We spent years teaching employees how to use AI. Now we are teaching AI how to do their jobs.
But before we hand over the keys to our business processes, perhaps we should ask who is still responsible for driving.
We have moved from AI answering questions to AI executing tasks. From copilots recommending actions to autonomous agents making decisions, accessing enterprise systems and coordinating with other agents.
The promise? Speed, efficiency and a significant reduction in operational friction.
The uncomfortable reality?
We are delegating execution faster than we are defining accountability.
And having spent years in enterprise architecture and transformation, I find that a rather dangerous way to redesign an operating model.
The UAE is making a bold move. And asking an important question.
In April 2026, the UAE announced its ambition to transition 50% of government sectors, services and operations towards Agentic AI within two years.
That is not another chatbot initiative. It is an ambition to fundamentally change how government operates.
Imagine a resident applying for a government service. Instead of navigating multiple departments, an AI agent could retrieve documents, verify eligibility, coordinate approvals and initiate the next steps.
Fewer handovers. Faster decisions. Less bureaucracy.
The potential is enormous.
And I particularly welcome the parallel ambition to develop Agentic AI capabilities among 80,000 federal government employees. Because technology transformation without human capability building is rarely sustainable.
But here is where I would challenge the conversation.
If 50% of government operations become agentic, who is accountable when an agent makes the wrong decision?
Consider a hypothetical scenario.
An AI agent retrieves outdated information from a government database. Another agent uses it to assess eligibility. A third updates the records and rejects an application.
All three agents complete their tasks successfully.
The dashboard is green.
The citizen’s application is rejected.
And the decision is wrong.
Technically successful. Operationally disastrous.
Who verified the information? Who authorised the decision? Who can reverse it?
More importantly, could the error have been prevented before the agents acted?
This is the difference between automating a workflow and governing an autonomous decision.
Your AI agent has access. But does it have authority?
We have spent decades designing enterprise systems around access controls, segregation of duties, approval hierarchies and accountability.
Now we are introducing agents that can access those same systems, interpret information and initiate actions.
Yet many organisations still approach Agentic AI as another technology implementation.
It isn’t.
Agentic AI is an operating-model transformation with a technology component.
An agent summarising a document is not the same as an agent transferring money, signing off a supplier or making a decision affecting someone’s access to public services.
The risk changes with the action.
So should the governance.
And here is another problem: an agent can be fully logged, correctly authorised and still act on information that is fundamentally wrong.
We could end up with a perfectly documented mistake.
An audit trail is important. But it is not a substitute for verifying the evidence before a consequential action is taken.
We have enough AI governance principles. Can we operationalise them?
The EU AI Act, NIST AI Risk Management Framework, ISO/IEC 42001, ISO/IEC 23894, OECD AI Principles and UNESCO’s AI ethics recommendations, alongside GDPR and applicable national and sector-specific requirements, provide important foundations for responsible AI.
They are not interchangeable. Some impose legal obligations, while others offer standards, guidance or principles.
But none is a ready-made execution architecture for every autonomous agent workflow.
That is where I believe the conversation needs to move.
From AI Governance to AI Governance Operations (AIGovOps).
From documenting what AI should do to embedding controls into what AI is permitted to execute.
This thinking also led me to develop TARA-A, the Enterprise Runtime for Agentic Accountability, through HumanAIze.
Rather than adding another compliance checklist, the approach considers major governance and regulatory requirements and translates them into risk-based authority, evidence verification, human accountability and technical controls at the point of execution.
It has evolved beyond a whitepaper into a live governance maturity assessment and a built Governance-as-Code kernel, with a broader Governance Operating System in development.
But TARA-A is not the point of this article.
The larger question is whether our governance mechanisms can operate at the speed of the autonomy we are introducing.
A policy that says an agent must not exceed its authority is one thing. An architecture that prevents it is quite another.
Are we automating value, or just activity?
There is another question I wish more boards would ask.
What business problem are we actually solving?
I have seen enough enterprise transformations to know that automating a poorly designed process rarely fixes the underlying problem.
It simply makes the dysfunction faster.
Before deploying agents, understand the current state. Map the process, its dependencies, exceptions and decision rights. Establish the business case.
Then ask whether autonomy is even necessary.
Sometimes conventional automation is the better answer.
And when Agentic AI does make sense, measure the outcomes.
Not the number of agents deployed.
Not the number of tasks completed.
Not how impressive the demo looks.
Measure cycle time, cost, quality, risk, customer or citizen outcomes and actual P&L impact.
If the agent delivers no measurable value, why are we scaling it?
If it introduces risks we cannot control, why are we deploying it?
AI adoption may be inevitable. Business value is not.
The real opportunity for the UAE, and the rest of us
I admire the UAE’s appetite for ambitious transformation. Having worked extensively within German enterprise environments and now building in the UAE, I see the value of both worlds.
German governance discipline and UAE execution agility need not be competing philosophies.
The next generation of enterprise AI needs both.
Because speed without control is fragile.
And governance that cannot keep pace with execution becomes a bottleneck.
The UAE has an opportunity not only to lead in Agentic AI adoption, but to help establish what accountable autonomy looks like at scale.
For enterprises, the lesson is equally relevant.
The question is no longer simply whether AI can perform the task.
It is whether we understand the authority we are handing over, the consequences of getting it wrong and the value we expect in return.
We should be designing that accountability before deployment, not reconstructing it after an incident.
An AI agent may take over the task. It should never take over accountability.
Because autonomy must be earned.
And accountability cannot be delegated.
