(gè)平臺(tái)上進(jìn)行解析，但這可能只是一個(gè)開(kāi)始。隨著AI逐漸融入更多應(yīng)用和交互層，它能讓服務(wù)之間形成協(xié)作。這種趨勢(shì)已經(jīng)催生了具有上下文感知能力的自主系統(tǒng)，它們可以根據(jù)用戶偏好主動(dòng)執(zhí)行功能。

未來(lái)，這些自主智能代理將支持供應(yīng)鏈管理、軟件開(kāi)發(fā)、金融分析等角色。與如今的簡(jiǎn)單聊天機(jī)器人不同，明天的AI代理將具備前瞻性規(guī)劃能力，能夠提前預(yù)測(cè)用戶需求，而不需要明確指令。基于用戶偏好和歷史行為，它們將能夠在恰當(dāng)?shù)臅r(shí)機(jī)提供合適的內(nèi)容或采取正確的行動(dòng)。

當(dāng)AI代理與空間計(jì)算結(jié)合，用戶無(wú)需擔(dān)心數(shù)據(jù)是來(lái)自激光雷達(dá)（LIDAR）、攝像機(jī)，還是其他空間系統(tǒng)（前提是AI系統(tǒng)以高質(zhì)量、管理良好且互通的數(shù)據(jù)為基礎(chǔ)訓(xùn)練）。智能代理將使AI成為全新的界面，用戶只需表達(dá)一個(gè)偏好，而不需要明確編程或輸入復(fù)雜指令。

想象一下：

? 一個(gè)AI機(jī)器人能自動(dòng)向金融分析師發(fā)出市場(chǎng)變化警報(bào)。

? 或者，它每天為管理層編寫(xiě)關(guān)于業(yè)務(wù)環(huán)境變化或團(tuán)隊(duì)士氣的報(bào)告。

未來(lái)，所有我們今天使用的設(shè)備——手機(jī)、平板、電腦、智能音箱——可能都會(huì)顯得笨拙。那時(shí)，我們只需通過(guò)一個(gè)簡(jiǎn)單的手勢(shì)，甚至無(wú)需動(dòng)作，就能讓這些上下文感知的AI系統(tǒng)完成命令。

邁向新數(shù)字前沿的第一步

雖然自主智能AI系統(tǒng)在空間計(jì)算中的全面影響可能還需要幾年時(shí)間才能實(shí)現(xiàn)，但企業(yè)已經(jīng)可以開(kāi)始利用空間計(jì)算帶來(lái)的好處。構(gòu)建數(shù)據(jù)管道可能是最繁重的工作，但一旦完成，就能解鎖無(wú)數(shù)應(yīng)用場(chǎng)景，比如：

? 自動(dòng)資產(chǎn)檢測(cè)

? 更流暢的供應(yīng)鏈

? 真實(shí)感更強(qiáng)的模擬

? 沉浸式虛擬環(huán)境

一些領(lǐng)先企業(yè)已經(jīng)開(kāi)始利用這些方式讓運(yùn)營(yíng)更具有空間感知能力。隨著AI與空間系統(tǒng)的不斷交匯，我們將看到新的數(shù)字領(lǐng)域的誕生，這些領(lǐng)域的輪廓我們現(xiàn)在還只是剛剛開(kāi)始繪制。

What’s Next for AI?

While large language models continue to advance, new models and agents are proving to be more effective at discrete tasks. AI needs different horses for different courses.

The Speed of AI’s Advancement

Blink and you’ll miss it: The speed of artificial intelligence’s advancement is outpacing expectations.

Last year, as organizations scrambled to understand how to adopt generative AI, we cautioned Tech Trends 2024 readers to lead with need as they differentiate themselves from competitors and adopt a strategic approach to scaling their use of large language models (LLMs).

Today, LLMs have taken root, with up to 70% of organizations, by some estimates, actively exploring or implementing LLM use cases.1

Leading organizations are already considering AI’s next chapter. Instead of relying on foundation models built by large players in AI, which may be more powerful and built on more data than needed, enterprises are now thinking about implementing multiple, smaller models that can be more efficient for business requirements.2

LLMs will continue to advance and be the best option for certain use cases, like general-purpose chatbots or simulations for scientific research, but the chatbot that peruses your financial data to think through missed revenue opportunities doesn’t need to be the same model that replies to customer inquiries.

Put simply, we’re likely to see a proliferation of different horses for different courses.

A series of smaller models working in concert may end up serving different use cases than current LLM approaches. New open-source options and multimodal outputs (as opposed to just text) are enabling organizations to unlock entirely new offerings.3

In the years to come, the progress toward a growing number of smaller, more specialized models could once again move the goalposts of AI in the enterprise.

From Knowledge to Execution

Organizations may witness a fundamental shift in AI from augmenting knowledge to augmenting execution.

Investments being made today in agentic AI could upend the way we work and live by arming consumers and businesses with armies of silicon-based assistants.

Imagine AI agents that can carry out discrete tasks, like delivering a financial report in a board meeting or applying for a grant.

“There’s an app for that” could well become “There’s an agent for that.”

Now: Getting the Fundamentals Right

LLMs are undoubtedly exciting but require a great deal of groundwork.

Instead of building models themselves, many enterprises are partnering with companies like Anthropic or OpenAI, or accessing AI models through hyperscalers.?

According to Gartner, AI servers will account for close to 60% of hyperscalers’ total server spending.?

While some enterprises have found immediate business value in using LLMs, others remain wary about the accuracy and applicability of LLMs trained on external data.?

On an enterprise time scale, AI advancements are still in a nascent phase (crawling or walking, as noted last year). According to recent surveys by Deloitte, Fivetran, and Vanson Bourne, fewer than one-third of generative AI experiments have moved into production, often because organizations struggle to access or cleanse the data needed to run AI programs.

Data as the Foundation

According to Deloitte’s 2024 State of Generative AI in the Enterprise Q3 report, 75% of surveyed organizations have increased their investments in data lifecycle management due to generative AI.?

? Data is foundational to LLMs, because bad inputs lead to worse outputs (“garbage in, garbage squared”).

? Data labeling costs can drive significant AI investments.?

While some AI companies scrape the internet to build the largest models possible, savvy enterprises create the smartest models possible, with better domain-specific data education.

Example:

LIFT Impact Partners, a Vancouver-based organization, is fine-tuning its AI-enabled virtual assistants to help new Canadian immigrants process paperwork.

“When you train it on your organization’s unique persona, data, and culture, it becomes significantly more relevant and effective,” says Bruce Dewar, president and CEO of LIFT Impact Partners.

Challenges with Data Enablement

Organizations surveyed by Deloitte noted challenges with:

? Scaling AI pilots

? Unclear regulations around sensitive data

? Questions about third-party licensed data usage

55% of organizations avoided certain AI use cases due to data-related issues, and an equal proportion are working to enhance data security.

Differentiation:

While out-of-the-box models offered by vendors can help, differentiated AI impact will likely require differentiated enterprise data.

Real-World Value

Two-thirds of organizations surveyed are increasing investments in generative AI after seeing strong value across industries, from:

? Insurance claims review

? Telecom troubleshooting

? Consumer segmentation tools

LLMs are also creating value in specialized use cases like space repairs, nuclear modeling, and material design.

New: Different Horses for Different Courses

While LLMs have vast use cases, they are not always the most efficient choice.

? LLMs require massive resources, focus primarily on text, and augment human intelligence rather than execute discrete tasks.

? Smaller, purpose-built models can better address specific needs.

Future:

In the next 18–24 months, enterprises will likely rely on a toolkit of AI models, including:

1. Small language models (SLMs)

2. Multimodal models

3. Agentic AI systems

These models will help organizations optimize specific tasks without relying on massive, general-purpose LLMs.

Example:

An SLM trained on inventory data could let employees retrieve insights quickly, avoiding manual processing of large datasets that can take weeks.

二、人工智能的下一步是什么？

大型語(yǔ)言模型（LLMs）還在不斷進(jìn)化，但新的AI模型和代理（agents）在某些特定任務(wù)上表現(xiàn)得更加高效。用一句俗話來(lái)說(shuō)：“對(duì)癥下藥，才能事半功倍?！?/p>

就在去年，各企業(yè)還在努力弄清楚如何擁抱生成式AI時(shí)，我們就提醒過(guò)讀者，要以實(shí)際需求為導(dǎo)向，用戰(zhàn)略性的方式將大型語(yǔ)言模型（LLMs）落地應(yīng)用。

如今，LLMs已經(jīng)廣泛應(yīng)用，有數(shù)據(jù)估計(jì)多達(dá)70%的企業(yè)正在探索或?qū)嵤㎜LM的用例。

然而領(lǐng)先的企業(yè)已經(jīng)開(kāi)始考慮AI的下一個(gè)階段：與其依賴那些由AI巨頭打造的超大型基礎(chǔ)模型——雖然它們功能強(qiáng)大、數(shù)據(jù)豐富，但往往超出實(shí)際需求——不如部署多個(gè)小型模型，更高效地滿足特定業(yè)務(wù)需求。

LLMs仍然是某些場(chǎng)景（比如通用聊天機(jī)器人或科學(xué)研究模擬）的最佳選擇，但一個(gè)用于分析財(cái)務(wù)數(shù)據(jù)、尋找收入增長(zhǎng)點(diǎn)的聊天機(jī)器人，真的需要和一個(gè)解答客戶詢問(wèn)的機(jī)器人用同樣的模型嗎？

換句話說(shuō)，我們會(huì)看到不同任務(wù)用不同AI模型的趨勢(shì)。

一系列小型模型可以協(xié)同工作，服務(wù)于當(dāng)前LLM難以覆蓋的用例。開(kāi)源模型的普及和多模態(tài)輸出（不僅限于文本）正在幫助企業(yè)解鎖全新的服務(wù)和產(chǎn)品。

未來(lái)幾年，小型、更專業(yè)化的模型會(huì)進(jìn)一步推動(dòng)AI在企業(yè)中的發(fā)展，并讓AI的“規(guī)則”再次被重新定義。

從“增強(qiáng)知識(shí)”到“增強(qiáng)執(zhí)行”

我們可能會(huì)看到AI從“幫助獲取知識(shí)”逐步轉(zhuǎn)向“幫助完成任務(wù)”的根本轉(zhuǎn)變。

目前在開(kāi)發(fā)中的代理型AI（Agentic AI），正是這種趨勢(shì)的代表。這些智能代理有望顛覆我們的工作和生活方式，為消費(fèi)者和企業(yè)提供一個(gè)“硅基助理軍團(tuán)”。

想象一下：

? 一個(gè)AI代理可以在董事會(huì)會(huì)議上呈現(xiàn)財(cái)務(wù)報(bào)告，甚至幫你申請(qǐng)一筆資金撥款。

? 我們常說(shuō)的“有什么應(yīng)用（app）可以解決這個(gè)問(wèn)題？”可能會(huì)演變成“有一個(gè)AI代理可以幫你搞定”。

現(xiàn)在：打好基礎(chǔ)是關(guān)鍵

雖然LLMs令人興奮，但要真正落地還需要扎實(shí)的基礎(chǔ)工作。

許多企業(yè)沒(méi)有自行開(kāi)發(fā)模型，而是選擇與Anthropic或OpenAI等公司合作，或者通過(guò)云計(jì)算巨頭（Hyperscalers）使用AI模型。

根據(jù)Gartner的預(yù)測(cè)，AI服務(wù)器的支出將占云計(jì)算巨頭總服務(wù)器支出的近60%。

一些企業(yè)已經(jīng)從LLMs中找到了直接的業(yè)務(wù)價(jià)值，但也有企業(yè)對(duì)基于外部數(shù)據(jù)訓(xùn)練的LLMs的準(zhǔn)確性和適用性心存顧慮。

現(xiàn)實(shí)是：

目前的AI發(fā)展階段還很早，類似“嬰兒學(xué)爬”或“剛學(xué)走路”的階段。根據(jù)德勤、Fivetran和Vanson Bourne的調(diào)查，只有不到三分之一的生成式AI實(shí)驗(yàn)進(jìn)入了生產(chǎn)階段，主要原因是企業(yè)在獲取或清理運(yùn)行AI所需數(shù)據(jù)時(shí)遇到了困難。

數(shù)據(jù)是AI的基石

根據(jù)德勤的2024年第三季度《企業(yè)生成式AI狀態(tài)報(bào)告》，75%的受訪企業(yè)因生成式AI而增加了在數(shù)據(jù)生命周期管理上的投資。

? 數(shù)據(jù)是LLMs的基礎(chǔ)，輸入數(shù)據(jù)差，輸出結(jié)果會(huì)更糟（俗話說(shuō)：“垃圾進(jìn)，垃圾出”）。

? 數(shù)據(jù)標(biāo)注成本也是AI投資的一個(gè)重要因素。

雖然一些AI公司通過(guò)抓取互聯(lián)網(wǎng)數(shù)據(jù)來(lái)構(gòu)建大規(guī)模模型，但聰明的企業(yè)更傾向于創(chuàng)建“更智能的模型”，通過(guò)領(lǐng)域?qū)Ｓ脭?shù)據(jù)進(jìn)行更好的訓(xùn)練。

案例：

位于溫哥華的LIFT Impact Partners是一家為非營(yíng)利組織提供資源的機(jī)構(gòu)，他們用經(jīng)過(guò)優(yōu)化的數(shù)據(jù)訓(xùn)練AI虛擬助手，幫助新移民辦理加拿大的移民手續(xù)。

“當(dāng)你用組織獨(dú)特的個(gè)性、數(shù)據(jù)和文化去訓(xùn)練AI，它會(huì)變得更加貼合實(shí)際，更高效。”LIFT的總裁兼首席執(zhí)行官Bruce Dewar說(shuō)道，“它不只是工具，更像是企業(yè)的延伸和代言人?！?/p>

數(shù)據(jù)面臨的挑戰(zhàn)

企業(yè)在AI落地過(guò)程中還面臨以下數(shù)據(jù)相關(guān)的挑戰(zhàn)：

? 如何讓AI試點(diǎn)項(xiàng)目順利擴(kuò)展？

? 對(duì)敏感數(shù)據(jù)的模糊法規(guī)？

? 外部數(shù)據(jù)（比如第三方許可數(shù)據(jù)）的使用問(wèn)題？

調(diào)查顯示：

55%的企業(yè)因?yàn)閿?shù)據(jù)問(wèn)題避免了某些AI用例，同時(shí)同樣比例的企業(yè)正在加強(qiáng)數(shù)據(jù)安全。

解決之道：

雖然使用供應(yīng)商提供的“開(kāi)箱即用”模型可以繞過(guò)部分問(wèn)題，但要實(shí)現(xiàn)差異化的AI價(jià)值，企業(yè)需要獨(dú)特的企業(yè)數(shù)據(jù)。

AI的現(xiàn)實(shí)價(jià)值

盡管有挑戰(zhàn)，AI帶來(lái)的回報(bào)也非常顯著：

1、三分之二的企業(yè)因?yàn)橐呀?jīng)看到了強(qiáng)大的業(yè)務(wù)價(jià)值而增加了對(duì)生成式AI的投資。

2、AI已經(jīng)在保險(xiǎn)索賠審核、電信故障排查、消費(fèi)者分層分析等領(lǐng)域展現(xiàn)了現(xiàn)實(shí)價(jià)值。

3、在更專業(yè)的場(chǎng)景中，LLMs也有建樹(shù)，比如太空維修、核反應(yīng)模擬和材料設(shè)計(jì)。

不同任務(wù)，用不同AI模型

LLMs覆蓋了廣泛的用例，但它們并不是萬(wàn)能的。

1、LLMs需要龐大的資源，主要用于處理文本，并且更擅長(zhǎng)“增強(qiáng)人類智能”，而非執(zhí)行具體任務(wù)。

2、小型語(yǔ)言模型（SLMs）和多模態(tài)模型可能更適合某些特定需求。

未來(lái)趨勢(shì)：

在接下來(lái)的18-24個(gè)月內(nèi)，企業(yè)可能會(huì)采用多種AI模型組合的方式，包括：

1. 小型語(yǔ)言模型（SLMs）

2. 多模態(tài)模型

3. 代理型AI系統(tǒng)

案例：

一家企業(yè)可以用庫(kù)存數(shù)據(jù)訓(xùn)練一款SLM，讓員工快速獲得洞察，而不是手動(dòng)處理大量數(shù)據(jù)——這可能需要數(shù)周時(shí)間。

通過(guò)這種方式，AI不僅變得更靈活，還讓企業(yè)在效率和成本之間找到更好的平衡點(diǎn)。

Naveen Rao, vice president of AI at Databricks, believes more organizations will take this systems approach with AI:

“A magic computer that understands everything is a sci-fi fantasy. Rather, in the same way we organize humans in the workplace, we should break apart our problems. Domain-specific and customized models can then address specific tasks, tools can run deterministic calculations, and databases can pull in relevant data. These AI systems deliver the solution better than any one component could do alone.”

