On April 23, His Highness Sheikh Mohammed bin Rashid Al Maktoum, Vice President and Prime Minister of the UAE and Ruler of Dubai, announced that half of UAE federal sectors, services and operations would move to agentic AI within two years. Implementation sits with His Highness Sheikh Mansour bin Zayed Al Nahyan, Vice President, Deputy Prime Minister and Chairman of the Presidential Court, and the task force is chaired by Mohammad Abdullah Al Gergawi, Minister of Cabinet Affairs, with AI training running through every federal employee. The percentage is what most coverage led with, but the buried fact is the posture: a government has said out loud that AI is moving inside its operations from advice to authorized action, and named the people accountable for making that move happen.
A memo I wrote a few weeks back, on board oversight of AI adoption, argued that boards should stop asking ‘what’s our AI strategy’ and instead ask for four reads on what’s already happening inside the company: how AI is being used across teams, where output has changed, where the risk envelope has moved, and whether the working chart of the company has shifted. Those four reads still hold. The agentic shift adds a layer they don’t cover, because the company has stopped being the actor for some set of decisions, and the board pack hasn’t been built to read that layer yet.
Agentic deployments are different from AI as a tool on one axis. The system initiates the action, not just the recommendation. A human’s no longer the first reader and the final mover. In most of the deployments running right now, that change has happened workflow by workflow, vendor by vendor, with no central list and no defined revisit. Ask a CEO which decisions the company has already authorized a system to make without human review, and the answer in 2026 is usually some version of ‘I’d have to check’.
The inventory is the missing piece. Boards already keep registers for things the company is exposed to even when leadership isn’t watching, like the related party transactions log or the litigation register. The agentic register is the same kind of thing. It names every decision the company has let a system make without a human in the loop, the workflow it sits inside, the owner, what the system’s allowed to act on, and what triggers a human stepping back in. In most companies the answer to ‘do we have one’ is no. The board’s job is to ask, every quarter, whether one exists and whether it’s current, not to write the register itself. Public board oversight guidance from places like Nasdaq has been pointing the same direction for AI generally, asking who’s accountable and how that accountability shows up in reporting and decision logs. The agentic version of the question is sharper, because actions taken by systems are harder to reconstruct after the fact than recommendations made by them.
The next read on that register is what undoing a wrong call costs. A duplicate refund is cheap to reverse, but an agent that approves a non conforming supplier shipment or files a regulatory return on a wrong number is something the company will spend a year explaining. Reversibility is the property that separates safe agentic deployment from unsafe, and it isn’t the same thing as accuracy. Two systems at the same accuracy can sit in different companies depending on what an error costs to undo: ten thousand monthly decisions at fifty thousand dollars an undo is a different business from the same volume at fifty dollars an undo, and the boards overseeing them have different jobs even when the dashboards look the same. The board’s question is whether management has classified the register by undo cost, and whether the company’s moving up that ladder on purpose or by drift. The second year curve of any agentic program is when the harder decisions get handed over.
The audit trail is the next question, and most companies haven’t faced it yet. When an agent acts, then another agent acts on its output, then a third system commits the result, credentials usually get reissued at each hop, and the line from the original human authorization to the final action breaks. A regulator or counsel asking the company to reconstruct that line in 2026 will, in most companies, get an apology. The way to think about it is that the company is running without a general ledger for one slice of decisions. Nobody calls it that yet, and the cost of not calling it that grows fastest in regulated functions.
The earlier memo on the four reads rested on the idea that a human was the first reader of model output, and could feel a model getting worse. Agentic systems take that reader away. Drift is no longer a thing a senior person notices in an inbox. It shows up later, when a customer complains or a regulator writes in, sometimes not until the year end audit. Accuracy is a vendor question. The board’s question, the one without a vendor to ask, is who in the company is responsible for noticing when accuracy has changed and how often they look. If the answer’s ‘the vendor will tell us’, the company’s outsourced its read on itself, which the four reads memo already pushed back on in a different context.
The regulatory surface a board oversees in this region is going to move with the program. Eleven days after the federal announcement, His Highness Sheikh Hamdan bin Mohammed bin Rashid Al Maktoum, Crown Prince of Dubai, Deputy Prime Minister and Minister of Defence of the UAE, launched a parallel program pulling Dubai’s private sector toward agentic AI over the same two years through Dubai Chamber training tracks and dedicated incubators and funds for agentic AI companies. The signal has already moved past public sector services and into private sector operations. As both sides start running agentic at scale, the rules for how companies interface with each other and with the public infrastructure will get specified, and probably faster than the usual timeline. Procurement and compliance questions that have been informal will become contractual, including the identity and credentialing questions agentic systems force. The labor pool will move with it. Every federal employee going through AI training means that inside twenty four months the country’s mid career operating bench is the most AI fluent it’s ever been, and the expectation a junior brings into a private sector role about how their work is done changes with that. The companion memo on the hiring bar after AI walks through that side of it. Neither shift is the board’s direct responsibility. Both shape the environment the board is overseeing.
The lesson sits one level above the federal program. A government has made a public commitment to a posture most private companies haven’t decided whether they’re taking, and the board pack hasn’t been built to read that posture inside their own company yet. The four reads from the previous memo are still the spine. What gets added on top is what’s been delegated, what undo costs, whether the chain can be reconstructed, and who notices drift. Boards that put those four into the pack now will know, eighteen months from now, where their company actually stands. Boards that wait for someone outside to ask first will spend that period catching up.