Agentic AI, systems that can plan, act across tools and complete multi-step work with limited supervision, has moved from research into the operating plans of every major bank in the region. The demonstrations are impressive. The production results are more uneven, and the pattern of where they succeed is instructive.
Where it is working
KYC refresh and onboarding. Document collection, extraction, screening and case assembly are high-volume, rule-heavy and already measured. Agents that assemble the case for a human decision are cutting cycle times from days to hours in banks that have redesigned the underlying process.
The winning question is not what the agent can do. It is which process has a baseline, an owner and a regulator who will accept the answer.
Credit operations. Covenant monitoring, financial spreading and early-warning triage are natural fits. The work is repetitive, the data is structured and the value of speed is clear to the chief risk officer.
Reconciliations and regulatory reporting. Break investigation and narrative drafting for returns are being handled by agents with a human sign-off, releasing analyst time that had been spent on assembly rather than judgement.
Three conditions for a return
First, a baseline. If the process has not been measured, the agent's impact cannot be evidenced and the programme will be judged on anecdote. Second, an owner. Somebody in the business, not in technology, has to own the outcome and the exceptions. Third, governance the regulator will accept. Model risk management, audit trails and explainability are not afterthoughts; SAMA, the CBUAE and the QCB will ask, and the answer has to exist.
Where it is not working
Agents deployed on processes that have never been redesigned. Agents in customer-facing roles where the tolerance for error is near zero and the brand risk is real. And programmes run by the technology function alone, without the operations leadership that knows where the exceptions live.
A practical sequence
Build the digital twin of the operation. Remove the waste. Pick the two processes with the clearest baseline and owner. Deploy agents with human-in-the-loop controls, measure weekly, and expand only when the return is banked. It is less exciting than the demo. It is also how the money gets made.