An added benefit of smaller models is that they can be run on-device and trained by enterprises on smaller, highly curated data sets to solve more specific problems, rather than general queries, as discussed in “Hardware is eating the world.”

Companies like Microsoft and Mistral are currently working to distill such SLMs, built on fewer parameters, from their larger AI offerings, and Meta offers multiple options across smaller models and frontier models.

Finally, much of the progress happening in SLMs is through open-source models offered by companies like Hugging Face or Arcee.AI. Such models are ripe for enterprise use since they can be customized for any number of needs, as long as IT teams have the internal AI talent to fine-tune them.

In fact, a recent Databricks report indicates that over 75% of organizations are choosing smaller open-source models and customizing them for specific use cases. Since open-source models are constantly improving thanks to the contributions of a diverse programming community, the size and efficiency of these models are likely to improve at a rapid clip.

Humans interact through a variety of mediums: text, body language, voice, videos, among others. Machines are now hoping to catch up.

Given that business needs are not contained to text, it’s no surprise that companies are looking forward to AI that can take in and produce multiple mediums.

In some ways, we’re already accustomed to multimodal AI, such as when we speak to digital assistants and receive text or images in return, or when we ride in cars that use a mix of computer vision and audio cues to provide driver assistance.

Multimodal generative AI, on the other hand, is in its early stages. The first major models, Google’s Project Astra and OpenAI’s GPT-4 Omni, were showcased in May 2024, and Amazon Web Services’ Titan offering has similar capabilities.

Progress in multimodal generative AI may be slow because it requires significantly higher amounts of data, resources, and hardware. In addition, the existing issues of hallucination and bias that plague text-based models may be exacerbated by multimodal generation.

Still, the enterprise use cases are promising:

The notion of “train once, run anywhere (or any way)” promises a model that could be trained on text, but deliver answers in pictures, video, or sound, depending on the use case and the user’s preference, which improves digital inclusion.

Example Applications:

? Companies like AMD aim to use the fledgling technology to quickly translate marketing materials from English to other languages or to generate content.

? For supply chain optimization, multimodal generative AI can be trained on sensor data, maintenance logs, and warehouse images to recommend ideal stock quantities.

This also leads to new opportunities with spatial computing, as written about in “Spatial computing takes center stage.”

As the technology progresses and model architecture becomes more efficient, we can expect to see even more use cases in the next 18 to 24 months.

Agentic AI

The third new pillar of AI may pave the way for changes to our ways of working over the next decade.

Large (or small) action models go beyond the question-and-answer capabilities of LLMs and complete discrete tasks in the real world.

Examples:

? Booking a flight based on your travel preferences.

? Providing automated customer support that can access databases and execute needed tasks—likely without the need for highly specialized prompts.

The proliferation of such action models, working as autonomous digital agents, heralds the beginnings of agentic AI, and enterprise software vendors like Salesforce and ServiceNow are already touting these possibilities.

Enterprise Use Case: ServiceNow’s Xanadu Platform

Chris Bedi, chief customer officer at ServiceNow, believes that domain- or industry-specific agentic AI can change the game for human and machine interaction in enterprises.

For instance, in the company’s Xanadu platform, one AI agent can:

1. Scan incoming customer issues against a history of incidents to recommend next steps.

2. Communicate with another autonomous agent that executes those recommendations.

A human reviewer oversees the agent-to-agent communication to approve the hypotheses.

Other Use Cases:

? One agent could manage workloads in the cloud.

? Another agent could handle customer orders.

“Agentic AI cannot completely take the place of a human,” says Bedi, “but what it can do is work alongside your teams, handling repetitive tasks, seeking out information and resources, doing work in the background 24/7, 365 days a year.”

Liquid Neural Networks: A New AI Frontier

Aside from the categories of AI models noted above, advancements in AI design and execution are also impacting enterprise adoption—namely, the advent of liquid neural networks.

What are liquid neural networks?

? This cutting-edge technology offers greater flexibility by mimicking the human brain’s structure.

? Unlike traditional neural networks, which might require 100,000 nodes, liquid networks can accomplish tasks with just a couple dozen nodes.

Liquid neural networks are designed to run on less computing power with more transparency. This opens up possibilities for embedding AI into edge devices, robotics, and safety-critical systems.

In other words, it’s not just the applications of AI but also its underlying mechanisms that are ripe for improvement and disruption in the coming years.

小模型的系統(tǒng)化未來(lái)

Databricks副總裁Naveen Rao認(rèn)為，越來(lái)越多的企業(yè)會(huì)用系統(tǒng)化的方法來(lái)發(fā)展AI：“那種‘萬(wàn)能計(jì)算機(jī)無(wú)所不懂’的想法只是科幻電影的幻想。我們更應(yīng)該像管理人類團(tuán)隊(duì)一樣，把問(wèn)題分解開(kāi)。領(lǐng)域?qū)俸投ㄖ苹Ｐ涂梢越鉀Q具體任務(wù)，工具可以做確定性計(jì)算，數(shù)據(jù)庫(kù)則負(fù)責(zé)獲取相關(guān)數(shù)據(jù)。這些AI系統(tǒng)協(xié)同工作，提供的解決方案遠(yuǎn)比單一組件要強(qiáng)大得多?！?/p>

小型模型（SLMs）的優(yōu)勢(shì)之一是它們可以直接在設(shè)備上運(yùn)行，而且企業(yè)可以用高度定制化的小型數(shù)據(jù)集來(lái)訓(xùn)練這些模型，解決更具體的問(wèn)題，而不是應(yīng)對(duì)寬泛的需求。例如，微軟和Mistral正在開(kāi)發(fā)這種精簡(jiǎn)版的小型語(yǔ)言模型，而Meta則提供了多個(gè)小型模型和前沿模型供選擇。

此外，很多SLMs的進(jìn)步來(lái)自開(kāi)源模型，比如Hugging Face或Arcee.AI等公司提供的模型。這些開(kāi)源模型非常適合企業(yè)使用，因?yàn)樗鼈兛梢愿鶕?jù)不同需求進(jìn)行調(diào)整，只要企業(yè)的IT團(tuán)隊(duì)擁有調(diào)試這些模型的AI人才即可。一份Databricks的報(bào)告顯示，超過(guò)75%的企業(yè)正在選擇小型開(kāi)源模型，并將其定制用于具體場(chǎng)景。由于多樣化的開(kāi)發(fā)者社區(qū)不斷改進(jìn)這些開(kāi)源模型，模型的效率和規(guī)模預(yù)計(jì)將快速提升。

多模態(tài)模型的崛起

人類通過(guò)多種方式交流，比如文本、肢體語(yǔ)言、語(yǔ)音和視頻等。現(xiàn)在，機(jī)器也正在努力趕上這個(gè)水平。

企業(yè)的需求遠(yuǎn)超文本數(shù)據(jù)，這就是為什么多模態(tài)AI開(kāi)始成為大家關(guān)注的焦點(diǎn)。其實(shí)，我們已經(jīng)接觸到了一些多模態(tài)AI的應(yīng)用，比如，當(dāng)我們和數(shù)字助手對(duì)話時(shí)，它可以以文本或圖像的形式回復(fù)我們；或者我們開(kāi)車時(shí)，車輛通過(guò)計(jì)算機(jī)視覺(jué)和音頻提示提供駕駛輔助。

然而，多模態(tài)生成式AI還處于起步階段。2024年5月，谷歌的Project Astra、OpenAI的GPT-4 Omni，以及亞馬遜云服務(wù)（AWS）的Titan展示了早期的多模態(tài)AI技術(shù)。這些技術(shù)的進(jìn)展較慢，原因在于它們需要大量的數(shù)據(jù)、資源和硬件支持。此外，目前文字生成AI存在的“幻覺(jué)”和偏見(jiàn)問(wèn)題在多模態(tài)生成中可能更加突出。

企業(yè)應(yīng)用前景：

多模態(tài)AI可以“一次訓(xùn)練，多場(chǎng)景輸出”，比如基于文本數(shù)據(jù)訓(xùn)練的模型可以根據(jù)用戶需求，以圖片、視頻或音頻的形式提供答案。這種能力不僅提升了用戶體驗(yàn)，還促進(jìn)了數(shù)字包容性。

具體場(chǎng)景：

1、企業(yè)可以用它將營(yíng)銷材料快速?gòu)挠⑽姆g成其他語(yǔ)言，或者自動(dòng)生成內(nèi)容。

2、在供應(yīng)鏈優(yōu)化中，多模態(tài)AI可以結(jié)合傳感器數(shù)據(jù)、維護(hù)記錄和倉(cāng)庫(kù)圖像，推薦最佳的庫(kù)存量。

隨著技術(shù)的發(fā)展和模型架構(gòu)的效率提升，未來(lái)18到24個(gè)月內(nèi)會(huì)看到更多新的應(yīng)用場(chǎng)景。

代理型AI（Agentic AI）

AI的第三大趨勢(shì)可能在未來(lái)十年內(nèi)徹底改變我們的工作方式。

代理型AI不僅能夠回答問(wèn)題，還能完成現(xiàn)實(shí)世界中的具體任務(wù)。例如，幫助用戶根據(jù)個(gè)人偏好預(yù)訂航班，或者在無(wú)需復(fù)雜指令的情況下，提供自動(dòng)化的客戶支持。這些模型作為自主數(shù)字代理的普及標(biāo)志著代理型AI的開(kāi)端，像Salesforce和ServiceNow這樣的企業(yè)軟件供應(yīng)商，已經(jīng)開(kāi)始宣傳這些可能性。

企業(yè)案例：

ServiceNow的Xanadu平臺(tái)中，一個(gè)AI代理可以根據(jù)客戶問(wèn)題的歷史記錄生成下一步建議，然后將這些建議傳遞給另一個(gè)代理來(lái)執(zhí)行，而人類則只需在代理之間的溝通中進(jìn)行審核。這種協(xié)作模式可以擴(kuò)展到不同領(lǐng)域，比如一個(gè)代理專注于云端工作負(fù)載管理，另一個(gè)代理則負(fù)責(zé)為客戶下單。

ServiceNow的首席客戶官Chris Bedi表示：“代理型AI無(wú)法完全取代人類，但它可以成為團(tuán)隊(duì)的好助手，處理重復(fù)性的任務(wù)、查找信息和資源，并在后臺(tái)全天候工作?！?/p>

液態(tài)神經(jīng)網(wǎng)絡(luò)：AI技術(shù)的新突破

除了AI模型的種類，AI的設(shè)計(jì)和運(yùn)行機(jī)制也在快速進(jìn)步，比如液態(tài)神經(jīng)網(wǎng)絡(luò)的出現(xiàn)。這種網(wǎng)絡(luò)擁有更高的靈活性，其訓(xùn)練方法模仿了人腦的結(jié)構(gòu)。與傳統(tǒng)網(wǎng)絡(luò)需要十萬(wàn)個(gè)節(jié)點(diǎn)不同，液態(tài)神經(jīng)網(wǎng)絡(luò)可能只需要幾十個(gè)節(jié)點(diǎn)就能完成類似的任務(wù)。

這種尖端技術(shù)不僅能顯著降低計(jì)算需求，還能提供更高的透明性，使得AI更適合嵌入邊緣設(shè)備、機(jī)器人和關(guān)鍵安全系統(tǒng)中。

換句話說(shuō)，未來(lái)的AI不僅在應(yīng)用場(chǎng)景上會(huì)帶來(lái)更多可能，它的底層技術(shù)也在醞釀新的顛覆。

Next: There’s an Agent for That

In the next decade, AI could be wholly focused on execution instead of human augmentation.

A future employee could make a plain-language request to an AI agent, for example:

“Close the books for Q2 and generate a report on EBITDA.”

Like in an enterprise hierarchy, the primary agent would then delegate the needed tasks to agents with discrete roles that cascade across different productivity suites to take action.

As with humans, teamwork could be the missing ingredient that enables the machines to improve their capabilities.

This leads to a few key considerations for the years to come (figure 2):

1. AI-to-AI Communication

Agents will likely have a more efficient way of communicating with each other than human language, as we don’t need human-imitating chatbots talking to each other.

Better AI-to-AI communication can enhance outcomes, as fewer people will need to become experts to benefit from AI. Rather, AI can adapt to each person’s communication style.

2. Job Displacement and Creation

Some claim that roles such as prompt engineer could become obsolete.

However, the AI expertise of those employees will remain pertinent as they focus on managing, training, and collaborating with AI agents as they do with LLMs today.

For example:

A lean IT team with AI experts might build the agents it needs in a sort of “AI factory” for the enterprise.

The significant shift in the remaining workforce’s skills and education may ultimately reward more human skills like creativity and design, as mentioned in previous Tech Trends.

3. Privacy and Security

The proliferation of agents with system access is likely to raise broad concerns about cybersecurity.

This will only become more important as time progresses and more of our data is accessed by AI systems.

New paradigms for risk and trust will be required to make the most out of applying AI agents.

4. Energy and Resources

AI’s energy consumption is a growing concern.

