hiw3c 日报精选
MIT Tech Review AI RSS 全文

构建可在整个企业范围内运行的AI

原文标题 · Building AI that can work across the enterprise

多年来,企业一直将AI视为一系列独立的实验。如今,当AI代理开始融入支撑业务的工作流程中时,这种模式开始显得不够适用。从孤立的试点项目转向AI代理和自主工作流程的组织正面临一个更深层的问题:重新思考数据、业务流程、治理……的运作方式。

来源:MIT Tech Review AI · www.technologyreview.com

正文为英文,可一键机器翻译(仅首次需要等待)

For years, enterprises have approached AI as a series of discrete experiments. Now, as AI agents begin to move into the workflows that run the business, that model is starting to look inadequate. Organizations moving from isolated pilots toward AI agents and autonomous workflows are confronting a deeper challenge: rethinking how data, business processes, governance, and people fit together, turning AI from a collection of tools into part of the enterprise operating model.

For Thorsten Leiduck, SVP and global head of SAP Business AI Platform Ecosystem, that shift begins with treating AI as an enterprise platform rather than a collection of disconnected projects. “AI is absolutely dependent on the data foundation,” he says, arguing that companies need a connected, semantically rich layer that reflects their business before they can reliably deploy AI across processes. The stakes also change as AI moves into mission-critical workflows. For some everyday applications, “80% might be good enough,” he says, but in areas such as supply chains or financial close, that level of accuracy is “a very terrible result.”

That foundation also determines which AI applications are worth pursuing. Satinder Pal Singh, associate VP North America for SAP BAIP and SAP AI Strategy at Infosys, frames AI agents as “digital coworkers” and recommends evaluating potential use cases through three lenses: impact, feasibility, and scalability. Companies should identify where an agent can move meaningful business metrics, determine whether their data and talent can support the use case, and then establish whether successful applications can be replicated. “A pilot that can’t scale is just an experiment,” he says.

Scaling also brings governance into sharper focus. Clear policies, access controls, monitoring, and oversight are necessary to ensure agents remain trustworthy and aligned with business requirements. “The absence of governance doesn’t speed up AI adoption,” Singh says. “It creates a risk backlog” that can ultimately force a more disruptive stop and audit.

Those considerations will become more important as AI agents take on more responsibilities as digital coworkers. Both executives see a future in which agentic systems coordinate work across applications, while customer-specific knowledge graphs and governed data become critical enterprise assets. For Singh, the longer-term opportunity is a shift from AI as a productivity tool to “AI as a decision-making partner,” helping leaders understand what is happening, anticipate what comes next, and determine how to respond.

This episode of Business Lab is produced in partnership with Infosys.

Full Transcript

Megan Tatum: From MIT Technology Review, I’m Megan Tatum, and this is Business Lab, the show that helps business leaders make sense of new technologies coming out of the lab and into the marketplace.

This episode is produced in partnership with Infosys.

Now, the development and democratization of AI tools has the potential to reconfigure how enterprises operate. But, at a time when applications are being scaled up from that initial phase of experimentation, we’re still seeing senior leaders grapple with questions around AI’s value, risk, and the complexity it adds. It’s becoming clear that, unlike past ways of automation, AI requires a real mindset shift. Put simply, the technology can’t simply be added into an enterprise’s existing toolkit. It requires a complete reconfiguration of what’s inside and how it’s used.


Two words for you: conscious configuration.

My guests today are Thorsten Leiduck, SVP and global head of SAP Business AI Platform Ecosystem, and Satinder Pal Singh, Associate VP for North America, SAP Business AI Platform and SAP AI Strategy.

Welcome, Thorsten and Satinder.

Thorsten Leiduck: Yeah, thank you for having us. Really a pleasure to be here with our valued partner, Infosys. Looking forward to the conversation.

Satinder Pal Singh: Hi, Megan. Thanks for inviting for the conversation today. Looking forward to the discussion.

Megan: Me too. Thank you both so much for being here. Thorsten, if I could start with you. We know that many senior leaders understand the transformative potential in AI, but they continue to struggle with aspects of its implementation. Many projects remain stuck at a pilot phase, for example. When you speak to clients, what are the biggest obstacles they face when it comes to integrating AI into their tech stack?

Thorsten: Yeah. Megan, as you mentioned, AI is on the agenda of every client conversation today, but many customers are really struggling how to convert that interest into an operationalized model and to unlock the value of AI across the enterprise. And the reason is that too many programs start as isolated POCs on different AI technologies and in an inconsistent fragmented data foundation.

This creates a lot of duplication, inconsistent performance, and no clear path to scale. At the same time, leaders are highly concerned about security, observability, reliability of agents, and some ethical implications when those get deployed. This is also shown in some studies. For example, BCG has launched a report, which shows that 74% of the organizations are stuck in pilot mode with no plan to scale. And in an Anthropic survey, data said that 42% are really struggling with getting access to the data and the quality to really, basically, scale the entire AI capabilities across the organization.

