Any clinical AI you deploy in Abu Dhabi now needs documented explainability, privacy controls, and a named line of accountability.
The DOH Responsible AI Standard (V1, 2025) makes explainability, data-privacy controls and named accountability a condition of deploying clinical AI.
What a board member needs before the next meeting on this.
- The line is drawn at deployment. A pilot in one clinic is tolerated; a tool your clinicians rely on across sites needs a documented governance file.
- Explainability is a procurement filter. A black-box model that cannot show its reasoning is now a vendor you cannot defend — ask for the file before you sign.
- Vendor data handling is your liability. DPIA, contractual audit rights, and knowing where patient data goes stay with you, not the vendor.
- Accountability means one named person per tool. If you cannot name who signs for the ED model today, the deployment is not compliant.
Before you deploy another clinical AI tool in Abu Dhabi, someone has to sign for it. The DOH Responsible AI Standard (V1, 2025) makes documented explainability, data-privacy controls, and a named line of accountability a condition of compliant deployment. It layers on the DOH's 2018 policy on AI in the healthcare sector and its six principles: transparency, user assistance, safety and security, privacy, ethics, and accountability. A triage algorithm, an imaging read, an ambient scribe in your clinics; each now carries a governance file, not just a licence.
The pilot-to-production gap is where groups get caught
It is easy to trial an AI scribe in one clinic or let a radiology vendor run an imaging model in the background. It is a different thing to deploy it across a hospital group as part of the care your clinicians rely on. The Responsible AI Standard draws its line at deployment. Once a tool informs clinical decisions on real patients, it needs a documented governance file: what it does, how it reaches its output, what data it touches, and who is answerable when it is wrong. A tool your teams are already quietly using without that file is a compliance gap you may not have on your risk register.
Explainability is now a deployment condition, not a nice-to-have
The 2025 standard fleshes out transparency and explainability requirements on top of the 2018 principles. In plain terms, a clinician and a regulator have to be able to see why the system produced a given recommendation and to challenge it. A model that returns a triage priority or a flagged nodule with no traceable reasoning is hard to defend under this standard. That narrows which vendors you can deploy, because a black box that cannot explain itself is now a procurement risk, not only a clinical one. Ask for the explainability documentation before you sign, not after an incident.
The vendor's data handling is your liability
An ambient scribe records the consultation. An imaging model ingests the study. Both touch protected health information, and the accountability does not transfer to the vendor because you bought a licence. You are expected to run a data protection impact assessment, hold contractual audit rights over the third party, and know where the data goes and how it is secured. Malaffi participation and DOH privacy expectations already bind you. An AI vendor that quietly trains on your patients' data, or stores it outside approved bounds, becomes your exposure, in your name, at your next accreditation review.
Accountability means a named person, not a committee
This is where good intentions usually fail. Someone has to own each deployed AI tool, sign that it is fit for use, monitor its performance, and answer for its errors. The 2018 policy already required governance structures, audits, and DOH reporting; the 2025 standard sharpens the expectation of a clear line of responsibility and a route for end-users to appeal an AI-driven decision. If you cannot name the person accountable for the model in your emergency department today, you do not yet have a compliant deployment, whatever the vendor's brochure says.
Four questions worth asking this month
- Which AI tools are running anywhere in the group today — including the ones clinicians adopted without asking?
- For each, can we produce the explainability and privacy file a DOH reviewer would ask for first?
- Who is the named accountable owner for the highest-risk model in use?
- Does our vendor contract give us audit rights over where patient data goes?
For a group running several sites, the fix is a single AI governance framework: an inventory of every tool in use, an explainability and privacy file for each, a named owner, and a review cadence that sits inside your quality and accreditation system rather than beside it. Build it once and every future deployment slots into it, which turns AI governance from a drag on adoption into the thing that lets you adopt safely and at pace. The alternative is discovering the gap during an accreditation review or after a patient-safety event, when the file you should have written a year ago is the first document the regulator asks to see. If your group is deploying clinical AI ahead of the governance to hold it, that is where a conversation with an Avior principal about operational excellence begins.