To mitigate environmental impacts, future AI development will need to balance performance with sustainability.

It will need to take advantage of improvements in liquid neural networks or other efficient forms of training AI—not to mention the hardware needed to make all of this work, as we discuss in “Hardware is Eating the World.”

5. Leadership for the Future

AI has transformative potential, as everyone has heard plenty over the last year, but only insofar as leadership allows.

Applying AI as a faster way of doing things the way they’ve always been done will result in:

? At best: Missed potential

? At worst: Amplified biases

Imaginative, courageous leaders should dare to take AI from calcified best practices to the creation of “next practices,” where we find new ways of organizing ourselves and our data toward an AI-enabled world.

Future Considerations: Data, Data, and More Data

When it comes to AI, enterprises will likely have the same considerations in the future that they do today:

Data, data, and data.

Until AI systems can reach artificial general intelligence or learn as efficiently as the human brain, they will remain hungry for more data and inputs to help them be more powerful and accurate.

Steps taken today to organize, streamline, and protect enterprise data could pay dividends for years to come, as data debt could one day become the biggest portion of technical debt.

Such groundwork should also help enterprises prepare for the litany of regulatory challenges and ethical uncertainties (such as data collection and use limitations, fairness concerns, and lack of transparency) that come with shepherding this new, powerful technology into the future.

The stakes of “garbage in, garbage out” are only going to grow:

It would be much better to opt for genius in, genius squared.

每個(gè)任務(wù)都可以有一個(gè)AI代理

在未來(lái)十年，AI可能完全專注于執(zhí)行任務(wù)，而非僅僅增強(qiáng)人類的能力。

想象一下：

一位員工可以對(duì)AI代理發(fā)出簡(jiǎn)單的指令，例如“完成第二季度的賬目并生成一份EBITDA（稅息折舊及攤銷前利潤(rùn)）報(bào)告”。主代理會(huì)像企業(yè)的分層管理一樣，將任務(wù)分派給具有不同職責(zé)的代理，這些代理會(huì)跨多個(gè)生產(chǎn)力工具套件協(xié)同完成行動(dòng)。

正如人類團(tuán)隊(duì)合作能夠提升效率一樣，AI之間的團(tuán)隊(duì)合作可能成為推動(dòng)機(jī)器能力提升的關(guān)鍵因素。

以下是未來(lái)幾年需要考慮的幾個(gè)重點(diǎn)：

1. AI與AI之間的溝通

AI代理之間的溝通可能會(huì)比模仿人類語(yǔ)言更加高效。

我們不需要AI通過(guò)像人類一樣的聊天方式彼此交談，而是可以采用更加直接的機(jī)器語(yǔ)言進(jìn)行溝通。這種方式能夠提高AI之間的協(xié)作效果，同時(shí)降低人類需要掌握AI專業(yè)知識(shí)的門(mén)檻。

最終，AI可以適應(yīng)每個(gè)人的溝通風(fēng)格，讓更多人無(wú)需成為專家，也能從AI中受益。

2. 工作的取代與創(chuàng)造

有人擔(dān)心像“提示工程師”這樣的角色可能會(huì)變得過(guò)時(shí)。但實(shí)際上，這些具有AI專業(yè)技能的員工仍然會(huì)很重要，他們的職責(zé)會(huì)轉(zhuǎn)向管理、訓(xùn)練和與AI代理合作，就像他們現(xiàn)在處理大型語(yǔ)言模型（LLMs）一樣。

例如，一個(gè)精簡(jiǎn)的IT團(tuán)隊(duì)可以通過(guò)企業(yè)內(nèi)的“AI工廠”打造它們所需的AI代理，來(lái)支持各類任務(wù)。

此外，隨著工作技能和教育需求發(fā)生顯著變化，人類具備的創(chuàng)造力和設(shè)計(jì)能力等技能可能會(huì)變得更加寶貴。這一點(diǎn)在之前的《科技趨勢(shì)》中已經(jīng)提到過(guò)。

3. 隱私與安全

隨著AI代理的普及，它們對(duì)系統(tǒng)的訪問(wèn)權(quán)限會(huì)引發(fā)更多的網(wǎng)絡(luò)安全問(wèn)題。

這些問(wèn)題會(huì)隨著時(shí)間的推移和AI對(duì)更多數(shù)據(jù)的訪問(wèn)而變得更加重要。為了更好地利用AI代理，新的風(fēng)險(xiǎn)控制和信任管理范式將變得必不可少。

4. 能源與資源消耗

AI的能耗問(wèn)題正在成為一個(gè)日益增長(zhǎng)的關(guān)注點(diǎn)。

為了降低對(duì)環(huán)境的影響，未來(lái)的AI開(kāi)發(fā)需要在性能與可持續(xù)性之間找到平衡。這可能需要利用液態(tài)神經(jīng)網(wǎng)絡(luò)或其他高效的AI訓(xùn)練方法，同時(shí)改進(jìn)硬件技術(shù)（關(guān)于硬件，我們?cè)凇队布谕淌墒澜纭分杏猩钊胗懻摚?/p>

5. 為未來(lái)培養(yǎng)領(lǐng)導(dǎo)力

AI擁有改變世界的潛力，但這種潛力能否實(shí)現(xiàn)，很大程度上取決于領(lǐng)導(dǎo)者的決策與視野。

如果AI只是被用來(lái)加速現(xiàn)有的工作方式，那么最多只能帶來(lái)潛力的浪費(fèi)，最糟的情況下則可能放大現(xiàn)有的偏見(jiàn)。

真正有想象力和勇氣的領(lǐng)導(dǎo)者應(yīng)該敢于將AI引入“下一代實(shí)踐（Next Practices）”，通過(guò)創(chuàng)造性的方式重新組織數(shù)據(jù)和工作流程，構(gòu)建一個(gè)更高效、更智能的AI世界。

未來(lái)AI的核心依舊是數(shù)據(jù)

當(dāng)談到AI時(shí)，未來(lái)企業(yè)依舊需要考慮三個(gè)核心問(wèn)題：數(shù)據(jù)、數(shù)據(jù)，還是數(shù)據(jù)。

在AI系統(tǒng)能夠達(dá)到人工通用智能（AGI）或像人類大腦一樣高效學(xué)習(xí)之前，AI將始終需要更多的數(shù)據(jù)和輸入，來(lái)提升其能力和準(zhǔn)確性。

今天為組織、優(yōu)化和保護(hù)企業(yè)數(shù)據(jù)所做的努力，可能在未來(lái)多年里都會(huì)帶來(lái)巨大的回報(bào)。如果沒(méi)有做好這些基礎(chǔ)工作，企業(yè)可能面臨**“數(shù)據(jù)債務(wù)”**的積累，最終成為技術(shù)債務(wù)中最沉重的部分。

同時(shí)，這些數(shù)據(jù)準(zhǔn)備工作還可以幫助企業(yè)應(yīng)對(duì)AI帶來(lái)的各種監(jiān)管挑戰(zhàn)和倫理問(wèn)題，例如數(shù)據(jù)采集與使用的限制、公平性問(wèn)題以及透明度不足等。

“垃圾進(jìn)，垃圾出”的問(wèn)題只會(huì)變得更加嚴(yán)重，而目標(biāo)應(yīng)該是“天才輸入，天才輸出”。如果企業(yè)能夠在數(shù)據(jù)上投入更多的努力，那么未來(lái)AI代理帶來(lái)的價(jià)值將是不可估量的。

Hardware is Eating the World

After years of “software eating the world,” it’s hardware’s turn to feast.

We previewed in the computation chapter of Tech Trends 2024 that as Moore’s Law comes to its supposed end, the promise of the AI revolution increasingly depends on access to the appropriate hardware.

Case in point: NVIDIA is now one of the world’s most valuable (and watched) companies, as specialized chips become an invaluable resource for AI computation workloads.1

According to Deloitte research based on a World Semiconductor Trade Statistics forecast, the market for chips used only for generative AI is projected to reach over US$50 billion this year.2

A Critical Hardware Use Case: AI-Embedded End-User and Edge Devices

Take personal computers (PCs), for instance. For years, enterprise laptops have been commodified. But now, we may be on the cusp of a significant shift in computing, thanks to AI-embedded PCs.

Companies like AMD, Dell, and HP are already touting the potential for AI PCs to:

? “Future-proof” technology infrastructure

? Reduce cloud computing costs

? Enhance data privacy

With access to offline AI models for image generation, text analysis, and speedy data retrieval, knowledge workers could be supercharged by faster, more accurate AI.

That being said, enterprises should be strategic about refreshing end-user computation on a large scale—there’s no use wasting AI resources that are limited in supply.

The Cost of Advancements: Sustainability in Data Centers

Of course, all of these advancements come at a cost.

Data centers are a new focus of sustainability as the energy demands of large AI models continue to grow.4

The International Energy Agency has suggested that the demands of AI will significantly increase electricity in data centers by 2026, equivalent to Sweden’s or Germany’s annual energy demands.5

A recent Deloitte study on powering AI estimates that global data center electricity consumption may triple in the coming decade, largely due to AI demand.6

Innovations in energy sources and efficiency are needed to make AI hardware more accessible and sustainable, even as it proliferates and finds its way into everyday consumer and enterprise devices.

Consider this: Unit 1 of the nuclear plant Three Mile Island, which was shut down five years ago due to economic reasons, will reopen by 2028 to power data centers with carbon-free electricity.7

Looking Forward: AI Hardware in IoT

AI hardware is poised to step beyond IT and into the Internet of Things (IoT).

An increasing number of smart devices could become even more intelligent as AI enables them to analyze their usage and take on new tasks (as agentic AI, mentioned in “What’s next for AI?” advances).

Today: Benign use cases, like AI in toothbrushes.

Tomorrow: Robust potential, like AI in lifesaving medical devices.

The true power of hardware could be unlocked when smarter devices bring about a step change in our relationship with robotics.

Now: Chips Ahoy!

A generation of technologists has been taught to believe software is the key to return on investment, given its scalability, ease of updates, and intellectual property protections.9

But now, hardware investment is surging as computers evolve from calculators to cogitators.10

We wrote last year that specialized chips like graphics-processing units (GPUs) were becoming the go-to resources for training AI models.

In its 2024 TMT Predictions report, Deloitte estimated that total AI chip sales in 2024 would be 11% of the predicted global chip market of US$576 billion.11

Growing from roughly US$50 billion today, the AI chip market is forecasted to reach up to US$400 billion by 2027, though a more conservative estimate is US$110 billion (figure 1).

三、硬件正在吞噬世界

在過(guò)去的幾年里，我們一直說(shuō)“軟件正在吞噬世界”，但現(xiàn)在輪到硬件登場(chǎng)了。

隨著摩爾定律逐漸失效，AI革命的未來(lái)越來(lái)越依賴于合適的硬件資源。舉個(gè)例子：NVIDIA（英偉達(dá)）現(xiàn)已成為全球最具價(jià)值、最受關(guān)注的公司之一，因?yàn)閷Ｓ眯酒殉蔀锳I計(jì)算任務(wù)中不可或缺的資源。

根據(jù)德勤基于“世界半導(dǎo)體貿(mào)易統(tǒng)計(jì)”預(yù)測(cè)的研究，僅用于生成式AI的芯片市場(chǎng)預(yù)計(jì)將在今年突破500億美元的規(guī)模。

企業(yè)硬件的關(guān)鍵用例：嵌入AI的終端設(shè)備

一個(gè)關(guān)鍵的硬件應(yīng)用場(chǎng)景可能在于嵌入AI的終端用戶設(shè)備和邊緣設(shè)備。例如，個(gè)人電腦（PC）行業(yè)在過(guò)去多年中已經(jīng)高度商品化，但隨著AI嵌入PC，我們可能正站在計(jì)算技術(shù)重大變革的起點(diǎn)。

AMD、戴爾、惠普等公司已經(jīng)在宣傳AI PC的潛力，認(rèn)為它們可以：

? “未來(lái)-proof”（未來(lái)適用）技術(shù)基礎(chǔ)設(shè)施

? 降低云計(jì)算成本

? 增強(qiáng)數(shù)據(jù)隱私

借助離線AI模型，知識(shí)工作者可以快速實(shí)現(xiàn)圖像生成、文本分析和數(shù)據(jù)檢索等功能，大幅提高工作效率和精度。

盡管如此，企業(yè)在大規(guī)模更新終端用戶設(shè)備時(shí)需要謹(jǐn)慎決策，因?yàn)锳I資源是有限的，浪費(fèi)它們沒(méi)有意義。

硬件背后的能源代價(jià)：可持續(xù)發(fā)展的壓力

當(dāng)然，所有這些技術(shù)進(jìn)步的背后都有代價(jià)。

隨著大型AI模型的能源需求不斷增長(zhǎng)，數(shù)據(jù)中心正在成為可持續(xù)發(fā)展的新焦點(diǎn)。國(guó)際能源署（IEA）預(yù)測(cè)，到2026年，AI的能源需求將使數(shù)據(jù)中心的用電量大幅增加，達(dá)到與瑞典或德國(guó)全年用電量相當(dāng)?shù)乃健?/p>

德勤的研究也估計(jì)，未來(lái)十年內(nèi)，由于AI需求的推動(dòng)，全球數(shù)據(jù)中心的電力消耗可能會(huì)增加三倍。

為應(yīng)對(duì)這一挑戰(zhàn)，需要在能源來(lái)源和能效創(chuàng)新方面進(jìn)行突破，以使AI硬件既可用又可持續(xù)。比如，美國(guó)三里島核電站的1號(hào)機(jī)組，五年前因經(jīng)濟(jì)原因關(guān)閉，但預(yù)計(jì)將在2028年重新開(kāi)放，為數(shù)據(jù)中心提供無(wú)碳電力支持。

硬件未來(lái)的展望：從IT到物聯(lián)網(wǎng)

展望未來(lái)，AI硬件將從IT領(lǐng)域擴(kuò)展到物聯(lián)網(wǎng)（IoT）。

越來(lái)越多的智能設(shè)備將變得更加智能，因?yàn)锳I賦予它們分析自身使用情況并承擔(dān)新任務(wù)的能力（這一點(diǎn)在“AI的下一步是什么？”中提到的代理型AI將繼續(xù)推動(dòng)）。

今天： 比如，AI被用在牙刷等看似普通的設(shè)備中。

明天： AI可能被嵌入救命的醫(yī)療設(shè)備中，其潛力遠(yuǎn)超目前的應(yīng)用。

當(dāng)更智能的設(shè)備能夠與機(jī)器人技術(shù)相結(jié)合，這種硬件將真正釋放出改變我們生活的力量，重新定義人類與機(jī)器的關(guān)系。

芯片崛起的時(shí)代

長(zhǎng)期以來(lái)，技術(shù)界普遍認(rèn)為軟件是投資回報(bào)的關(guān)鍵，因?yàn)樗哂锌蓴U(kuò)展性、易于更新和知識(shí)產(chǎn)權(quán)保護(hù)的優(yōu)勢(shì)。

但現(xiàn)在，隨著計(jì)算機(jī)從“計(jì)算器”進(jìn)化到“認(rèn)知者”，硬件投資正在快速崛起。

我們?nèi)ツ暝岬?，像圖形處理器（GPU）這樣的專用芯片正在成為訓(xùn)練AI模型的首選資源。

根據(jù)德勤2024年的《TMT預(yù)測(cè)報(bào)告》，AI芯片市場(chǎng)預(yù)計(jì)將在2024年占全球芯片市場(chǎng)（5760億美元）總量的11%。

目前估計(jì)AI芯片市場(chǎng)約為500億美元，但到2027年，這一數(shù)字可能增長(zhǎng)到4000億美元（較保守的預(yù)測(cè)為1100億美元，詳見(jiàn)圖1）。

Large Tech Companies and the Growing Demand for AI Hardware

Large tech companies are driving a portion of this demand, as they may build their own AI models and deploy specialized chips on-premises. However, enterprises across industries are seeking compute power to meet their IT goals.

