Agentic AI in the Enterprise: Governance, Data, and an Honest Look at ROI

On each of the six big platform keynotes in 2026, a future was painted in which autonomous agents do much of the enterprise’s work. I take such stage promises with respect — and with scepticism. Because the more interesting question than “what is possible?” is, for decision-makers: “what of this actually lands in the company today, and at what price?”

This article is the honest look at the impact of agentic AI — beyond the demos. It belongs to my series on the 2026 keynote season; the full overview and the strategic context are available separately.

The sobering figure of the season

The most honest statement came, of all places, from one of the loudest vendors. ServiceNow cited its own research on stage: six in ten companies already use agentic AI, but only one in ten has built something genuinely autonomous — and 95 percent cannot quantify the ROI of their AI. On average, ServiceNow said, 367 applications run per company, AI mostly bolted on without governance.

That is not a technology problem. It is a competence and governance problem. The platforms are mature enough; what is missing is the organisations’ ability to create value and measure it. The real hurdle is almost never the technology — this season confirms it.

Where the impact is real

Despite the sobering headline figure, there are solid examples — which I deliberately mark as vendor claims, because that is how they reached the stage. Salesforce reported that strong adopters save up to 20 hours a week with the agent in Slack. Oracle cited a 30 percent reduction in call-centre handling time at its customer TIM Brazil. ServiceNow stated it had recovered 2.3 million employee hours internally.

The pattern behind the numbers matters more than the numbers themselves: the impact is greatest where agents take on clearly bounded, high-volume tasks — service, support, routine analysis. Not as spectacular full automation, but as consistent relief in the right places.

Why governance decides the pace

The strongest image of the season came from ServiceNow with a horror scenario: an agent deleting production databases and backups in nine seconds. The lesson is not “stay away” but: without a control plane, speed becomes risk. Accordingly, every vendor built one — from the kill switch in the AI Control Tower to Workday’s Agent Passport.

For decision-makers this means, paradoxically: governance is not a brake but the enabler of speed. Only once an agent has an identity, clear rights and a complete audit trail can you let it loose on real processes with a clear conscience. How autonomy and human control can be balanced cleanly I explore in “Claude Managed Agents”; why machine identities are the blind spot here, in “Agentic AI in Cybersecurity”.

Data first — or the agent hallucinates

Oracle called data leakage between agents a “fatal problem” and deliberately moved access security down to the database layer, so an agent sees only the data the respective user is allowed to see. That is more than a technical detail: an agent working on dirty or unsecured data produces confidently wrong results — and is therefore more dangerous than no automation at all.

The practical conclusion: before you invest in agents, invest in data quality, the access model and the semantic layer. This sequence is not negotiable — it is the precondition for the measured ROI to come out positive later.

The talent double-blow

Two pressures converge in 2026. On one side, SAP promises to cut the effort of ERP migrations by up to 50 percent with AI tooling — so the transformation pressure rises. On the other, ServiceNow cited a demographic gap of up to 50 million workers by 2030. Agents are the answer to both — but someone has to build, ground and supervise them.

This makes the bottleneck human, not technological. Which new roles companies have to build or buy for this I describe in “The New Enterprise Skills for 2026”.

A checklist for decision-makers

Before you launch the next agent initiative, I would answer six questions honestly:

• Data foundation: are the relevant data clean, accessible and semantically described?
• Access model: does the agent really see only what it is allowed to — at the data level, not just in the interface?
• Governance: is there identity, audit and a kill switch for every agent?
• Scope: is the use case clearly bounded and high-volume enough to create value?
• Measurement: is it defined, before launch, how you measure success and ROI?
• Competence: do you have the roles in-house — or a plan to build or buy them?

Whoever cannot answer these six questions is very likely to end up among the 95 percent who cannot quantify their ROI. Whoever can answer them belongs to the minority that draws real, measurable value from agentic AI.

Translating hype into measurable outcomes — with the right sequence of data, governance and enablement — is exactly what I do. If you want to put your agent strategy on a solid foundation, get in touch.