Decision Support, Not Decision Making
Co-Founder & CEO of CLRT
You run a hospital group or a clinic network in the Emirates, and in the past year you have been sold clinical AI by everyone: the surgical platform, the imaging vendor, the ambient scribe, the predictive risk model. The pitch is that the data is finally there, and it is. Abu Dhabi's health information exchange, Malaffi, reported 3.5 billion clinical records across 12.7 million patient profiles and 3,072 connected facilities in August 2025. Dubai's exchange, NABIDH, unifies more than 9.53 million patient records across more than 1,500 facilities. No health system in the region, and few anywhere, has a spine like it. Now try to find one published clinical outcome from AI running on that spine. Not a partnership, not a principle, an outcome. The search comes back empty, and the reason it does tells you where to point your own programme.
Start with the spine, because its growth is the argument. In August 2022 the Department of Health Abu Dhabi and Malaffi launched an AI-powered Patient Risk Profile, displaying for each patient a risk score for diabetes, heart failure, kidney disease, hypertension, heart attack and stroke. At launch the exchange held 900 million records for 7 million patients across 2,000 facilities. By September 2022 it passed 1 billion records. By October 2023, 2 billion records and 7.9 million patients across 2,736 facilities. By August 2025, 3.5 billion records, 12.7 million patient profiles and 3,072 facilities, with the risk profile still listed among the platform's capabilities. The launch release published no accuracy figure for the model, and no milestone release since has published an outcome: no readmission avoided, no diagnosis brought forward, no cost per case moved. Three years of a predictive model in production, nearly four times the records beneath it, and the only numbers that grew were the numbers describing the data.
Then look at the flagship. On 22 May 2026 the Department of Health announced an emirate-wide intelligent surgical network on Johnson and Johnson's Polyphonic platform, built with AWS, NVIDIA and Core42, that will connect operating rooms across Cleveland Clinic Abu Dhabi, PureHealth, Mediclinic and NMC. The release names four technology partners and four hospital groups. It publishes no count of operating rooms, no count of procedures and no outcome measure of any kind. What it does publish, in the regulator's own words as reported by Gulf News that day, is a doctrine. The DoH undersecretary put it in one sentence: "AI here is decision support, not decision making. The surgeon is always in charge." All surgical data, she added, stays within Abu Dhabi's sovereign infrastructure, anonymised and governed by DoH. That is not a vendor's marketing line. It is the regulator drawing the boundary every operator in the emirate will be held to.
The regulator's behaviour says the same thing more quietly than its words. On 21 July 2026 the Department of Health Abu Dhabi and the Abu Dhabi Public Health Centre, with Abu Dhabi Health Data Services, announced the completion of twelve disease registries built on Malaffi data, from mental health and obesity to cancer, rare diseases and multiple sclerosis, and placed AI in the next phase, to be incorporated later as part of the Population Health Intelligence programme. Registries first, then the model. In Dubai, the first AI the Dubai Health Authority announced inside NABIDH, in April 2025, was not a clinician at all. Built with Imprivata, it monitors and prevents unauthorised access to patient records, on a platform where 82 percent of the medical workforce is active. The region's two most data-rich health authorities, with every incentive to publish a clinical win, sequenced AI behind the registries and deployed it first as an auditor. That is a signal, not timidity.
Put the three together and the shape is unmistakable. The spine has run years ahead of the workflow. The constraint on clinical AI in the UAE is not data, which exists at a scale few countries can match. It is not compute and it is not policy, since the DoH's Policy on Use of Artificial Intelligence in the Healthcare Sector dates from 2018 and remains the only AI policy on the regulator's policies page. The constraint is the layer between the record and the decision: who receives a risk flag, what they do with it, whether the action can be checked against the record before it ships, who is accountable when the model is confidently wrong. That layer lives inside the hospital, in its workflows, and no exchange or partnership can build it from outside. Which is why the measurable work sits there, in the administrative and decision-support workflows around the record rather than in the theatre.
For an operator this settles the order of the programme. The theatre is where the announcements are, and also where measurement is hardest, liability highest, and the model's role already fixed at support. The workflows around the record are the opposite on every count. Prior authorisation, coding, discharge and referral paperwork, access monitoring, the triage of risk flags into a queue a clinician signs: each produces an output that can be checked against the record, each has a cycle time and an error rate a board can read within a quarter, and each leaves the clinical decision precisely where the undersecretary put it. The case for separating clinical from clerical work used to rest on principle. Now the regulator has said it, and the growth figures show how far the data has outrun the outcome. The first published number from AI in an Emirati hospital will almost certainly come from a workflow nobody put on a stage.
The spine ran three years ahead of the workflow. The gap was never the data.
A deeper dive
The mechanism behind the missing outcome is worth understanding, because it repeats inside every hospital that buys a model. A risk score is not an intervention. When the Patient Risk Profile flags a patient for heart failure, nothing has happened to that patient until a clinic receives the flag, decides what it means for this person, acts, and later measures whether the action changed anything. Each of those four steps is a workflow, and every one of them lives inside the provider, not inside the exchange. The exchange can grow from 900 million records to 3.5 billion without a single one of those workflows existing, which is precisely what the growth series shows. An outcome is produced by the loop, and the loop is the unglamorous part: routing the flag to the right person, defining what a wrong flag costs, checking the model's output against the record before anyone acts on it, and keeping an account of what happened next. Most clinical-AI procurement buys the score and assumes the loop. The loop is the product.
The second-order trap is that the regulator's doctrine, correctly applied, can erase the saving it was meant to protect. Decision support means a clinician reviews the model's output before acting, every time. If the output arrives as one more item in an already crowded screen, review becomes a click, and the arrangement quietly becomes decision making with a signature on it, which is the reversal the doctrine exists to prevent. If the review is genuine, the minutes the model saved are re-spent verifying it, and the dashboard reports utilisation while the clinician's day is no shorter. Both failures are invisible from the vendor's metrics. The way out is to engineer the review as a workflow with its own instrumentation: what the clinician sees, what they can check it against, how long the check takes, and what the system records when they override it. That is a design and governance problem long before it is a modelling problem, and it is the work that decides whether an operator's first AI number is an outcome or an anecdote.
Work with CLRT
Deciding where AI sits in relation to the record and the clinician, and building the loop that turns a model's output into a number a board can read, is the work CLRT does with healthcare operators in the Emirates. We start with the question the regulator has already answered for the theatre and nobody has answered for your workflows: which ones can be checked, measured and governed now. If you want that map before the next vendor arrives, run the CLRT Ascent diagnostic, or bring us your roadmap and we will draw it with you.
Vishal Sachar is the Co-Founder and CEO of CLRT, where he helps UAE businesses make sense of applied agentic AI and put it to work. He writes on agentic systems, AI governance, and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.