The path forward is really to treat AI not just as a project, but more holistically, as an enterprise platform, one that really consistently connects an unified data layer with the process context of the applications, and then embed AI capabilities governed across the entire life cycle. And to make this real, we in SAP launched the SAP autonomous enterprise framework, so organizations can move from pilots to scaled outcomes.

Megan: Yeah, really, to provide that support. And, Satinder, would you agree, why is it that we’re seeing enterprises get a touch of almost AI paralysis here? What’s the mindset shift that needs to happen and, perhaps, why is this something organizations need to be thinking about now, as well?

Satinder: Absolutely, Megan. And, if anything, the picture sharpens when you look at, specifically, at agentic AI. Infosys and HFS research surveyed over 500 global 2,000 enterprises and found only 14% have successfully scaled AI. Eighty-six percent remain in early exploration, proof of concept, or fragmented pilot phases. For a technology at the top of every boardroom agenda for two years, that’s a striking failure rate. And at the core of that 86% is what I would describe as AI as a department trap.

What happens is that an organization creates an AI team or a center of excellence and walls it off from business. That team builds technically excellent pilots the business doesn’t recognize itself in, and you end up with AI solutions searching for problems, rather than the other way around. Mindset shift is clear. Move from treating AI as a technology initiative to treating it as an operating model transformation. AI stops being something IT does for business and starts being something the whole organization does together. Every function, every workflow, embedded from the start.

And what makes this a now problem? Early movers are building an advantage. You cannot close by buying the same platform later. Eighteen months of agent behavior data, calibrated knowledge graph, workforce that has genuinely internalized AI, that is a structural advantage that compounds. Infosys Knowledge Institute AI Maturity Index shows that only 8% of the companies qualify as AI native today, yet they achieve four times the revenue contribution from the AI versus laggards. If you’re not building AI today, you’re not standing still, you’re falling behind at an accelerating rate.

Megan: Right. There’s a real polarization there, isn’t there, between the front-runners and the laggards, as you say. As we shift from pilot to scaled-up applications, perhaps the first step is evaluating when AI is the right tool. AI has quickly become the default solution within many enterprises. Is there a tendency to use it as a bit of a blunt instrument, nowadays, Thorsten? At the outset, how can senior leaders better evaluate what use cases are worthy of their time and investment, in order to strip out some of that later complexity?

Thorsten: Yeah, I think, Megan, what everybody needs to be aware is adopting and rolling out AI across an organization is as much a change management journey as it’s a technology rollout.

Now, when you have this in mind, it makes absolutely sense to put AI in front as many teams as possible, because you want to build formality, and spark ideas, and also people should see the benefit of AI in an organization. When you think about travel assistance, meeting summaries via Copilot, or AI-enabled workflows, increase the productivity, everybody feels this in their real job, in their real daily business. And, obviously, this creates creativity and also provides the organization from a bottom-up perspective with some innovation scale.

Now, however, as majority grows, the conversation, and this is also what we discuss with many of the leaders, and our clients, shifts more to economics and governance. Things like token costs, model performance, security, and accountability become much more relevant. And when you then take the approach and think about AI agents as digital workers that execute specific tasks within a business process, then this means, from an AI perspective, not every task justifies nation.

In some cases, human effort remains more economical or effective. The smart approach, and that’s what we really recommend to our clients, is to start with a defined, maybe small set of high-value use cases with defined success metrics upfront. So, for example, such as cycle time reduction, accuracy gains, or cost of per transaction, or top line improvement, and then scale what works. Those use cases can be tied to a line of business domains or industry scenarios.

And then, what we are doing from an SAP perspective, together with Infosys, we really help customers to be successful. We provide, as part of our autonomous suite, the according domain assistance, which really run business processes end-to-end and which deliver find ROIs for single end-to-end business process steps.

And then, last but not least, in many scenarios, I think, which has been very common pattern in particular around large industry AI use cases, is a joint approach, and we do this with partners like Infosys, but also customers around forward deployed engineering, so that you really define very closely the AI-specific use case with the defined outcomes together with the customer, and deliver this as the initial projects. And then, ideally, scale this to other customers and take it back like a best practice, which can be rolled out to other organizations, as well.

Megan: Right. It’s been really clear, isn’t it, on what you’re looking to achieve and what’s worth your time, as I say. And what would your advice, Satinder, be here, on refining an AI strategy to ensure organizations can focus on those areas, where AI can deliver the highest returns?

Satinder: Sure, Megan. I think Thorsten’s framing of AI agents as digital coworkers forces a precision that agentic AI strategy language often lacks. When you ask which digital coworkers do we need and where do they deliver the most value, the answer starts to become clearer right from the first step. At Infosys, through our Topaz AI-first framework, we push clients through the three lenses simultaneously.

