AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Building AI that can work across the enterprise

Collected Nov 16, 2026

In a Business Lab episode produced by MIT Technology Review in partnership with Infosys, two executives outlined why enterprise AI programs stall and what a scaled alternative looks like. Thorsten Leiduck, SVP and global head of SAP Business AI Platform Ecosystem, and Satinder Pal Singh, associate VP for North America for SAP BAIP and SAP AI Strategy at Infosys, spoke with host Megan Tatum. Their central argument is that AI should be run as an enterprise platform rather than a series of disconnected proof-of-concept projects, and that the data layer beneath it decides what is possible.

The numbers they cite describe a wide gap between interest and results. Leiduck points to a BCG report finding that 74% of organizations are stuck in pilot mode with no plan to scale, and an Anthropic survey in which 42% cited difficulty getting access to data of sufficient quality. Singh cites an Infosys and HFS Research survey of more than 500 global 2000 enterprises in which only 14% had successfully scaled AI, leaving 86% in early exploration, proof-of-concept or fragmented pilots. He calls the pattern the "AI as a department trap": an AI team or center of excellence builds technically strong pilots that the business does not recognize itself in, producing AI solutions in search of problems. Separately, the Infosys Knowledge Institute AI Maturity Index shows only 8% of companies qualifying as AI native, yet those firms generate four times the revenue contribution from AI compared with laggards, according to Singh.

Leiduck's view is that fragmented data foundations are the root cause. Running pilots on different AI technologies across different parts of an organization, he says, creates duplication, inconsistent performance and no clear route to scale. What companies need first is a connected, semantically rich data layer that reflects their own business, so agents are grounded in company reality rather than generic patterns. For developers, that is a useful way to think about retrieval and grounding: a semantic layer over ERP and line-of-business data is the difference between an agent that knows what a specific customer, order or cost center means and one that guesses from general text.

The executives draw a sharp line on accuracy. Leiduck says roughly 80% accuracy may be good enough for consumer-style tasks or drafting slides and email with a human in the loop, but in supply chains or financial close it is "a very terrible result." The implication is that tolerance for error should be set per workflow, not per vendor or model, and that business-process logic and industry context need to be embedded in both the data and the models.

Singh offers a three-lens filter for choosing use cases: impact, meaning whether an agent moves revenue, cost, working capital or risk; feasibility, meaning whether the current data foundation, platform and talent can actually execute it; and scalability, meaning whether a successful deployment can be replicated. Infosys maintains a library of more than 300 enterprise use cases across industries and, through its Topaz AI-first framework, pushes clients through all three lenses at once. He cites one engagement that cut complex SAP specification documents from 86 person hours to 6, a 92% reduction. His advice is to run the AI portfolio like an investment portfolio, with quick wins funding the program, medium-term bets and one or two bold long-term positions.

Governance comes up as a scaling constraint rather than a brake. Clear policies, access controls, monitoring and oversight are needed for agents to stay trustworthy and aligned with business requirements. Singh argues that skipping governance does not accelerate adoption but builds a "risk backlog" that can eventually force a more disruptive stop and audit. Leiduck adds that as deployments grow, the conversation shifts toward token costs, model performance, security and accountability, and that not every task justifies automation because human effort can remain more economical or effective.

On the partnership itself, Singh describes SAP as supplying the platform foundation, including Business Data Cloud for the semantically rich enterprise, a generative AI hub for governed model access and an agent hub for lifecycle governance. Infosys supplies implementation depth, industry context and accelerators, with more than 50 pre-configured AI use cases on the Business AI Platform that are production-ready without building from scratch. He reports a global S/4HANA transformation that completed over 480 RICEF objects in five months using AI-assisted delivery tooling, and an AI platform that auto-classifies more than 120,000 SAP incidents monthly, and says the combined effect cuts time to value from 12 to 18 months down to four to six months. Leiduck also references SAP's autonomous enterprise framework and domain assistants intended to run business processes end to end, and a forward-deployed engineering model in which use cases are defined with customers around specific outcomes before being reused elsewhere.

Looking ahead, both executives expect agentic systems to coordinate work across applications, with customer-specific knowledge graphs and governed data becoming core enterprise assets. Singh frames the longer-term shift as moving from AI as a productivity tool to AI as a decision-making partner that helps leaders see what is happening, anticipate what comes next and decide how to respond.

Why it matters: Teams building internal agents should expect scrutiny less on model choice than on data grounding, per-workflow accuracy thresholds, and the monitoring and access controls around each agent, since those are the areas the executives tie to scaling. The survey figures and customer examples come from SAP and Infosys and should be read as vendor-reported; my inference is that buyers will increasingly be asked to justify pilots by named business metrics and replication plans rather than by demo quality. The time-to-value compression claimed here, if it holds broadly, would shorten the payback window product teams can credibly promise.

Read at MIT Technology Review · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

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,…