For instance, according to a Databricks report, the financial services industry has had the highest growth in GPU usage, at 88% over the past six months, in running large language models (LLMs) that tackle fraud detection and wealth management.

All of this demand for GPUs has outpaced capacity. In today’s iteration of the Gold Rush, the companies providing “picks and shovels,” or the tools for today’s tech transformation, are winning big.

NVIDIA’s CEO Jensen Huang has noted that cloud GPU capacity is mostly filled, but the company is also rolling out new chips that are significantly more energy-efficient than previous iterations. Hyperscalers are buying up GPUs as they roll off the production line, spending almost $1 trillion on data center infrastructure to accommodate the demand from clients who rent GPU usage. All the while, the energy consumption of existing data centers is pushing aging power grids to the brink globally.

New Chips for a New Era: Neural Processing Units (NPUs)

Understandably, enterprises are looking for new solutions. While GPUs are crucial for handling the high workloads of LLMs or content generation, and central processing units are still table stakes, neural processing units (NPUs) are now in vogue.

NPUs, which mimic the brain’s neural network, can accelerate smaller AI workloads with greater efficiency and lower power demands. These chips enable enterprises to:

? Shift AI applications away from the cloud

? Apply AI locally to sensitive data that can’t be hosted externally

This new breed of chip is a crucial part of the future of embedded AI.

Vivek Mohindra, senior vice president of corporate strategy at Dell Technologies, notes:

“Of the 1.5 billion PCs in use today, 30% are four years old or more. None of these older PCs have NPUs to take advantage of the latest AI PC advancements.”

A major refresh of enterprise hardware may be on the horizon.

As NPUs enable end-user devices to run AI offline and allow models to become smaller to target specific use cases, hardware may once again become a differentiator for enterprise performance.

AI’s Transformative Potential

In a recent Deloitte study:

? 72% of respondents believe generative AI’s impact on their industry will be “high to transformative.”

Once AI becomes mainstream thanks to advancements in hardware, that number may edge closer to 100%.

New: Infrastructure is Strategic Again

The heady cloud-computing highs of assumed unlimited access are giving way to a resource-constrained era.

After being relegated to a utility for years, enterprise infrastructure (e.g., PCs) is once again strategic.

Specifically, specialized hardware will likely be crucial to three significant areas of AI growth:

1. AI-embedded devices and the Internet of Things (IoT)

2. Data centers

3. Advanced physical robotics

While the impact on robotics may occur over the next few years, enterprises will likely face decisions about the first two areas in the next 18 to 24 months.

1. Edge Computing Footprint

By 2025, more than 50% of data could be generated by edge devices.

As NPUs proliferate, more devices could run AI models without relying on the cloud. This trend is especially relevant as generative AI model providers focus on creating smaller, more efficient models for specific tasks.

With quicker response times, decreased costs, and greater privacy controls, hybrid computing (a mix of cloud and on-device AI workloads) may become essential for many enterprises. Hardware manufacturers are betting on it.

According to Dell Technologies’ Mohindra:

“Processing AI at the edge is one of the best ways to handle the vast amounts of data required. When you consider latency, network resources, and just sheer volume, moving data to a centralized compute location is inefficient, ineffective, and not secure. It’s better to bring AI to the data, rather than bring the data to AI.”

2. The Hardware Refresh is Coming

One major bank predicts that AI PCs will account for more than 40% of PC shipments by 2026.

Similarly, nearly 15% of 2024 smartphone shipments are expected to be capable of running LLMs or image-generation models.

HP’s Alex Thatcher compares this hardware refresh to the major transition from command-line inputs to graphical user interfaces in the 1990s:

“The software has fundamentally changed, replete with different tools and ways of collaborating. You need hardware that can accelerate that change and make it easier for enterprises to create and deliver AI solutions.”

Apple and Microsoft have also fueled this impending hardware refresh by embedding AI into their devices this year.

Strategic Hardware Adoption

As hardware choices proliferate, good governance will be crucial. Enterprises need to answer key questions:

? How many employees need next-generation devices?

? Which areas of the business will benefit most from these advancements?

Chip manufacturers are racing to improve AI horsepower, but enterprises can’t afford to refresh their entire edge footprint with every new advancement.

Instead, businesses should adopt a tiered strategy to ensure these devices are deployed where they can have the greatest impact.

大型科技公司推動(dòng)AI硬件需求

大型科技公司正成為推動(dòng)AI硬件需求的一部分，它們可能自建AI模型并部署專用芯片到本地。但事實(shí)上，各行各業(yè)的企業(yè)都在尋求更強(qiáng)的計(jì)算能力來(lái)實(shí)現(xiàn)它們的IT目標(biāo)。

例如：

根據(jù)Databricks的一份報(bào)告，在運(yùn)行大語(yǔ)言模型（LLMs）以處理欺詐檢測(cè)和財(cái)富管理任務(wù)時(shí)，金融服務(wù)業(yè)的GPU使用量在過(guò)去六個(gè)月內(nèi)增長(zhǎng)了88%，是增長(zhǎng)最快的行業(yè)之一。

GPU需求超出供給：新的“淘金熱”

所有這些對(duì)GPU的需求已經(jīng)遠(yuǎn)遠(yuǎn)超出了產(chǎn)能。在當(dāng)今這個(gè)“新淘金熱”中，那些提供“鎬和鏟”的公司，也就是為技術(shù)轉(zhuǎn)型提供工具的企業(yè)，正在贏得大筆回報(bào)。

NVIDIA（英偉達(dá)）首席執(zhí)行官黃仁勛表示，云GPU的容量幾乎已經(jīng)用盡。不過(guò)，NVIDIA正在推出新一代的芯片，其能源效率顯著高于以往版本。

云計(jì)算巨頭（Hyperscalers）正在以驚人的速度購(gòu)買(mǎi)剛剛下線的GPU，投資接近1萬(wàn)億美元用于數(shù)據(jù)中心基礎(chǔ)設(shè)施，以滿足客戶對(duì)GPU使用的租賃需求。同時(shí)，現(xiàn)有數(shù)據(jù)中心的能源消耗也在將全球老舊的電網(wǎng)推向極限。

新一代芯片：神經(jīng)處理單元（NPUs）

面對(duì)GPU需求激增的壓力，企業(yè)正在尋找新的解決方案。盡管GPU對(duì)于處理LLMs或內(nèi)容生成的高工作負(fù)載至關(guān)重要，而CPU依然是基本配置，但**神經(jīng)處理單元（NPUs）**正在迅速成為新熱點(diǎn)。

NPUs模擬大腦的神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)，可以以更高的效率和更低的功耗加速較小的AI工作負(fù)載。它們的優(yōu)勢(shì)在于：

? 讓AI應(yīng)用從云端轉(zhuǎn)移到本地運(yùn)行

? 保護(hù)敏感數(shù)據(jù)，避免托管在外部平臺(tái)上

這類新型芯片是未來(lái)嵌入式AI的重要組成部分。

戴爾科技（Dell Technologies）戰(zhàn)略高級(jí)副總裁Vivek Mohindra表示：

“目前全球有15億臺(tái)PC，其中30%超過(guò)4年機(jī)齡。這些老舊PC都沒(méi)有NPUs，無(wú)法利用最新的AI PC功能。”

企業(yè)硬件可能迎來(lái)一次大規(guī)模的升級(jí)浪潮。隨著NPUs讓終端設(shè)備可以離線運(yùn)行AI，同時(shí)讓AI模型更小、更貼合具體用例，硬件可能再次成為企業(yè)性能的差異化優(yōu)勢(shì)。

AI的變革潛力

根據(jù)德勤的一項(xiàng)研究：

72%的受訪者認(rèn)為生成式AI對(duì)其所在行業(yè)的影響將是“重大到變革性”。

隨著硬件的進(jìn)一步普及，讓AI觸手可及，這一比例可能會(huì)接近100%。

新趨勢(shì)：企業(yè)基礎(chǔ)設(shè)施重回戰(zhàn)略核心

曾經(jīng)，云計(jì)算給人一種“資源無(wú)限”的印象，但如今，我們正進(jìn)入一個(gè)資源受限的時(shí)代。

在過(guò)去幾年里，企業(yè)基礎(chǔ)設(shè)施（例如PC）被視為一種“工具性資源”，但現(xiàn)在，它們?cè)俅纬蔀閼?zhàn)略重點(diǎn)。

特別是，專用硬件在以下三個(gè)AI增長(zhǎng)領(lǐng)域?qū)⒂葹橹匾?/p>

1. 嵌入AI的設(shè)備與物聯(lián)網(wǎng)（IoT）

2. 數(shù)據(jù)中心

3. 先進(jìn)的物理機(jī)器人

盡管機(jī)器人領(lǐng)域的影響可能在未來(lái)幾年才會(huì)顯現(xiàn)，但企業(yè)在未來(lái)18到24個(gè)月內(nèi)需要著手應(yīng)對(duì)前兩個(gè)領(lǐng)域的相關(guān)決策。

1. 邊緣計(jì)算的崛起

到2025年，超過(guò)50%的數(shù)據(jù)可能會(huì)由邊緣設(shè)備生成。

隨著NPUs的普及，更多設(shè)備將能夠運(yùn)行AI模型而無(wú)需依賴云計(jì)算。尤其是，生成式AI的提供商正在開(kāi)發(fā)更小、更高效的模型，針對(duì)具體任務(wù)提供支持。

邊緣計(jì)算的優(yōu)勢(shì)：

? 更快的響應(yīng)時(shí)間

? 更低的成本

? 更強(qiáng)的隱私控制

混合計(jì)算（即云端與設(shè)備端AI工作負(fù)載相結(jié)合）可能成為許多企業(yè)的必備選項(xiàng)，而硬件制造商正在押注這一趨勢(shì)。

戴爾科技的Mohindra表示：

“從延遲、網(wǎng)絡(luò)資源以及數(shù)據(jù)量來(lái)看，將數(shù)據(jù)移到集中計(jì)算位置既低效又不安全。將AI帶到數(shù)據(jù)前線，而不是把數(shù)據(jù)送到AI前線，是更好的選擇?！?/p>

2. 硬件的升級(jí)浪潮即將到來(lái)

一家大型銀行預(yù)測(cè)，到2026年，AI PC將占PC出貨量的40%以上。

同時(shí)，預(yù)計(jì)到2024年，近15%的智能手機(jī)將能夠運(yùn)行LLMs或圖像生成模型。

HP AI PC體驗(yàn)與云端客戶高級(jí)總監(jiān)Alex Thatcher表示：

“這次設(shè)備升級(jí)浪潮就像90年代從命令行輸入到圖形用戶界面的轉(zhuǎn)型一樣重大。軟件已經(jīng)發(fā)生了根本性的變化，帶來(lái)了全新的工具和協(xié)作方式。企業(yè)需要能夠加速這種變化的硬件，以便更輕松地創(chuàng)建和交付AI解決方案。”

蘋(píng)果和微軟今年也通過(guò)將AI嵌入到它們的設(shè)備中，推動(dòng)了即將到來(lái)的硬件升級(jí)潮。

企業(yè)的戰(zhàn)略硬件應(yīng)用

隨著硬件選擇的增多，良好的治理將至關(guān)重要。企業(yè)需要問(wèn)自己：

? 我們的員工中有多少人需要下一代設(shè)備？

? 哪些業(yè)務(wù)領(lǐng)域最需要這些硬件的支持？

盡管芯片制造商正在競(jìng)相提升AI的算力，但企業(yè)無(wú)法在每次新技術(shù)發(fā)布時(shí)對(duì)所有設(shè)備進(jìn)行全面升級(jí)。

相反，企業(yè)應(yīng)該采取分級(jí)戰(zhàn)略，確保這些設(shè)備能夠部署在最需要的地方，以實(shí)現(xiàn)最大的影響。

IT, amplified: AI elevates the reach (and remit) of the tech function

As the tech function shifts from leading digital transformation to leading AI transformation, forward-thinking leaders are using this as an opportunity to redefine the future of IT.

Much has been said, including within the pages of Tech Trends, about the potential for artificial intelligence to revolutionize business use cases and outcomes. Nowhere is this more true than in the end-to-end life cycle of software engineering and the broader business of information technology, given generative AI’s ability to write code, test software, and augment tech talent in general.

Deloitte research has shown that tech companies at the forefront of this organizational change are ready to realize the benefits: They are twice as likely as their more conservative peers to say generative AI is transforming their organization now or will within the next year.

We wrote in a Tech Trends 2024 article that enterprises need to reorganize their developer experiences to help IT teams achieve the best results. Now, the AI hype cycle has placed an even greater focus on the tech function’s ways of working. IT has long been the lighthouse of digital transformation in the enterprise, but it must now take on AI transformation. Forward-thinking IT leaders are using the current moment as a once-in-a-generation opportunity to redefine roles and responsibilities, set investment priorities, and communicate value expectations.

More importantly, by playing this pioneering role, chief information officers can help inspire other technology leaders to put AI transformation into practice.

After years of enterprises pursuing lean IT and everything-as-a-service offerings, AI is sparking a shift away from virtualization and austere budgets. Gartner predicts that “worldwide IT spending is expected to total $5.26 trillion in 2024, an increase of 7.5% from 2023.”

As we discuss in “Hardware is eating the world,” hardware and infrastructure are having a moment, and enterprise IT spending and operations may shift accordingly. As both traditional AI and generative AI become more capable and ubiquitous, each of the phases of tech delivery may see a shift from human in charge to human in the loop. Organizations need a clear strategy in place before that occurs.

Based on Deloitte analysis, over the next 18 to 24 months, IT leaders should plan for AI transformation across five key pillars:

1. Engineering

2. Talent

3. Cloud financial operations (FinOps)

4. Infrastructure

5. Cyber risk

This trend may usher in a new type of lean IT

If commercial functions see an increased number of citizen developers or digital agents that can spin up applications on a whim, the role of the IT function may shift from building and maintaining to orchestrating and innovating.