The first one is impact. Where does deploying an agent move the numbers that matter? Revenue, cost, working capital risk. We have built a library of over 300 enterprise use cases across industries, so we never begin from a blank page. The second is feasibility. Given your data foundation platform and talent today, which high-impact use cases can you actually execute? Feasibility determines your sequence. The high-impact use case on a broken data foundation is not viable. When the foundation is right, the results are dramatic. One Infosys engagement cut complex SAP-specification documents from 86 person hours to 6, a 92% reduction.

Third is scalability. Once you have proven a use case with clear ROI metrics, cycle time reduction, accuracy gains, cost per transaction, what’s the architecture that lets you replicate it? A pilot that can’t scale is just an experiment. My practical advice: treat your AI portfolio like an investment portfolio, quick wins that fund the program, medium-term bets, one or two bold, long-term positions that balance keep stakeholders engaged at every stage and not just in year three.

Megan: That’s such a useful framework for people to use there. And given the urgency here, Thorsten, how can organizations approach and manage AI in a more structured, efficient way? To start with, what are the fundamentals that some organizations perhaps miss at the moment? I know Infosys and SAP talk about platform and data, then AI, for example.

Thorsten: Yeah, Megan, I think the fundamental, which many organizations are overlooking, is data and the data is a foundation for AI. Because, at the end, without the data structures and the ability to access the right information at the right time and basically bringing different kind of structured, unstructured data sources together, in a meaningful governed and semantic way, there will be no successful AI. AI is absolutely dependent on the data foundation. And I think this is often overlooked, and that’s why I said when you start POCs in different organization, or different parts of the organization with different technologies, it’s very hard to establish an end-to-end consistent model on the data side, as well as on the AI side.

One of the key goals companies need to achieve is, and that’s why we really talk about you need to start with data first and the platform first, is to establish and semantically-rich connected data layer that really reflects your business, so that the agents are really grounded in the company’s reality and not just in generic patterns.

And then, I think, the second element that’s important is accuracy. In the enterprise context, it really matters. For example, what might pass in a consumer setting, where you have 80% accuracy, and even in a workforce environment where you build PowerPoints or you do email drafts, which you then, with a human in the loop, optimize, 80% might be good enough. But, when you really think about mission-critical or business-critical processes in the supply chain, or closing the books, 80% is a very terrible result.

The stakes are much too high and you need to have much more accuracy. And that’s why it’s important that you have business process logic and industry context embedded into the data and the models which build the foundation, if you will, the backbone for every AI and, obviously, for the data in the ERP and the line of business system.

And then, last but not least, the other critical element, and you see this obviously in the news with all the agents going rogue, are governance, compliance, and security. Very critical in an enterprise context. You need to set clear policies, access controls. You need to have monitoring ability, capabilities, to really make sure that what then the AI agenda capabilities or agents are executing is fully aligned, is trustworthy, and scalable.

Megan: Right. Which is quite a lot of things for organizations to be thinking about. Satinder, how do SAP and Infosys work with organizations to streamline some of these processes, perhaps and make unlocking AI’s capability a little bit easier?

Satinder: Megan, I think at the onset, I want to acknowledge that, at Infosys, we are uniquely positioned for our relationship with SAP, because we are a customer to SAP, as well as one of their key global suppliers. And the way SAP and Infosys work together is designed around the architecture Thorsten described, that defines a coherent partnership.

SAP brings the platform foundation, business data cloud for the semantically-rich enterprise, generative AI hub for governed LLMXs, and agent hub for the lifecycle governance, enterprise grade from day one. Infosys brings implementation depth, industry context, and use case accelerators. Through Infosys Topaz, we have built over 50 pre-configured AI use cases on Business AI Platform, production-ready configurations deployable without building from scratch. We co-develop use cases within SAP’s industry AI engagement with clear customer-specific FTE metrics. That shared accountability for results is different from a typical implementation engagement.

In practice, one global S/4HANA transformation program used AI-assisted delivery tooling to complete over 480 RICEF objects in five months, a pace conventional delivery would not sustain. In another account, an AI-powered platform auto-classifies over 120,000 SAP incidents monthly, with zero analyst charge. The combined effect time to value drops from 12 to 18 months, down to four to six months.

What I would emphasize is change management damage. Technology is increasingly deployable, and Thorsten also spoke about it earlier. What’s hard is helping an organization internalize that their agents are digital coworkers, not automations. That mindset shift is 70% of the implementation challenge that platform cannot solve alone, and Infosys and SAP run joint adoption programs addressing both the dimensions in parallel.

Megan: Fantastic. Some great real-world examples there, as well. Thorsten, can you tell us a little bit more about SAP’s Business AI Platform, or BAIP, in particular? How does that work and what sets it apart from other platforms?

Thorsten: Yeah, so SAP Business AI Platform, if you will, is the foundation for our SAP autonomous enterprise framework, which really consists of a unified user interface and is enabled as an agentic