Who Can Switch Off Your AI Agents? A Control Checklist

Most businesses can tell you how many laptops they own. Far fewer can say how many AI agents are working inside their finance function, who set each one up, or how to stop one quickly. New research suggests that gap is real, and it is widest where adoption is fastest.

For SME owners and CFOs, the fix is not more technology. It is the same discipline you already apply to bank mandates and signing authority, extended to software that acts on your behalf.

What the research found

In a survey of 685 chief information officers across eight markets, including the UAE, carried out by The Harris Poll for Dataiku in July 2026, 84% agreed that employees are creating AI agents and applications faster than IT can govern them. And 81% said they lack complete oversight of agents built outside approved systems or formal channels.

Tracking is not the same as checking. While 90% of CIOs said they were confident they track those agents, 72% could not consistently confirm whether the agents were delivering the business outcomes they were built for. Some 60% said they have no central governance layer across their apps, tools and IT environments.

Responsibility is blurred too. When an agent goes wrong, CIOs were split on who answers for it: shared teams (23%), central IT (21%), the data or AI team (20%), and security, risk and compliance (18%).

The UAE picture: fast adoption, slow brakes

The National's report on the UAE figures, published on 5 October, shows how quickly agents are spreading in the Gulf. In the UAE, 62% of CIOs said their companies have more than 50 AI agents at work, and 15% said they run more than 500. That is the highest share across the eight markets, against a 9% global average.

The brakes have not kept pace. Only 5% of UAE CIOs said they could reliably identify and contain a problematic agent across environments within one to two hours, compared with 10% worldwide. And 80% had seen an agent act against business intent or policy while still working within its technical limits. The encouraging part: 62% of those caught the problem before it caused harm, the best result in the study.

That last finding matters most for finance. An agent can do exactly what it was technically allowed to do and still get it wrong, because nobody wrote down what "right" looks like.

What this means for your business

These are large-company numbers, but the same pattern shows up in smaller firms, often sooner. In a 40-person business there is rarely an IT team standing between a keen finance executive and a new AI tool.

For example, a Doha trading company might have one agent drafting payment reminders, a second matching bank lines, and a third, built by a sales manager, pulling overdue invoices into a spreadsheet. Each looks harmless on its own. Together, no single person knows what is running, what data each one touches, or who to call if reminders start going to customers who have already paid.

In India, where many SMEs juggle GST filings, TDS and bank uploads across several tools, the same risk applies to anything an agent can submit, send or post. The test is simple: if this agent misbehaved at 10am, could we stop it before lunch?

Treat agents the way you treat people with access to money and data. Every one needs a named owner, a written job description, clear limits and a way to be suspended.

A practical checklist for finance teams

Speaking to The National, Dataiku's head of AI strategy, Kurt Muehmel, recommended keeping one inventory of every agent, naming an owner for each, deciding in advance who can stop one, standardising how agents are tested, changed and retired, and practising containment like a cyber-security incident. Here is how that translates for an SME finance team:

• Build an agent register. One sheet: agent name, what it does, the systems and data it can reach, who built it, who owns it and when it was last reviewed. Include anything staff have set up themselves.
• Name one owner per agent. The owner answers for its output. It should not default to IT.
• Write down what "correct" means. In one or two lines, state what a good outcome looks like and what the agent must never do, such as emailing a customer whose invoice is in dispute.
• Decide who can stop it, and how. Name a primary and a backup person. Know the exact step: revoke the login, switch off the integration or pause the workflow.
• Rehearse once a quarter. Pick one agent and time how long it takes to pause it. If the answer is "we are not sure", fix that first.
• Check results, not just activity. Each month, sample the agent's work, whether reminders sent, matches made or entries posted, and compare it with what a person would have done.
• Retire what you no longer use. An old agent with live access is risk with no benefit.
• Keep sensitive actions behind a person. Anything that moves money, files a return or changes supplier bank details should need human sign-off.

If you use NexCFO for AR collections or bank reconciliation, put those agents on the same register and give each a named owner, as you would any other tool.

AI agents can save your finance team hours every week, but only if you can name every one of them and stop any of them before lunch.