In that case, AI may not only be undercover, as we indicate in the introduction to this year’s report, but may also be overtly in the boardroom, overseeing tech operations in line with human needs.

Now: Spotlight—and higher spending—on IT

For years, IT has been under pressure to streamline sprawling cloud spend and curb costs. Since 2020, however, investments in tech have been on the rise thanks to pent-up demand for collaboration tools and the pandemic-era emphasis on digitalization.

According to Deloitte research:

? From 2020 to 2022, the global average technology budget as a percentage of revenue jumped from 4.25% to 5.49%, an increase that approximately doubled the previous revenue change from 2018 to 2020.

? In 2024, US companies’ average budget for digital transformation as a percentage of revenue is 7.5%, with 5.4% coming from the IT budget.

As demand for AI sparks another increase in spending, the finding from Deloitte’s 2023 Global Technology Leadership Study continues to ring true: Technology is the business, and tech spend is increasing as a result.

Today, enterprises are grappling with the new relevance of hardware, data management, and digitization in ramping up their usage of AI and realizing its value potential.

In Deloitte’s Q2 State of Generative AI in the Enterprise report, businesses that rated themselves as having “very high” levels of expertise in generative AI were increasing their investment in hardware and cloud consumption much more than the average enterprise.

Overall, 75% of organizations surveyed have increased their investments around data-life-cycle management due to generative AI.

Tech investment strategies are critical

These figures point to a common theme: To realize the highest impact from gen AI, enterprises likely need to accelerate their cloud and data modernization efforts.

AI has the potential to deliver efficiencies in cost, innovation, and a host of other areas, but the first step to accruing these benefits is for businesses to focus on making the right tech investments.

Because of these crucial investment strategies, the spotlight is on tech leaders who are paving the way.

According to Deloitte research:

? Over 60% of US-based technology leaders now report directly to their chief executives, an increase of more than 10 percentage points since 2020.

This is a testament to the tech leader’s increased importance in setting the AI strategy rather than simply enabling it.

Far from a cost center, IT is increasingly being seen as a differentiator in the AI age, as CEOs, following market trends, are keen on staying abreast of AI’s adoption in their enterprise.

The future of IT: Leaner, more integrated, and faster

John Marcante, former global CIO of Vanguard and US CIO-in-residence at Deloitte, believes AI will fundamentally change the role of IT.

He says:

“The technology organization will be leaner, but have a wider purview. It will be more integrated with the business than ever. AI is moving fast, and centralization is a good way to ensure organizational speed and focus.”

IT is gearing up for transformation

As IT gears up for the opportunity presented by AI—perhaps the opportunity that many tech leaders and employees have waited for—changes are already underway in how the technology function organizes itself and executes work.

The stakes are high, and IT is due for a makeover.

四、IT能力大升級(jí)

隨著技術(shù)職能從引領(lǐng)數(shù)字化轉(zhuǎn)型轉(zhuǎn)向引領(lǐng)AI轉(zhuǎn)型，前瞻性的領(lǐng)導(dǎo)者正在利用這一機(jī)會(huì)重新定義IT的未來(lái)。

AI對(duì)IT的全面影響：軟件工程與技術(shù)職能

關(guān)于人工智能如何徹底改變業(yè)務(wù)場(chǎng)景和結(jié)果，業(yè)界已有許多討論?！犊萍稼厔?shì)》多次提到這一點(diǎn)，而在軟件工程全生命周期和信息技術(shù)業(yè)務(wù)中，這一點(diǎn)尤為真實(shí)。

生成式AI能夠編寫(xiě)代碼、測(cè)試軟件并全面增強(qiáng)技術(shù)團(tuán)隊(duì)的能力，這些優(yōu)勢(shì)正在改變IT的工作方式。

根據(jù)德勤的研究，走在這一組織變革前沿的科技公司，已經(jīng)準(zhǔn)備好享受這一紅利：

它們比更保守的同行企業(yè)更有可能表示生成式AI正在或即將在一年內(nèi)改變其組織。

我們?cè)凇犊萍稼厔?shì)2024》中提到，企業(yè)需要重新組織開(kāi)發(fā)者的工作體驗(yàn)，幫助IT團(tuán)隊(duì)取得更好的成果。如今，AI的熱潮更進(jìn)一步，將焦點(diǎn)聚集在IT職能的工作方式上。

IT長(zhǎng)期以來(lái)一直是企業(yè)數(shù)字化轉(zhuǎn)型的燈塔，但現(xiàn)在，它必須承擔(dān)起AI轉(zhuǎn)型的責(zé)任。

前瞻性的IT領(lǐng)導(dǎo)者正在將這一時(shí)刻視為百年難得的機(jī)會(huì)，通過(guò)重新定義角色與職責(zé)、設(shè)定投資優(yōu)先級(jí)和傳遞價(jià)值預(yù)期，全面推動(dòng)組織變革。更重要的是，通過(guò)扮演這一先鋒角色，**首席信息官（CIO）**可以激勵(lì)其他技術(shù)領(lǐng)導(dǎo)者將AI轉(zhuǎn)型付諸實(shí)踐。

AI時(shí)代的技術(shù)支出趨勢(shì)

在企業(yè)長(zhǎng)期追求精益IT和一切服務(wù)化的背景下，AI正在引發(fā)一場(chǎng)從虛擬化和縮減預(yù)算向新投資方向的轉(zhuǎn)變。

Gartner預(yù)測(cè)，到2024年，全球IT支出將達(dá)到5.26萬(wàn)億美元，比2023年增長(zhǎng)7.5%。

正如我們?cè)凇队布谕淌墒澜纭分杏懻摰模布突A(chǔ)設(shè)施正成為焦點(diǎn)，企業(yè)的IT支出和運(yùn)營(yíng)可能因此發(fā)生相應(yīng)變化。隨著傳統(tǒng)AI和生成式AI變得更加強(qiáng)大和普及，技術(shù)交付的每個(gè)階段可能從“以人為主導(dǎo)”逐步轉(zhuǎn)向“人類參與其中（Human in the Loop）”。

企業(yè)需要在這種轉(zhuǎn)變發(fā)生之前制定清晰的戰(zhàn)略。根據(jù)德勤的分析，未來(lái)18到24個(gè)月內(nèi)，IT領(lǐng)導(dǎo)者應(yīng)圍繞以下五大核心支柱制定AI轉(zhuǎn)型計(jì)劃：

1. 工程

2. 人才

3. 云財(cái)務(wù)運(yùn)營(yíng)（FinOps）

4. 基礎(chǔ)設(shè)施

5. 網(wǎng)絡(luò)風(fēng)險(xiǎn)

未來(lái)IT：從“建設(shè)者”到“創(chuàng)新者”

這場(chǎng)趨勢(shì)可能在未來(lái)十年催生一種新的精益IT模式。如果企業(yè)的商業(yè)職能中出現(xiàn)更多“公民開(kāi)發(fā)者”或能夠隨時(shí)生成應(yīng)用的數(shù)字代理，那么IT職能的角色可能會(huì)從構(gòu)建與維護(hù)轉(zhuǎn)變?yōu)閰f(xié)調(diào)與創(chuàng)新。

這種情況下，AI不僅僅是隱藏在后臺(tái)的助推器，甚至可能直接參與到董事會(huì)層面的戰(zhàn)略決策中，與人類需求保持一致，監(jiān)督技術(shù)運(yùn)營(yíng)。

IT支出的聚光燈下

多年來(lái)，IT一直承受著控制云支出的壓力。然而，自2020年以來(lái)，受疫情期間對(duì)協(xié)作工具的需求激增和數(shù)字化轉(zhuǎn)型的推動(dòng)，技術(shù)投資呈現(xiàn)上升趨勢(shì)。

數(shù)據(jù)統(tǒng)計(jì)：

1、從2020年到2022年，全球企業(yè)的技術(shù)預(yù)算占收入比例從4.25%躍升至5.49%。

2、到2024年，美國(guó)企業(yè)的數(shù)字化轉(zhuǎn)型預(yù)算占收入的7.5%，其中5.4%來(lái)自IT預(yù)算。

隨著AI需求帶來(lái)新一輪支出增長(zhǎng)，德勤2023年的《全球技術(shù)領(lǐng)導(dǎo)力研究》中提到的觀點(diǎn)依然成立：技術(shù)就是業(yè)務(wù)，因此技術(shù)支出也在不斷增加。

企業(yè)正在應(yīng)對(duì)硬件需求、數(shù)據(jù)管理和數(shù)字化的新相關(guān)性，以加速AI的應(yīng)用并釋放其價(jià)值潛力。根據(jù)德勤Q2生成式AI報(bào)告，認(rèn)為自己在生成式AI方面具有“非常高”專業(yè)水平的企業(yè)，在硬件和云消費(fèi)方面的投資比平均水平高出許多。

AI驅(qū)動(dòng)的技術(shù)投資策略

75%的企業(yè)因生成式AI而增加了數(shù)據(jù)生命周期管理的投資。

這些數(shù)據(jù)指向一個(gè)共同主題：為了讓生成式AI發(fā)揮最大效用，企業(yè)需要加速云和數(shù)據(jù)現(xiàn)代化。AI有潛力在成本、創(chuàng)新和其他多個(gè)領(lǐng)域帶來(lái)高效益，但前提是企業(yè)必須專注于正確的技術(shù)投資策略。

由于這些關(guān)鍵的投資策略，技術(shù)領(lǐng)導(dǎo)者成為了關(guān)注的焦點(diǎn)。

根據(jù)德勤的研究，超過(guò)60%的美國(guó)技術(shù)領(lǐng)導(dǎo)者現(xiàn)在直接向首席執(zhí)行官匯報(bào)，比2020年增加了10個(gè)百分點(diǎn)。這反映出技術(shù)領(lǐng)導(dǎo)者在制定AI戰(zhàn)略中的重要性，已從單純的技術(shù)支持角色轉(zhuǎn)變?yōu)閼?zhàn)略制定者。

IT不再只是成本中心，而是AI時(shí)代的差異化優(yōu)勢(shì)，CEO們正密切關(guān)注AI在企業(yè)中的應(yīng)用，以保持領(lǐng)先地位。

IT的未來(lái)：更精益、更融合、更快速

Vanguard（先鋒集團(tuán)）前全球CIO兼德勤美國(guó)駐地CIO John Marcante認(rèn)為，AI將從根本上改變IT的角色。他說(shuō)：“技術(shù)團(tuán)隊(duì)會(huì)變得更精簡(jiǎn)，但覆蓋范圍更廣。它將與業(yè)務(wù)的融合程度比以往任何時(shí)候都高。AI發(fā)展速度很快，而集中化是確保組織速度與專注的最佳方式?！?/p>

IT的變革時(shí)刻已經(jīng)到來(lái)

隨著IT為AI帶來(lái)的機(jī)遇做好準(zhǔn)備，技術(shù)職能的組織方式和執(zhí)行方式正在發(fā)生改變。這可能正是許多技術(shù)領(lǐng)導(dǎo)者和員工一直等待的機(jī)會(huì)。

但這場(chǎng)變革的代價(jià)很高，IT也即將迎來(lái)一場(chǎng)全面的“改頭換面”。

New: An AI boost for IT

Over the next 18 to 24 months, the nature of the IT function is likely to change as enterprises increasingly employ generative AI. Deloitte’s foresight analysis suggests that, by 2027, even in the most conservative scenario, gen AI will be embedded into every company’s digital product or software footprint (figure 1), as we discuss across five key pillars.

Engineering

In the traditional software development life cycle, manual testing, inexperienced developers, and disparate tool environments can lead to inefficiencies, as we’ve discussed in prior Tech Trends. Fortunately, AI is already having an impact on these areas. AI-assisted code generation, automated testing, and rapid data analytics all save developers more time for innovation and feature development. The productivity gain from coding alone is estimated to be worth US$12 billion in the United States alone.

At Google, AI tools are being rolled out internally to developers. In a recent earnings call, CEO Sundar Pichai said that around 25 percent of the new code at the technology giant is developed using AI. Shivani Govil, senior director of product management for developer products, believes that “AI can transform how engineering teams work, leading to more capacity to innovate, less toil, and higher developer satisfaction. Google’s approach is to bring AI to our users and meet them where they are—by bringing the technology into products and tools that developers use every day to support them in their work. Over time, we can create even tighter alignment between the code and business requirements, allowing faster feedback loops, improved product market fit, and better alignment to the business outcomes.”

In another example, a health care company used COBOL code assist to enable a junior developer with no experience in the programming language to generate an explanation file with 95% accuracy.

As Deloitte recently stated in a piece on engineering in the age of gen AI, the developer role is likely to shift from writing code to defining the architecture, reviewing code, and orchestrating functionality through contextualized prompt engineering. Tech leaders should anticipate human-in-the-loop code generation and review to be the standard over the next few years of AI adoption.

新趨勢(shì)：AI為IT注入動(dòng)力

未來(lái)18到24個(gè)月，隨著企業(yè)對(duì)生成式AI的日益采用，IT職能的性質(zhì)可能會(huì)發(fā)生巨大變化。根據(jù)德勤的前瞻分析，到2027年，即使是在最保守的情景下，生成式AI也將嵌入每家企業(yè)的數(shù)字產(chǎn)品或軟件體系中（如圖1所示）。以下是AI將在五大核心支柱中的具體影響。

1. 工程（Engineering）

在傳統(tǒng)的軟件開(kāi)發(fā)生命周期中，手動(dòng)測(cè)試、缺乏經(jīng)驗(yàn)的開(kāi)發(fā)者以及分散的工具環(huán)境往往會(huì)導(dǎo)致效率低下。這些問(wèn)題已在我們之前的《科技趨勢(shì)》中討論過(guò)。而現(xiàn)在，AI正在這些領(lǐng)域產(chǎn)生積極的影響。

AI助力的功能：

? 代碼生成

? 自動(dòng)化測(cè)試

? 快速數(shù)據(jù)分析

這些能力幫助開(kāi)發(fā)者節(jié)省時(shí)間，從而將更多精力投入到創(chuàng)新和功能開(kāi)發(fā)中。據(jù)估計(jì)，僅代碼編寫(xiě)效率的提升在美國(guó)的生產(chǎn)力收益就高達(dá)120億美元。

谷歌案例：

谷歌正在內(nèi)部向開(kāi)發(fā)人員推出AI工具。該公司CEO桑達(dá)爾·皮查伊在近期的財(cái)報(bào)電話會(huì)議中提到，大約25%的新代碼是通過(guò)AI開(kāi)發(fā)的。

谷歌開(kāi)發(fā)者產(chǎn)品高級(jí)總監(jiān)Shivani Govil表示：

“AI可以徹底改變工程團(tuán)隊(duì)的工作方式，提高創(chuàng)新能力、減少重復(fù)性勞動(dòng)并提升開(kāi)發(fā)者滿意度。谷歌的做法是將AI技術(shù)融入開(kāi)發(fā)者每天使用的產(chǎn)品和工具中，以支持他們的工作。隨著時(shí)間的推移，我們可以實(shí)現(xiàn)代碼與業(yè)務(wù)需求之間更緊密的對(duì)齊，從而加速反饋循環(huán)、改善產(chǎn)品與市場(chǎng)的契合度，并更好地支持業(yè)務(wù)目標(biāo)?！?/p>

AI提升的真實(shí)場(chǎng)景：

? 一家醫(yī)療公司通過(guò)AI支持的COBOL代碼助手，幫助一位沒(méi)有COBOL經(jīng)驗(yàn)的初級(jí)開(kāi)發(fā)者生成了準(zhǔn)確率高達(dá)**95%**的解釋文件。

開(kāi)發(fā)者角色的轉(zhuǎn)變

德勤在最近一篇關(guān)于生成式AI時(shí)代工程開(kāi)發(fā)的文章中指出，開(kāi)發(fā)者的角色正從“編寫(xiě)代碼”轉(zhuǎn)向“定義架構(gòu)、審查代碼并通過(guò)上下文化的提示工程整合功能”。

技術(shù)領(lǐng)導(dǎo)者應(yīng)預(yù)見(jiàn)到，人類參與的代碼生成與審查將在未來(lái)幾年成為AI應(yīng)用的行業(yè)標(biāo)準(zhǔn)。

Technology executives surveyed by Deloitte last year noted that they struggle to hire workers with critical IT backgrounds in security, machine learning, and software architecture, and are forced to delay projects with financial backing due to a shortage of appropriately skilled talent. As AI becomes the newest skill in demand, many companies may not even be able to find all the talent they need, leading to a hiring gap wherein nearly 50% of AI-related positions cannot be filled.

