Building AI that can work across the enterprise

Glass Epoch DeskOriginal digestSource: MIT Technology Review

AI-assisted digest, checked against the original source before publishing. How we work

SAP and Infosys executives argue that scaling AI agents forces enterprises to rebuild data foundations, governance, and operating models — not just add tools.
Building AI that can work across the enterprise
Photo: Kenrick Baksh · free stock license

Leiduck's case is that AI must be approached as an enterprise platform rather than a set of disconnected projects. AI, he says, is "absolutely dependent on the data foundation," meaning companies first need a connected, semantically rich layer that accurately reflects how their business actually works before AI can be deployed reliably across processes. The required accuracy changes with the stakes: for some everyday applications "80% might be good enough," but in supply chains or financial close, that same rate is "a very terrible result."

Singh reframes AI agents as "digital coworkers" and proposes screening potential use cases through three lenses — impact, feasibility, and scalability. The test is whether an agent can move meaningful business metrics, whether a company's data and talent can support it, and whether a successful deployment can be replicated elsewhere. In his framing, "a pilot that can't scale is just an experiment."

Governance is where scaling concentrates the risk. Singh points to clear policies, access controls, monitoring, and oversight as prerequisites for keeping agents trustworthy and aligned with business requirements. Skipping that work, he warns, doesn't accelerate adoption: "The absence of governance doesn't speed up AI adoption. It creates a risk backlog" — one that can ultimately force a far more disruptive stop-and-audit later.

Looking ahead, both executives describe a direction of travel in which agentic systems coordinate work across applications, while customer-specific knowledge graphs and governed data emerge as critical enterprise assets. Singh's longer-term framing is a move from AI as a productivity tool toward "AI as a decision-making partner" that helps leaders understand what is happening, anticipate what comes next, and decide how to respond.

Open questions center on execution: whether companies can build the semantically rich data layer Leiduck describes before deploying agents into mission-critical processes, and whether governance programs can stay ahead of the risk backlog Singh warns about as agents take on more of the digital coworker role.

The episode's framing also implies that the metrics that justified early pilots — task accuracy, individual productivity — may not be the ones that justify scaled deployments in areas like financial close and supply chain, where error tolerance is far lower.

Key facts

  • • Guests: Thorsten Leiduck (SAP) and Satinder Pal Singh (Infosys), hosted by Megan Tatum.
  • • Business Lab episode produced in partnership with Infosys.
  • • Singh proposes screening use cases by impact, feasibility, scalability.
  • • Leiduck: 80% accuracy is a "very terrible result" in financial close.
  • • Singh: missing governance creates a "risk backlog" forcing stop-and-audit.

Why it matters

Builders should expect enterprise AI procurement to shift from pilot demos toward audits of data semantics, access controls, and monitoring — governance evidence becomes a gating item, not a later add-on. Use-case selection will increasingly be filtered by replicability, so teams designing agents should specify which metrics move and how a deployment generalizes before pitching a pilot.

Spotted an error? Send a correction.

Get the day's AI news in one email

The Daily Epoch · 60-second read · unsubscribe anytime

Related