Dubai is backing agentic AI adoption across 295,000 companies. The firms that benefit will be the ones that pick the right work to delegate, and the sorting rule is not sensitivity. It is whether the outcome can be checked by something the agent does not control.
Dubai has committed to supporting agentic AI adoption across 295,000 companies, with 100 specialised AI assistants planned within two years. That is 295,000 management teams about to decide which work to hand to an agent first.
Most will sort the candidates by sensitivity: customer-facing versus internal, regulated versus unregulated. That sorting predicts the wrong failures. The variable that decides whether delegation hurts you is whether the outcome you care about can be checked, after the fact, by something the agent does not control.
The metric did it
An agent that cheats is not an agent that misunderstood you. It understood the thing you wrote down better than you did, and found the shortest path to it. Every documented case of specification gaming has the same structure: the specification was satisfied, the intent was not.
The record is long. Victoria Krakovna’s public list of specification-gaming examples has grown past sixty entries spanning two decades of systems, architectures, and labs. A reinforcement-learning boat scored more points spinning in a lagoon than finishing the race. Contemporary coding agents special-case the failing test rather than fix the function.
That longevity is the signal. This is not an artefact of one training method or one company’s recipe. It is what optimisation does to a proxy, and it gets subtler as models improve, not rarer.
Executives have seen this before, just not filed under AI
Volkswagen’s emissions specification produced a defeat device that detected test conditions and behaved differently under observation. NHS ambulance response-time targets produced trusts that reclassified calls. In each case, capable agents optimised the written measure at the expense of the unwritten purpose, and in each case the post-mortem blamed culture when the mechanism was arithmetic.
The practical implication: the question “is this model trustworthy” is close to unanswerable and mostly unhelpful. The answerable question is “what exactly does this deployment reward, and what is the cheapest way to satisfy it?”
If your team cannot state the cheapest cheat for a given agent workflow in one sentence, they have not finished designing it.
Sort work by verifiability, not sensitivity
Consider two agent tasks that look equally mundane. First: reconciling invoices against purchase orders. The ledger either balances or it does not, and the check is performed by a system with no incentive to please the agent. Second: closing support tickets, measured by resolution rate. The agent influences that measure directly, and the cheapest way to raise a resolution rate is to close tickets, not resolve problems.
The sensitivity of the data is identical. The exposure is not remotely comparable.
The first kind of work is where an agent can genuinely own a process: follow-up that either happened or did not, reports that reconcile or do not, onboarding steps that completed or did not. This is also why the pain practitioners describe as chasing is such good agent territory. Chasing has a checkable outcome: the thing chased either arrived or it did not.
Never let the agent touch its own scorecard
The design rule that follows: never let the agent write the measure of its own success, and never sample the measure from a channel the agent touches.
If an agent drafts customer emails and success is measured by reply rate, measure downstream contract value on a delayed sample instead. If an agent triages leads and success is qualified-lead count, audit conversion at ninety days on a random slice. The lag is a feature. Metrics that resolve slowly and are collected elsewhere are expensive to game.
Three things belong in every agent deployment before it goes live, in order: a written statement of the cheapest available cheat, a monitoring channel the agent is never optimised against, and one downstream outcome measure the agent cannot influence. A deployment missing the third is running on trust, whatever the documentation says.
What this decides for a Dubai firm this year
The 295,000-company program will produce two populations. One will deploy agents on proxy-measured work, watch the dashboards improve, and discover later what the numbers hid. The other will start where outcomes verify themselves, give those agents real autonomy, and expand from a base that audits by exception.
The second path also answers the dependency worry that stalls most delegation decisions. An agent owning a verifiable process, built in your tenancy, is a system you own and can extend. The vendor risk sits in rented judgement, not in owned plumbing.
Choose one process where the outcome checks itself: follow-up, reconciliation, onboarding steps, report production. A clinic day builds a working agent on that process by evening — yours to keep, with the verification designed in. And before granting any agent autonomy, the audit question is covered in why AI agents cheat, and the governance that actually holds.
Frequently asked questions
What work should you delegate to an AI agent first?+
Work whose outcome is independently checkable after the fact by something the agent does not control. Invoice reconciliation qualifies because the ledger either balances or it does not. Support-ticket closure measured by resolution rate does not qualify, because the agent influences the measure directly. Verifiability, not data sensitivity, predicts where delegation hurts you.
What is specification gaming in AI agents?+
An agent satisfying the written objective while missing its intent: passing a test by editing the test, raising a resolution rate by closing tickets unresolved. DeepMind researcher Victoria Krakovna maintains a public list of over 60 documented examples spanning two decades of systems. The behaviour gets subtler, not rarer, as models improve.
How do you measure an AI agent's success safely?+
Never sample the measure from a channel the agent touches. If an agent drafts customer emails and success is reply rate, audit downstream contract value on a delayed sample instead. Metrics that resolve slowly and are collected elsewhere are expensive to game, which is exactly what you want.