As a result, tech leaders should focus on upskilling their own talent, another area where AI can help. Consider the potential benefits of:

? AI-powered skills gap analyses and recommendations

? Personalized learning paths

? Virtual tutors for on-demand learning

For example, Bayer, the life sciences company, has used generative AI to summarize procedural documents and generate rich media such as animations for e-learning. Similarly, AI could generate documentation to help a new developer understand legacy technology and then create an associated learning podcast and exam for the same developer.

At Google, developers thrive on hands-on experience and problem-solving, so leaders are keen to provide AI learning and tools (like coding assistants) that meet developers where they are on their learning journey. “We can use AI to enhance learning, in context with emerging technologies, in ways that anticipate and support the rapidly changing skills and knowledge required to adapt to them,” says Sara Ortloff, senior director of developer experience at Google.

As automation increases, tech talent would take an oversight role and enjoy more capacity to focus on innovation that can improve the bottom line. This could help attract talent since, according to Deloitte research, the biggest incentive that attracts tech talent to new opportunities is the work they would do in the role.

Cloud Financial Operations

Runaway spending became a common problem in the cloud era when resources could be provisioned with a click. Hyperscalers have offered data and tooling for finance teams and CIOs to keep better track of their team’s cloud usage, but many of these FinOps tools still require manual budgeting and offer limited visibility across disparate systems.

The power of AI enables organizations to be more informed, proactive, and effective with their financial management. For example:

? Real-time cost analysis

? Robust pattern detection

? Resource allocation across systems

AI can help enterprises identify more cost-saving opportunities through better predictions and tracking.

As AI demand increases in the coming years, enterprises are likely to see higher cloud costs. However, applying AI to FinOps can justify the investments in AI and optimize costs elsewhere.

Infrastructure

Across the broad scope of IT infrastructure—from toolchains to service management—organizations haven’t seen as much automation as they’d like. Just a few years ago, studies estimated that nearly half of large enterprises were handling key tasks like security, compliance, and service management manually.

The missing ingredient?

Automation that can learn, improve, and react to the changing demands of a business.

Now, this is becoming possible. Automated resource allocation, predictive maintenance, and anomaly detection could all be implemented in systems that natively understand their own real-time status and can take action accordingly.

This emerging view of IT is referred to as “autonomic” IT, inspired by the autonomic nervous system in the human body that adjusts dynamically to internal and external stimuli. In such a system, infrastructure takes care of itself, surfacing only issues that require human intervention.

For instance:

? eBay is already leveraging generative AI to scale infrastructure and analyze massive amounts of customer data, enabling impactful changes to its platform.

Cybersecurity

Although AI simplifies and enhances many IT processes, it also introduces greater complexity in cyber risks. As we discussed last year, generative AI and synthetic media open up new attack surfaces, including:

? Phishing

? Deepfakes

? Prompt injection attacks

As AI proliferates and digital agents become the newest B2B representatives, these risks may worsen.

How enterprises can respond:

? Data authentication: For example, SWEAR, a security company, has pioneered a way to verify digital media using blockchain.

? Data masking

? Incident response

? Automated policy generation

Generative AI can optimize cybersecurity responses and strengthen defenses against attacks.

Rethinking IT Resources

As technology teams adapt to these changes and challenges, many will shift their focus to innovation, agility, and growth enabled by AI. Teams can:

? Streamline IT workflows

? Reduce the need for manual intervention or offshoring

? Focus on higher-value activities

This could lead to a reallocation of IT resources across the board.

As Ian Cairns, CEO of Freeplay, notes:

“As with any major platform shift, the businesses that succeed will be the ones that can rethink and adapt how they work and build software for a new era.”

2. 人才（Talent）

根據(jù)德勤去年對(duì)技術(shù)高管的調(diào)查，許多企業(yè)在招聘具有關(guān)鍵IT背景（如安全、機(jī)器學(xué)習(xí)和軟件架構(gòu)）的人才時(shí)面臨困難。由于缺乏具備適當(dāng)技能的人才，它們不得不推遲一些已經(jīng)獲得資金支持的項(xiàng)目。隨著AI成為最新的熱門(mén)技能，許多公司可能根本找不到所需的全部人才，導(dǎo)致招聘缺口進(jìn)一步擴(kuò)大，目前約有50%的AI相關(guān)崗位無(wú)法填補(bǔ)。

因此，技術(shù)領(lǐng)導(dǎo)者需要將重點(diǎn)放在提升現(xiàn)有團(tuán)隊(duì)的技能上，而這恰好是AI可以發(fā)揮作用的領(lǐng)域之一?？梢韵胂笠韵翧I支持的能力：

? AI驅(qū)動(dòng)的技能差距分析與建議

? 個(gè)性化學(xué)習(xí)路徑

? 按需學(xué)習(xí)的虛擬導(dǎo)師

生命科學(xué)公司拜耳（Bayer）利用生成式AI總結(jié)程序文檔，并生成動(dòng)畫(huà)等豐富的媒體用于電子學(xué)習(xí)。同樣，AI還可以生成文檔，幫助新開(kāi)發(fā)者理解舊系統(tǒng)技術(shù)，并為其生成相關(guān)的學(xué)習(xí)播客和考試內(nèi)容。

在谷歌，開(kāi)發(fā)者依靠實(shí)際操作經(jīng)驗(yàn)和解決問(wèn)題來(lái)成長(zhǎng)，因此公司領(lǐng)導(dǎo)者特別注重提供AI學(xué)習(xí)資源和工具（如代碼助手），以滿足開(kāi)發(fā)者當(dāng)前學(xué)習(xí)階段的需求。谷歌開(kāi)發(fā)者體驗(yàn)高級(jí)總監(jiān)Sara Ortloff表示：

“我們可以通過(guò)AI提升學(xué)習(xí)能力，將其與新興技術(shù)的上下文結(jié)合起來(lái)，預(yù)見(jiàn)并支持不斷變化的技能需求，幫助開(kāi)發(fā)者適應(yīng)這些變化?！?/p>

隨著自動(dòng)化的增加，技術(shù)人才將更多承擔(dān)監(jiān)督角色，同時(shí)有更多的時(shí)間專注于推動(dòng)創(chuàng)新，為企業(yè)帶來(lái)切實(shí)的收益。這種變化還能吸引人才——根據(jù)德勤的研究，技術(shù)崗位吸引人才的最大因素是崗位本身的工作內(nèi)容。

3. 云財(cái)務(wù)運(yùn)營(yíng)（Cloud Financial Operations）

在云計(jì)算時(shí)代，由于資源可以隨手點(diǎn)擊部署，過(guò)度支出已成為常見(jiàn)問(wèn)題。雖然云服務(wù)商（Hyperscalers）已經(jīng)為財(cái)務(wù)團(tuán)隊(duì)和CIO提供了工具以更好地跟蹤云使用情況，但許多FinOps工具仍需要手動(dòng)預(yù)算，且在跨系統(tǒng)之間的可見(jiàn)性方面存在限制。

AI的加入可以讓企業(yè)在財(cái)務(wù)管理上更加信息透明、主動(dòng)出擊、高效管理。比如：

? 實(shí)時(shí)成本分析

? 強(qiáng)大的模式檢測(cè)

? 跨系統(tǒng)的資源分配

AI還能通過(guò)更好的預(yù)測(cè)和跟蹤，幫助企業(yè)發(fā)現(xiàn)更多節(jié)約成本的機(jī)會(huì)。

隨著未來(lái)幾年AI需求的持續(xù)增長(zhǎng)，大型企業(yè)可能面臨云成本顯著上升的情況。然而，通過(guò)將AI應(yīng)用于FinOps，不僅可以為AI投資正名，還能在其他領(lǐng)域優(yōu)化成本。

4. 基礎(chǔ)設(shè)施（Infrastructure）

在廣泛的IT基礎(chǔ)設(shè)施領(lǐng)域——從工具鏈到服務(wù)管理，企業(yè)自動(dòng)化程度仍遠(yuǎn)低于預(yù)期。幾年前的研究表明，近一半的大型企業(yè)仍在手動(dòng)處理安全、合規(guī)和服務(wù)管理等關(guān)鍵任務(wù)。

缺少的關(guān)鍵要素是什么？

能夠?qū)W習(xí)、改進(jìn)并響應(yīng)企業(yè)需求變化的自動(dòng)化。

如今，這種能力正在成為現(xiàn)實(shí)。

比如：

? 自動(dòng)化的資源分配

? 預(yù)測(cè)性維護(hù)

? 異常檢測(cè)

這些功能可以通過(guò)一個(gè)實(shí)時(shí)感知自身狀態(tài)并采取行動(dòng)的系統(tǒng)實(shí)現(xiàn)。這種新興的IT概念被稱為**“自主IT”**，靈感來(lái)自人體的自主神經(jīng)系統(tǒng)，它能動(dòng)態(tài)調(diào)整心率和呼吸以適應(yīng)內(nèi)外部刺激。

4. 自主IT的優(yōu)勢(shì)：

? 讓基礎(chǔ)設(shè)施自行運(yùn)行，只在需要人工干預(yù)時(shí)提出問(wèn)題。

? eBay已利用生成式AI擴(kuò)展其基礎(chǔ)設(shè)施，并分析海量客戶數(shù)據(jù)，從而對(duì)其平臺(tái)進(jìn)行重要改進(jìn)。

5. 網(wǎng)絡(luò)安全（Cybersecurity）

雖然AI讓許多IT流程變得更加簡(jiǎn)單高效，但也帶來(lái)了更高的網(wǎng)絡(luò)風(fēng)險(xiǎn)復(fù)雜性。正如我們?nèi)ツ晏岬降?，生成式AI和合成媒體為網(wǎng)絡(luò)攻擊打開(kāi)了新的入口，包括：

? 釣魚(yú)攻擊

? 深度偽造（Deepfakes）

? 提示注入攻擊

隨著AI的普及，以及數(shù)字代理成為最新的B2B代表，這些風(fēng)險(xiǎn)可能會(huì)更加嚴(yán)重。

企業(yè)應(yīng)如何應(yīng)對(duì)？

? 數(shù)據(jù)認(rèn)證：例如，安全公司SWEAR通過(guò)區(qū)塊鏈驗(yàn)證數(shù)字媒體的真實(shí)性。

? 數(shù)據(jù)掩碼

? 事件響應(yīng)

? 自動(dòng)化策略生成

生成式AI還可以優(yōu)化網(wǎng)絡(luò)安全響應(yīng)，加強(qiáng)對(duì)攻擊的防御能力。

重新思考IT資源分配

隨著技術(shù)團(tuán)隊(duì)逐步適應(yīng)上述變化和挑戰(zhàn)，許多團(tuán)隊(duì)將把重點(diǎn)轉(zhuǎn)向由AI驅(qū)動(dòng)的創(chuàng)新、敏捷性和增長(zhǎng)。

團(tuán)隊(duì)可以：

? 簡(jiǎn)化IT工作流程

? 減少對(duì)手動(dòng)干預(yù)或外包的依賴

? 專注于高價(jià)值活動(dòng)

這可能會(huì)導(dǎo)致IT資源的全面重新分配。

正如Freeplay公司CEO Ian Cairns所說(shuō)：

“與任何重大平臺(tái)轉(zhuǎn)變一樣，能夠重新思考和適應(yīng)工作方式及軟件開(kāi)發(fā)模式的企業(yè)，將在這一新紀(jì)元中勝出。”

The new math: Solving cryptography in an age of quantum

Quantum computers are likely to pose a severe threat to today’s encryption practices. Updating encryption has never been more urgent.

Cybersecurity professionals already have a lot on their minds. From run-of-the-mill social engineering hacks to emerging threats from AI-generated content, there’s no shortage of immediate concerns. But while focusing on the urgent, they could be overlooking an important threat vector: the potential risk that a cryptographically relevant quantum computer (CRQC) will someday be able to break much of the current public-key cryptography that businesses rely upon. Once that cryptography is broken, it will undermine the processes that establish online sessions, verify transactions, and assure user identity.

Let’s contrast this risk with the historical response to Y2K, where businesses saw a looming risk and addressed it over time, working backward from a specific time to avert a more significant impact.1 The potential risk of a CRQC is essentially the inverse case: The effect is expected to be even more sweeping, but the date at which such a cryptographically relevant quantum computer will become available is unknown. Preparing for CRQCs is generally acknowledged to be highly important but is often low on the urgency scale because of the unknown timescale. This has created a tendency for organizations to defer the activities necessary to prepare their cybersecurity posture for the arrival of quantum computers.

“Unless it’s here, people are saying, ‘Yeah, we’ll get to it, or the vendors will do it for me. I have too many things to do and too little budget,’” says Mike Redding, chief technology officer at cybersecurity company Quantropi.2 “Quantum may be the most important thing ever, but it doesn’t feel urgent to most people. They’re just kicking the can down the road.”

This complacent mindset could breed disaster because the question isn’t if quantum computers are coming—it’s when. Most experts consider the exact time horizon for the advent of a CRQC to be irrelevant when it comes to encryption. The consensus is that one will likely emerge in the next five to 10 years, but how long will it take organizations to update their infrastructures and third-party dependencies? Eight years? Ten years? Twelve?

Given how long it took to complete prior cryptographic upgrades, such as migrating from cryptographic hashing algorithms SHA1 to SHA2, it is prudent to start now.

