AI for UAE government: running it day to day.

What AI for UAE government looks like after the launch event.

The strategy layer is genuinely unmatched. The federal portal runs U-Ask, a generative-AI assistant for government services. Dubai appointed 22 Chief AI Officers across its government entities in 2024 under its Universal Blueprint for AI. Abu Dhabi's Government Digital Strategy 2025-2027 commits AED 13 billion toward an AI-native government, with over 200 AI solutions planned across services.

Inside the average department, the systems that run every day are less photogenic: bilingual correspondence drafting, summarization of long case files, retrieval over the entity's own policies, minutes and recurring reports. These are the systems that quietly succeed or quietly rot, and nobody holds a launch event for either outcome.

The difference between the two outcomes is rarely the model. It is whether anyone owns the unglamorous routines: the classification gate, the review step, the terminology list, the log. Government raises the stakes on each, because the failure mode is a wrong answer in an official letter with a ministry crest on it.

The rules you operate under, emirate by emirate.

A recurring misunderstanding, especially in teams arriving from the private sector: the UAE's federal data protection law is not your frame. The PDPL (Federal Decree-Law No. 45 of 2021) excludes government entities and government data from its scope. What binds you instead is a stack of emirate-level and entity-level rules, and most of them are operational documents, not abstract principles.

Five routines that keep a system defensible.

1. 01:

   Gate on data class at the point of use

   The classification decision made during procurement is worthless if a case file can still be pasted into the wrong tool on a Tuesday. The gate has to live where the work happens: retrieval connected only to approved repositories, and a short written rule per workflow saying what may and may not go in.

2. 02:

   Name one reviewer per workflow

   The AI Charter's human-oversight principle becomes real as a name, not a committee. One person per workflow signs off anything that leaves the entity: letters, decisions, published answers. Internal drafts can run lighter. If you cannot name the reviewer, the workflow is not ready for official output.

3. 03:

   Decide disclosure conventions once

   Agree in writing when AI assistance is flagged: generally not for internal drafting a human has reviewed and owns, and explicitly for anything citizen-facing that answers on its own, the way U-Ask is clearly presented as an assistant. What you want to avoid is each department improvising its own answer under pressure.

4. 04:

   Log what an auditor will ask for

   Per run: the input, the model and version, which source documents retrieval pulled, who reviewed, and the final output. Most of it can be captured automatically. The day a decision is challenged, this log is the difference between an explanation and an incident.

5. 05:

   Re-evaluate on a schedule, not on a complaint

   Keep the evaluation set from your original model selection and re-run it quarterly and on every model update. Quality drift is silent, and staff stop reporting it; they just stop using the tool, which looks like success in the complaint log and failure in the usage log.

Bilingual output is an operations problem, not a launch problem.

Arabic handling gets evaluated carefully once, during selection, and then degrades unattended. Government Arabic is formal Modern Standard Arabic with fixed honorifics, letter conventions, and official translations for titles and program names, and all of those change: entities get renamed, programs launch, honorifics shift with appointments. The terminology list you injected into prompts at launch is a living document, and someone has to own it.

The open-weight Arabic models matter here for an operational reason, not a patriotic one: Jais (from G42's Inception with MBZUAI) and the Falcon family (from Abu Dhabi's Technology Innovation Institute) can run on sovereign infrastructure, which is what the stricter data classes require. Mixed estates are normal, and each part needs its own evaluation set and re-run schedule.

The standing checks worth keeping:

• Terminology drift. Sample recent outputs monthly against the official translation list. The model inventing its own rendering of a ministry's name is the single most common bilingual failure.
• Register drift. Have one of the people who writes official correspondence read ten recent outputs. Colloquial drift is obvious to them in minutes and invisible to a dashboard.
• Mixed-direction formatting. Numbers, dates, and English entity names inside Arabic text still break quietly after model updates. Keep three or four known-ugly documents in the evaluation set.
• Translation symmetry. Check both directions. Systems tuned on English-to-Arabic drafting are often noticeably worse in reverse, and half of real correspondence runs in reverse.

When the model changes under you.

Commercial models deprecate on the vendor's calendar, not the government's. A model that passed your Arabic evaluation and your security review can be retired with months of notice, and the replacement is not grandfathered into either. Treat model change as a standing category of work: pin versions where the platform allows it, watch deprecation notices, and budget a re-evaluation and, where accreditation names the model, a re-approval cycle before the forced migration date.

The same discipline applies to the humans. Every model change shifts tone and failure modes slightly, and staff trust, once lost, is expensive to rebuild. Announce changes, re-run the evaluation openly, and show the results. The mechanics of carrying a large organization through tooling changes are the same ones we describe in AI change management for large teams; government entities are large teams with sharper consequences.

Reporting upward without inflating.

Every entity operates under pressure to show AI progress, and the temptation is to report activity: prompts issued, staff trained, tools deployed. The number that survives scrutiny is per-workflow and boring: correspondence turnaround before and after, case-file summary time, retrieval answers accepted without rework. Baseline it honestly at launch and report the same number every quarter. The method in our workflow audit guide produces exactly this shape of baseline, which is one reason to audit before you build.

This whole layer (gates, reviews, logs, evaluations, model migrations, reporting) is what we mean by operations, and it is the part entities most often under-resource, because the build got the budget and the running did not. It is also precisely what our managed AI operations service carries: keeping accredited systems inside their rules, evaluated, and reported on.

Do it yourself, or bring us in.

Everything above is runnable internally: one owner per workflow, a classification gate, a named reviewer, a living terminology list, a log, and a quarterly evaluation calendar. The honest failure mode is not capability, it is attention. These routines are nobody's main job, so they hold for a quarter and then quietly stop.

We do this work on-site in Dubai and Abu Dhabi, and we have done it inside government rather than only written about it. If the system is not built yet, a fixed-scope audit is the right first contract; if it is live, the conversation starts with what you are running and what rules it answers to.

Frequently asked questions.

Keep reading.