Your company is filling with AI agents. Who is accountable when one acts?

Somewhere in your business this quarter, an AI agent moved from a pilot someone ran on a Friday to a process the team now quietly depends on. No one announced it. No one signed it off. It simply started doing the work, and it kept doing it. You tend to find these things out sideways, in a status update, and a small question forms that you cannot quite park: if that agent gets something wrong at scale, who answers for it?

Here is the shape of it. The agents are arriving faster than the rules for them. Deloitte's 2026 State of AI study, a survey of 3,235 business and technology leaders across 24 countries, finds agentic AI (software that plans and acts across several steps on its own, rather than answering one prompt at a time) already in at least moderate use at 23% of organisations, heading toward nearly three-quarters within two years. The oversight is not keeping pace: just 21% report a mature model for governing those agents. The advantage over the next two years will not go to whoever deploys the most agents. It will go to whoever can let them act with confidence.

Why is the value still "untapped" if everyone is adopting AI?

Because adoption is not the same as transformation. Deloitte's own phrase for this moment is the untapped edge: the potential is real, and mostly unclaimed. Two-thirds of organisations report productivity gains and 53% report better decisions, yet the depth varies wildly. 34% say they are deeply transforming the business, while 37% are still using AI at the surface, layered onto workflows that never changed.

This is the part worth holding onto. Most leaders are trying to install new software on broken hardware, and an agent placed on top of an unchanged process simply makes the old mistakes faster. The edge stays untapped because value comes from redesigning the work, done by only about 30% of organisations, rather than from the number of agents switched on. It is the same reason behind why most organisations struggle to turn AI into profit: the technology lands, the operating model stays still.

What happens when agent adoption outruns governance?

You concentrate a risk you cannot see. An agent that acts on its own is only as safe as the guardrails around it, and right now the guardrails lag. Deloitte's 21% mature-governance figure is the tell. And this is not a niche concern: Stanford's 2025 AI Index found organisations widely recognise the risks, with 64% naming inaccuracy, 63% regulatory compliance and 60% cybersecurity, yet not all of them act on what they name. Recognition without action is precisely where an unattended agent does its quiet damage.

"Across the enterprise, we're seeing massive ambition around AI, with organisations starting to pivot from experimentation to integrating AI into the core of the business," says Nitin Mittal, Deloitte's Global AI Leader. That pivot is the opportunity and the exposure in one move. The boards that are ahead here are the ones already asking the sharper questions about oversight and readiness, before the agents scale, not after an incident forces the conversation.

How do we let agents act with confidence?

Deliberately, and in an order. Confidence in an autonomous system is built the same way a clean audit trail is: by design, before scale. Work through these steps, because the order is where most agent programmes quietly lose control.

1. Put a named owner on every agent in production. Accountability travels with a person, not a tool. If no one owns it, no one governs it.
2. Govern the autonomy, not just the model. Decide what an agent may do unsupervised, where it must stop and ask, and how you would know it had drifted. Only 21% of organisations have this; being in that fifth is the advantage.
3. Redesign the process before you scale the agent. Your job is to design the loop the agent runs in, not to bolt it onto the old one. An agent inherits the flaws of the workflow you give it.
4. Build the workforce fluency to supervise. 48% of organisations are upskilling and 36% are hiring specialised talent. The ability to judge an agent's output, and catch it when it is confidently wrong, is the new core competency of the work.
5. Measure drift, not just output. Agree what "still working correctly" looks like and review it on a cadence. An agent that was right in March can be quietly wrong by June.

What does this change for me as a leader this quarter?

It moves governance from the back office to your table. Not "are we deploying enough agents", which is answering itself, but "can we say what each one is allowed to do, and who answers when it acts". That is a leadership question, and it sits with you rather than with the vendor or the platform team.

There is a longer arc behind this. Phase One, the Age of Effort: work hard, get a little more, linear growth. Phase Two, the Age of Scale: build once, sell to many, exponential growth. Phase Three, the Age of Acceleration: output decoupled from human effort almost entirely, the phase agentic AI begins to deliver. Agents are the first tools that act rather than merely answer, which is why governance stops being paperwork and becomes the thing that decides whether the acceleration is yours to keep. Build the conditions now, and the agents compound in your favour.

About the author

Thomas Green

British technology futurist, AI keynote speaker and advisor. Thirty years across enterprise technology and AI strategy, helping leaders navigate the future of work. The futurist who died.