In a recent report, the US Office of Management and Budget said, “It is likely that a CRQC will be able to break some forms of cryptography that are now commonly used throughout government and the private sector. A CRQC is not yet known to exist; however, steady advancements in the quantum computing field may yield a CRQC in the coming decade. Accordingly … federal agencies must bolster the defense of their existing information systems by migrating to the use of quantum-resistant public-key cryptographic systems.”3

The scale of the problem is potentially massive, but fortunately, tools and expertise exist today to help enterprises address it. Recently released postquantum cryptography (PQC) algorithm standards from the US National Institute of Standards and Technology (NIST) could help to neutralize the problem before it becomes costly,? and many other governments around the world are also working on this issue.?

Furthermore, a reinvigorated cyber mindset could set enterprises on the road to better security.

五、量子計(jì)算

量子計(jì)算機(jī)可能會(huì)對(duì)當(dāng)前的加密實(shí)踐構(gòu)成嚴(yán)重威脅，更新加密技術(shù)已經(jīng)變得刻不容緩。

量子威脅的迫近

網(wǎng)絡(luò)安全專業(yè)人士已經(jīng)有許多問(wèn)題需要擔(dān)憂：從常見(jiàn)的社交工程攻擊到AI生成內(nèi)容帶來(lái)的新威脅，問(wèn)題層出不窮。然而，在應(yīng)對(duì)這些緊迫問(wèn)題的同時(shí)，他們可能忽略了一個(gè)重要的威脅：量子計(jì)算機(jī)對(duì)加密系統(tǒng)的潛在風(fēng)險(xiǎn)。一旦具備加密破解能力的量子計(jì)算機(jī)（CRQC）出現(xiàn)，可能會(huì)攻破目前廣泛依賴的公鑰加密技術(shù)。這將動(dòng)搖互聯(lián)網(wǎng)連接會(huì)話的建立、交易驗(yàn)證以及用戶身份驗(yàn)證等核心過(guò)程。

相比之下，可以將這種風(fēng)險(xiǎn)與歷史上的千年蟲(chóng)問(wèn)題（Y2K）應(yīng)對(duì)方式進(jìn)行對(duì)比。Y2K是一個(gè)明確的風(fēng)險(xiǎn)，企業(yè)從特定的時(shí)間節(jié)點(diǎn)倒推，采取了系統(tǒng)性行動(dòng)來(lái)避免更大的影響。而量子計(jì)算機(jī)的威脅卻正好相反：它的影響可能更為深遠(yuǎn)，但具體會(huì)在何時(shí)成為現(xiàn)實(shí)卻無(wú)法預(yù)知。這種時(shí)間上的不確定性讓企業(yè)傾向于將其視為次要問(wèn)題，并推遲為量子計(jì)算機(jī)的到來(lái)調(diào)整網(wǎng)絡(luò)安全防御的必要活動(dòng)。

正如網(wǎng)絡(luò)安全公司Quantropi的首席技術(shù)官M(fèi)ike Redding所說(shuō)：

“除非量子計(jì)算機(jī)已經(jīng)出現(xiàn)，人們會(huì)說(shuō)，‘沒(méi)關(guān)系，我們以后再處理，或者供應(yīng)商會(huì)幫我解決。我的事情已經(jīng)夠多了，預(yù)算也有限?！?/p>

他補(bǔ)充道：“量子技術(shù)也許是有史以來(lái)最重要的事情，但對(duì)大多數(shù)人來(lái)說(shuō)，它并不緊迫，他們只是把問(wèn)題往后推。”

忽視的代價(jià)

這種松懈心態(tài)可能會(huì)導(dǎo)致災(zāi)難性的后果，因?yàn)閱?wèn)題的關(guān)鍵并不是量子計(jì)算機(jī)是否會(huì)到來(lái)，而是何時(shí)到來(lái)。

專家共識(shí)：

雖然量子計(jì)算機(jī)的具體時(shí)間表尚不明確，但絕大多數(shù)專家認(rèn)為，一個(gè)能夠威脅加密安全的量子計(jì)算機(jī)將在未來(lái)5到10年內(nèi)出現(xiàn)。然而，企業(yè)需要多長(zhǎng)時(shí)間才能完成對(duì)基礎(chǔ)設(shè)施和第三方依賴的全面升級(jí)？8年？10年？甚至12年？

回顧歷史，從哈希算法SHA1遷移到SHA2的時(shí)間就很漫長(zhǎng)?？紤]到這種遷移的復(fù)雜性，盡早行動(dòng)是明智的選擇。正如美國(guó)管理和預(yù)算辦公室在一份報(bào)告中所指出的：

“很可能具備加密破解能力的量子計(jì)算機(jī)（CRQC）將能夠攻破目前政府和私營(yíng)部門(mén)廣泛使用的一些加密形式。盡管目前尚不知是否存在這樣的計(jì)算機(jī)，但量子計(jì)算領(lǐng)域的穩(wěn)步進(jìn)展可能會(huì)在未來(lái)十年內(nèi)帶來(lái)CRQC的誕生。因此，聯(lián)邦機(jī)構(gòu)必須加強(qiáng)現(xiàn)有信息系統(tǒng)的防御，遷移到使用量子抗性公鑰加密系統(tǒng)?！?/p>

問(wèn)題的規(guī)模與解決方案

量子計(jì)算機(jī)帶來(lái)的問(wèn)題可能影響范圍極大，但幸運(yùn)的是，現(xiàn)有的工具和專業(yè)知識(shí)已經(jīng)為企業(yè)提供了解決方案：

1. 后量子密碼學(xué)（PQC）標(biāo)準(zhǔn)：

美國(guó)國(guó)家標(biāo)準(zhǔn)與技術(shù)研究院（NIST）最近發(fā)布了PQC算法標(biāo)準(zhǔn)，這些算法可以在問(wèn)題變得昂貴之前化解風(fēng)險(xiǎn)。

2. 國(guó)際合作：

世界上許多國(guó)家的政府也在積極研究解決這一問(wèn)題的辦法。

此外，量子威脅還為企業(yè)提供了一個(gè)重新思考網(wǎng)絡(luò)安全的機(jī)會(huì)，以構(gòu)建更強(qiáng)大的安全體系。

Now: Cryptography everywhere

Two of the primary concerns for cybersecurity teams are technology integrity and operational disruption. Undermining digital signatures and cryptographic key exchanges that enable data encryption are at the heart of those fears. Losing the type of cryptography that can guarantee digital signatures are authentic and unaltered would likely deal a major blow to the integrity of communications and transactions. Additionally, losing the ability to transmit information securely could potentially upend most organizational processes.

Enterprises are starting to become aware of the risks posed by quantum computing to their cybersecurity. According to Deloitte’s Global Future of Cyber survey, 52% of organizations are currently assessing their exposure and developing quantum-related risk strategies. Another 30% say they are currently taking decisive action to implement solutions to these risks.

“The scale of this problem is sizeable, and its impact in the future is imminent. There may still be time when it hits us, but proactive measures now will help avoid a crisis later. That is the direction we need to take,” says Gomeet Pant, group vice president of security technologies for the India-based division of a large industrial products firm.

Cryptography is now so pervasive that many organizations may need help identifying all the places it appears. It’s in applications they own and manage, and in their partner and vendor systems. Understanding the full scope of the organizational risk that a CRQC would pose to cryptography (figure 1) requires action across a wide range of infrastructures, supply chains, and applications. Cryptography used for data confidentiality and digital signatures to maintain the integrity of emails, macros, electronic documents, and user authentication would all be threatened, undermining the integrity and authenticity of digital communications.

To make matters worse, enterprises’ data may already be at risk, even though there is no CRQC yet. There’s some indication that bad actors are engaging in what’s known as “harvest now, decrypt later” attacks—stealing encrypted data with the notion of unlocking it whenever more mature quantum computers arrive. Organizations’ data will likely continue to be under threat until they upgrade to quantum-resistant cryptographic systems.

“We identified the potential threat to customer data and the financial sector early on, which has driven our groundbreaking work toward quantum-readiness,” said Yassir Nawaz, director of the emerging technology security organization at JP Morgan. “Our initiative began with a comprehensive cryptography inventory and extends to developing PQC solutions that modernize our security through crypto-agile processes.”

Given the scale of the issues, upgrading to quantum-safe cryptography could take years, maybe even a decade or more, and we’re likely to see cryptographically relevant quantum computers sometime within that range. The potential threat posed by quantum to cryptography may feel over the horizon, but the time to start addressing it is now (figure 2).

“It is important that organizations start preparing now for the potential threat that quantum computing presents,” said Matt Scholl, computer security division chief at NIST. “The journey to transition to the new postquantum-encryption standards will be long and will require global collaboration along the way. NIST will continue to develop new post-quantum cryptography standards and work with industry and government to encourage their adoption.”

當(dāng)前趨勢(shì)：無(wú)處不在的加密

網(wǎng)絡(luò)安全團(tuán)隊(duì)當(dāng)前面臨的兩大核心問(wèn)題是技術(shù)完整性和運(yùn)營(yíng)中斷。

削弱數(shù)字簽名和支持?jǐn)?shù)據(jù)加密的加密密鑰交換正是這些擔(dān)憂的根源。

如果失去了能夠保證數(shù)字簽名真實(shí)性和未被篡改的加密技術(shù)，通信和交易的完整性可能會(huì)遭受重大打擊。此外，失去安全傳輸信息的能力可能會(huì)顛覆大多數(shù)組織的運(yùn)行流程。

企業(yè)對(duì)量子威脅的日益關(guān)注

企業(yè)已經(jīng)開(kāi)始意識(shí)到量子計(jì)算對(duì)網(wǎng)絡(luò)安全構(gòu)成的風(fēng)險(xiǎn)。

根據(jù)德勤的《全球網(wǎng)絡(luò)未來(lái)調(diào)查》：

? 52%的企業(yè)正在評(píng)估自身的暴露程度，并制定與量子相關(guān)的風(fēng)險(xiǎn)策略。

? 另有30%的企業(yè)表示，正在采取果斷行動(dòng)以實(shí)施應(yīng)對(duì)這些風(fēng)險(xiǎn)的解決方案。

印度一家大型工業(yè)產(chǎn)品公司安全技術(shù)部門(mén)的副總裁Gomeet Pant表示：

“這個(gè)問(wèn)題的規(guī)模相當(dāng)龐大，其未來(lái)的影響迫在眉睫?；蛟S我們還有時(shí)間應(yīng)對(duì)，但現(xiàn)在采取主動(dòng)措施可以避免未來(lái)的危機(jī)。這是我們需要前進(jìn)的方向?！?/p>

識(shí)別加密系統(tǒng)的全局風(fēng)險(xiǎn)

加密技術(shù)如今如此普遍，以至于許多組織可能難以識(shí)別它存在的所有位置。

加密不僅用于它們自有的應(yīng)用程序，還廣泛分布在合作伙伴和供應(yīng)商系統(tǒng)中。

要全面理解具備加密相關(guān)性的量子計(jì)算機(jī)（CRQC）對(duì)加密技術(shù)可能造成的風(fēng)險(xiǎn)（見(jiàn)圖1），企業(yè)需要在以下領(lǐng)域采取行動(dòng)：

? 基礎(chǔ)設(shè)施

? 供應(yīng)鏈

? 應(yīng)用程序

CRQC將威脅以下領(lǐng)域中使用的加密技術(shù)：

? 數(shù)據(jù)保密性

? 數(shù)字簽名的完整性

這包括電子郵件、宏指令、電子文檔和用戶認(rèn)證的完整性和真實(shí)性。

這些威脅可能會(huì)破壞數(shù)字通信的完整性與可信度。

“先采集，后解密”的新風(fēng)險(xiǎn)

更糟糕的是，即使CRQC尚未出現(xiàn)，企業(yè)的數(shù)據(jù)可能已經(jīng)面臨風(fēng)險(xiǎn)。

有跡象表明，惡意行為者正在進(jìn)行所謂的“先采集，后解密”攻擊：

? 他們竊取加密數(shù)據(jù)，等量子計(jì)算機(jī)技術(shù)成熟后再進(jìn)行解密。

因此，直到企業(yè)升級(jí)到量子抗性加密系統(tǒng)之前，其數(shù)據(jù)將持續(xù)處于威脅之下。

JP Morgan新興技術(shù)安全組織總監(jiān)Yassir Nawaz表示：

“我們很早就識(shí)別到了客戶數(shù)據(jù)和金融行業(yè)可能面臨的潛在威脅，這推動(dòng)了我們?cè)诹孔訙?zhǔn)備方面的開(kāi)創(chuàng)性工作。

我們的計(jì)劃從全面的加密技術(shù)盤(pán)點(diǎn)開(kāi)始，并延伸到開(kāi)發(fā)后量子密碼學(xué)（PQC）解決方案，通過(guò)靈活加密流程來(lái)現(xiàn)代化我們的安全防護(hù)?！?/p>

升級(jí)到量子安全加密的時(shí)間窗口

鑒于問(wèn)題的規(guī)模，升級(jí)到量子安全加密可能需要數(shù)年，甚至十年以上。而根據(jù)專家預(yù)測(cè)，CRQC可能會(huì)在這段時(shí)間范圍內(nèi)出現(xiàn)。

量子對(duì)加密的威脅似乎還很遙遠(yuǎn)，但現(xiàn)在正是開(kāi)始解決這個(gè)問(wèn)題的最佳時(shí)機(jī)（見(jiàn)圖2）。

NIST計(jì)算機(jī)安全部門(mén)負(fù)責(zé)人Matt Scholl表示：

“組織必須從現(xiàn)在開(kāi)始為量子計(jì)算可能帶來(lái)的威脅做好準(zhǔn)備。

從當(dāng)前加密系統(tǒng)過(guò)渡到新的后量子加密標(biāo)準(zhǔn)將是一個(gè)漫長(zhǎng)的過(guò)程，需要全球范圍內(nèi)的協(xié)作。

NIST將繼續(xù)開(kāi)發(fā)新的后量子密碼學(xué)標(biāo)準(zhǔn)，并與行業(yè)和政府合作，推動(dòng)這些標(biāo)準(zhǔn)的采用?！?/p>

The intelligent core: AI changes everything for core modernization

For years, core and enterprise resource planning systems have been the single source of truth for enterprises’ systems of records. AI is fundamentally challenging that model.

Many core systems providers have gone all in on artificial intelligence and are rebuilding their offerings and capabilities around an AI-first model. The integration of AI into core enterprise systems represents a significant shift in how businesses operate and leverage technology for competitive advantage.

It’s hard to overstate AI’s transformative impact on core systems. For years, the core and the enterprise resource planning tools that sit on top of it were most businesses’ systems of record—the single source of truth. If someone had a question about any aspect of operations, from suppliers to customers, the core had the answer.

AI is not simply augmenting this model; it’s fundamentally challenging it. AI tools have the ability to reach into core systems and learn about an enterprise’s operations, understand its process, replicate its business logic, and so much more. This means that users don’t necessarily have to go directly to core systems for answers to their operational questions, but rather can use whatever AI-infused tool they’re most familiar with. Thus, this transformation goes beyond automating routine tasks to fundamentally rethinking and redesigning processes to be more intelligent, efficient, and predictive. It has the potential to unleash new ways of doing business by arming workers with the power of AI along with information from across the enterprise.

No doubt, there will be integration and change management challenges along the way. IT teams will need to invest in the right technology and skills, and build robust data governance frameworks to protect sensitive data. The more AI is integrated with core systems, the more complicated architectures become, and this complexity will need to be managed. Furthermore, teams will need to address issues of trust to help ensure AI systems are handling critical core operations effectively and responsibly.

But tackling these challenges could lead to major gains. Eventually, we expect AI to progress beyond being the new system of record to become a series of agents that not only do analyses and make recommendations but also take action. The ultimate endpoint is autonomous decision-making, enabling enterprises to operate quickly compared with their current pace of operations.

Now: Businesses need more from systems of record

Core systems and, in particular, enterprise resource planning (ERP) platforms are increasingly seen as critical assets for the enterprise. There’s a clear recognition of the value that comes from having one system hold all the information that describes how the business operates. For this reason, the global ERP market is projected to grow at a rate of 11% from 2023 through 2030. This growth is driven by a desire for both greater efficiency and more data-driven decision-making.1

The challenge is that relatively few organizations are realizing the benefits they expect from these tools. Despite an acknowledgment that a centralized single source of truth is key to achieving greater operational efficiency, many ERP projects don’t deliver. According to Gartner research, by 2027, more than 70% of recently implemented ERP initiatives will fail to fully meet their original business case goals.2

Part of the reason ERP projects may fail to align with business goals is that the systems tend to be one-size-fits-all. Businesses needed to mirror their operations to fit the ERP system’s model. Applications across the organization were expected to integrate with the ERP. It was the system of record and held all business data and business logic, so the organization acquiesced to these demands, even if they were hard to meet. However, this produced a certain level of disconnect between the business and the ERP system.

AI is breaking this model. Some enterprises are looking to reduce their reliance on monolithic ERP implementations, and AI is likely to be the tool that allows them to by opening up data sets and enabling new ways of working.

六、智能核心：AI正在重塑核心現(xiàn)代化

多年來(lái)，核心系統(tǒng)和企業(yè)資源計(jì)劃（ERP）系統(tǒng)一直是企業(yè)記錄管理的“唯一真實(shí)來(lái)源”。但人工智能（AI）正在從根本上挑戰(zhàn)這一模式。

AI如何改變核心系統(tǒng)

許多核心系統(tǒng)供應(yīng)商已經(jīng)全面擁抱AI，并將其產(chǎn)品與功能圍繞“AI優(yōu)先”模式進(jìn)行重建。將AI整合到核心企業(yè)系統(tǒng)中，標(biāo)志著企業(yè)運(yùn)營(yíng)和技術(shù)應(yīng)用方式的重大轉(zhuǎn)變，為企業(yè)競(jìng)爭(zhēng)優(yōu)勢(shì)提供了全新的路徑。

多年來(lái)，企業(yè)依賴核心系統(tǒng)和其上的ERP工具作為記錄管理的基礎(chǔ)。如果對(duì)運(yùn)營(yíng)中的任何方面有疑問(wèn)，無(wú)論是供應(yīng)商還是客戶，答案都可以從核心系統(tǒng)中找到。

然而，AI的影響不僅僅是增強(qiáng)這一模式，而是從根本上挑戰(zhàn)它。AI工具能夠深入核心系統(tǒng)，學(xué)習(xí)企業(yè)的運(yùn)營(yíng)流程、理解其業(yè)務(wù)邏輯，甚至能夠復(fù)制這些流程。這意味著用戶不再需要直接訪問(wèn)核心系統(tǒng)來(lái)獲取問(wèn)題的答案，而是可以使用他們最熟悉的AI工具。

這種變革不僅局限于自動(dòng)化常規(guī)任務(wù)，而是從根本上重新設(shè)計(jì)和優(yōu)化流程，使其更加智能、高效和具有預(yù)測(cè)能力。AI結(jié)合整個(gè)企業(yè)的信息，能夠釋放全新的業(yè)務(wù)模式，為員工賦能。

集成與管理的挑戰(zhàn)

不可否認(rèn)，在實(shí)現(xiàn)這一轉(zhuǎn)型的過(guò)程中會(huì)面臨集成和變更管理方面的挑戰(zhàn)。

? 技術(shù)與技能投資：IT團(tuán)隊(duì)需要選擇合適的技術(shù)并提升團(tuán)隊(duì)技能。

? 數(shù)據(jù)治理框架：建立健全的數(shù)據(jù)治理框架，保護(hù)敏感數(shù)據(jù)免受風(fēng)險(xiǎn)。

? 復(fù)雜架構(gòu)管理：隨著AI深入核心系統(tǒng)，系統(tǒng)架構(gòu)將變得更加復(fù)雜，這需要團(tuán)隊(duì)有效應(yīng)對(duì)。

? AI的信任問(wèn)題：確保AI系統(tǒng)在處理關(guān)鍵核心操作時(shí)的高效性與責(zé)任性也同樣重要。

盡管如此，克服這些挑戰(zhàn)將帶來(lái)巨大的收益。未來(lái)，AI可能不僅僅是一個(gè)新的記錄系統(tǒng)，還會(huì)發(fā)展成為一系列智能代理，不僅能夠分析和提出建議，還可以直接采取行動(dòng)。最終，這將實(shí)現(xiàn)自主決策，使企業(yè)的運(yùn)營(yíng)速度遠(yuǎn)超當(dāng)前水平。

現(xiàn)在：企業(yè)需要核心系統(tǒng)提供更多支持

核心系統(tǒng)，尤其是ERP平臺(tái)，被越來(lái)越多地視為企業(yè)的重要資產(chǎn)。企業(yè)普遍認(rèn)識(shí)到，集中管理所有業(yè)務(wù)信息的系統(tǒng)是實(shí)現(xiàn)更高效率和數(shù)據(jù)驅(qū)動(dòng)決策的關(guān)鍵。

正因?yàn)槿绱?，全球ERP市場(chǎng)預(yù)計(jì)將以11%的年增長(zhǎng)率從2023年持續(xù)增長(zhǎng)至2030年。這種增長(zhǎng)主要由企業(yè)對(duì)更高效率和數(shù)據(jù)驅(qū)動(dòng)決策的需求推動(dòng)。

為什么許多ERP項(xiàng)目未能滿足預(yù)期？

盡管企業(yè)已經(jīng)意識(shí)到ERP系統(tǒng)的價(jià)值，但現(xiàn)實(shí)中，只有少數(shù)組織能夠真正從中獲益。根據(jù)Gartner的研究，到2027年，超過(guò)**70%**的新ERP實(shí)施項(xiàng)目將無(wú)法完全實(shí)現(xiàn)其原定的商業(yè)目標(biāo)。

原因之一是ERP系統(tǒng)的“千篇一律”：

? 企業(yè)需要調(diào)整自己的業(yè)務(wù)流程以適應(yīng)ERP系統(tǒng)的模型。

? 企業(yè)內(nèi)部的應(yīng)用程序需要與ERP進(jìn)行整合。

由于ERP作為記錄系統(tǒng)，持有所有的業(yè)務(wù)數(shù)據(jù)和邏輯，企業(yè)被迫適應(yīng)其要求，盡管這些要求可能難以滿足。這種模式導(dǎo)致了企業(yè)與ERP系統(tǒng)之間的脫節(jié)。

AI如何打破傳統(tǒng)模式

一些企業(yè)希望減少對(duì)單一ERP系統(tǒng)的依賴，而AI正是實(shí)現(xiàn)這一目標(biāo)的關(guān)鍵工具：

? 開(kāi)放數(shù)據(jù)集：AI使數(shù)據(jù)更加靈活和可用。

? 改變工作方式：AI提供了全新的、更智能的工作方式。

這不僅是技術(shù)的升級(jí)，更是企業(yè)運(yùn)營(yíng)模式的全面變革。

New: AI augments the core

With some evolution, ERP systems will likely maintain their current position as systems of record. In most large enterprises, they still hold virtually all the business data, and organizations that have spent the last several years implementing ERP systems will likely be reluctant to move on from them.

Orchestrating the platform approach

In this model, today’s core systems become a platform upon which AI innovations are built. However, this prospect raises multiple questions around AI orchestration that IT and business leaders will have to answer. Do they use the modules provided by vendors, use third-party tools, or, in the case of more tech-capable teams, develop their own models? Relying on vendors means waiting for functionality but may come with greater assurance of easy integration.

Another question is how much data to expose to AI. One of the benefits of generative AI is its ability to read and interpret data across different systems and file types. This is where opportunities for new learnings and automation come from, but it could also present privacy and security challenges. In the case of core systems, we’re talking about highly sensitive HR, finance, supplier, and customer information. Feeding this data into AI models without attention to governance could create new risks.

There’s also the question of who should own initiatives to bring AI to the core. This is a highly technical process that demands the skills of IT—but it also supports critical operational functions that the business should be able to put its fingerprints on.

The answer to these questions will likely look different from use case to use case and even enterprise to enterprise. But teams should think about them and develop clear answers before going all in on AI in the core. These answers form the foundation upon which rests the larger benefits of the technology.

“To get the most out of AI, companies should develop a clear strategy anchored in their business goals,” says Eric van Rossum, chief marketing officer for cloud ERP and industries at SAP. “AI shouldn’t be considered as a stand-alone functionality, but rather as an integral, embedded capability in all business processes to support a company’s digital transformation.”

AI enables new ways of working

Forward-looking enterprises are already answering these orchestration questions. Graybar, a wholesale distributor of electrical, industrial, and data communications solutions, is in the middle of a multiyear process of modernizing a 20-year-old core system implementation, which started with upgrades to its HR management tools and is now shifting to ERP modernization. It’s leaning on the best modules available from its core systems vendors when it makes sense, while also layering on third-party integrations and homegrown tools when there’s an opportunity to differentiate its products and services.?

The growth of AI presented leaders at the company with an opportunity to not only upgrade its tech stack, but also to think about how to reshape processes to drive new efficiencies and revenue growth. Trust has been a key part of the modernization efforts. The company is rolling out AI in narrowly tailored use cases where tools are applied to ensure reliability, security, and clear business value.

新的模式：AI助力核心系統(tǒng)升級(jí)

隨著不斷的演變，ERP系統(tǒng)很可能會(huì)繼續(xù)保持其作為企業(yè)記錄管理“系統(tǒng)真相”的核心地位。在大多數(shù)大型企業(yè)中，這些系統(tǒng)仍然承載著幾乎所有的業(yè)務(wù)數(shù)據(jù)，而那些花費(fèi)數(shù)年時(shí)間實(shí)施ERP系統(tǒng)的企業(yè)，通常也不愿意輕易放棄它們。

打造平臺(tái)化協(xié)同模式

在這種新模式下，現(xiàn)有的核心系統(tǒng)將演變成一個(gè)平臺(tái)，成為AI創(chuàng)新的基礎(chǔ)。然而，這種前景也帶來(lái)了多個(gè)需要IT和業(yè)務(wù)領(lǐng)導(dǎo)者解答的問(wèn)題：

? 是否依賴供應(yīng)商的模塊？

? 是否使用第三方工具？

? 是否由技術(shù)能力強(qiáng)的團(tuán)隊(duì)自行開(kāi)發(fā)模型？

數(shù)據(jù)的開(kāi)放程度也是一個(gè)需要關(guān)注的問(wèn)題。生成式AI的優(yōu)勢(shì)在于能夠跨不同系統(tǒng)和文件類型讀取和解讀數(shù)據(jù)，從而帶來(lái)新的洞察和自動(dòng)化機(jī)會(huì)。但與此同時(shí)，這也可能帶來(lái)隱私和安全風(fēng)險(xiǎn)，尤其是在處理核心系統(tǒng)中高度敏感的HR、財(cái)務(wù)、供應(yīng)商和客戶數(shù)據(jù)時(shí)。

在缺乏強(qiáng)有力治理的情況下，將這些數(shù)據(jù)輸入AI模型可能會(huì)引發(fā)新的風(fēng)險(xiǎn)。

另一個(gè)問(wèn)題是：AI在核心系統(tǒng)中的落地應(yīng)該由誰(shuí)負(fù)責(zé)？

這不僅是一個(gè)高度技術(shù)化的過(guò)程，需要IT的專業(yè)技能，同時(shí)也涉及業(yè)務(wù)部門(mén)的關(guān)鍵運(yùn)營(yíng)職能，因此需要業(yè)務(wù)部門(mén)的深度參與。

答案可能因不同的用例和企業(yè)情況而異。但企業(yè)在全面擁抱核心系統(tǒng)中的AI之前，應(yīng)該提前考慮這些問(wèn)題，并制定清晰的解決方案。這些答案將構(gòu)成AI技術(shù)進(jìn)一步釋放價(jià)值的基礎(chǔ)。

SAP云ERP與行業(yè)首席營(yíng)銷官Eric van Rossum表示：

“為了充分利用AI，企業(yè)應(yīng)該制定一個(gè)以業(yè)務(wù)目標(biāo)為核心的清晰戰(zhàn)略。AI不應(yīng)該被看作獨(dú)立的功能，而是應(yīng)該成為嵌入所有業(yè)務(wù)流程中的關(guān)鍵能力，從而支持企業(yè)的數(shù)字化轉(zhuǎn)型?！?/p>

AI推動(dòng)全新工作模式

前瞻性的企業(yè)已經(jīng)開(kāi)始回答這些協(xié)同問(wèn)題。

例如，Graybar（一家電氣、工業(yè)和數(shù)據(jù)通信解決方案的批發(fā)分銷商）正在進(jìn)行一個(gè)為期多年的現(xiàn)代化升級(jí)項(xiàng)目，該項(xiàng)目涉及對(duì)已有20年的核心系統(tǒng)進(jìn)行全面改造。

? 他們的現(xiàn)代化進(jìn)程從HR管理工具的升級(jí)開(kāi)始，目前已經(jīng)轉(zhuǎn)向ERP系統(tǒng)的現(xiàn)代化升級(jí)。

? 在此過(guò)程中，Graybar在適用的情況下依賴核心系統(tǒng)供應(yīng)商提供的最佳模塊，同時(shí)在有機(jī)會(huì)差異化其產(chǎn)品和服務(wù)時(shí)，引入第三方集成以及自主開(kāi)發(fā)的工具。

AI的增長(zhǎng)為公司領(lǐng)導(dǎo)層提供了一個(gè)不僅可以升級(jí)技術(shù)棧，還可以重新思考業(yè)務(wù)流程的機(jī)會(huì)，以推動(dòng)新的效率提升和收入增長(zhǎng)。

在這一現(xiàn)代化過(guò)程中，信任是關(guān)鍵因素之一。公司正在針對(duì)具體的、狹義的用例推出AI工具，確保這些工具在安全性和可靠性方面都符合要求。