Xami Workers · Compliance · Explainer

An AI worker runs every written check, on every item, every time, and records that it did. It hands a person a short list to scan and approve, exceptions first. The person keeps the judgement and the sign-off. The business carries less risk and gets more of what compliance pays back: claims and invoices accepted the first time.

Most compliance failures are not failures of judgement. Nobody decided to break a rule. A check was skipped on a busy afternoon, or remembered a day late, or done for nine items out of ten. This piece is about that kind of failure, and about the narrow job an AI worker can do against it without anyone giving up the decision.

The examples below come from UAE healthcare and UAE e-invoicing, because that is where our own pages have read the regulators’ documents. Each rule is cited under Sources, with its date. The pattern applies to any business that runs on written rules.

Compliance fails on the checks nobody remembers

Ask a compliance lead where things go wrong and the answer is rarely a hard call. It is the dull, repeated step that has to happen on every patient, every claim or every invoice, and quietly doesn’t:

• A reassessment that wasn’t booked by day 42. The patient got better, but nobody measured it on the day the rule counts.
• An approval that expired. The patient came in for a session the insurer’s authorisation no longer covered.
• A claim missing a field. The right treatment, refused because one licence number was not on the line.
• An e-invoice without a buyer TRN. The sale happened; the tax invoice, legally, did not.

None of these needs expertise. Each needs someone to check the same thing every time, under load, without being told. That is the part people are worst at, through no fault of their own, and it is the part a machine does not get tired of.

What an AI worker does instead

An AI worker is given a role: the queue it owns, the rules it checks, the actions it may take, and the point where a person takes over. Inside that role it does four things.

1. Runs every written check, on every item. Not a sample at month end. Every patient on a clock, every open approval, every unpaid claim, every invoice before it goes out.
2. Records that it did. Each check leaves a line: what was checked, against which rule, what it found, and when.
3. Hands a person a short list. The items that passed are summarised. The exceptions come first, each with the evidence: the field that is missing, the date that is close, the record it read.
4. Waits for the sign-off. The person approves, edits or rejects. The worker acts on that, and the approval goes on the record with the person’s name.

The tedious work comes off people’s desks. The judgement does not. Instead of rebuilding a tracker by hand, the person reads the exceptions and decides what to do about them.

“AI does not forget”, said precisely

It is tempting to say an AI worker never forgets. The accurate version is narrower, and more useful. The rules are written into the worker’s role and into the system of record: the approval’s end date is a field, the day-42 clock starts when the plan is made, the buyer’s TRN is checked against the customer record. So the worker never skips a check because it was busy, or because it was Friday, or because the person who usually does it was on leave.

That is not the same as never being wrong. The model can misread a scanned note, match the wrong payment, or misjudge which rule applies to an unusual case. That is exactly why a person approves the result, and why the worker shows its evidence with every item instead of only a verdict. A check you cannot see is a check you have to trust. A check that shows its working is one a person can approve in seconds, or catch.

Assist before autonomy

The rule we work to is assist before autonomy. Every action starts in Assist: the worker prepares the work and a person approves it. An action runs on its own only when its record is good and the person who owns the role signs it off, one action at a time. After an error, a complaint or a dip in its numbers, it drops back to Assist. Some actions never leave Assist at all: anything clinical, anything that posts money, write-offs and legal notices.

In compliance work that means the worker usually starts as the one who prepares, checks and drafts, and a person signs. A reminder to a patient that an approval is running out may be promoted early. A letter to an insurer stays with a person until it has earned its way out. The client decides which actions, if any, the worker may take alone. The pattern is set out on our Xami Workers page.

Four worked examples

Each example says what the worker checks, what the person sees and approves, and what is recorded. The rules are summarised from the regulators’ own documents, through the Xamun pages that read them.

1. An Abu Dhabi rehab clinic and Jawda

Since the fourth quarter of 2026, outpatient rehab providers licensed by the Department of Health – Abu Dhabi report Jawda indicators every quarter. A patient counts only with a baseline written before therapy and a reassessment on the same measure at a fixed point, day 42 for most indicators. A patient whose reassessment was missed stays in the denominator and counts as not improved. In ClinicOS the check runs on every patient on a clock: baseline recorded, cluster tagged, reassessment booked before the due date. The physiotherapist sees who is due or overdue this week, with the patients pulling the figure down listed first, and confirms the bookings; Sam on recalls calls back missed visits and offers new times. The record holds each baseline, each reassessment with its date, and each recorded reason a patient was left out. Our Position Paper No. 10 sets out each indicator.

2. Prior approvals and extensions

In Abu Dhabi, a physiotherapy course usually runs on an insurer’s authorisation through the DoH Shafafiya system, with an authorised number of sessions and an end date. When those will not cover the plan of care, the clinic sends a request of type Extension. For rehabilitation therapy the Shafafiya dictionary names the ordering physician as the activity clinician, so the referring doctor’s licence belongs on the line. The worker checks sessions approved, used and left against the end date on every open approval, and that the claim line carries the right clinician and a signed note. The person sees the approvals running low, and a drafted extension built from the record: measures against baseline and goal, attendance, home practice. The physiotherapist adds a sentence and signs; Sam on claims and approvals drafts, and every letter to an insurer stays with a person until it has been promoted. The record keeps the draft, the edits and who signed. The detail is in our explainer on physiotherapy approvals in Abu Dhabi.

3. Claims follow-up

Every partly or fully denied claim in Abu Dhabi carries a denial code on the remittance advice, and the DoH claims standard sets a 45-day payment period. In Dubai, according to an Al Tamimi & Company summary of the 2025 claims directive, payment is due within 45 days for an initial claim and 30 days for a resubmission, and a disputed claim can be resubmitted twice. Sam on clinic finance follows up every rejected and unpaid claim, with its denial code and how long it has been open against those clocks, and matches insurer payments to the claims they pay. The person sees the claims near a deadline first, each with its denial code and the record behind it, and decides what to resubmit, what to appeal and what to write off. Anything sent to an insurer and anything that posts money stays with a person. The record shows every claim chased, when, and what was decided. The UAE rules are summarised on our InsureOS page.

4. UAE e-invoicing mandatory fields

From 1 January 2027, UAE businesses with revenue of AED 50 million or more issue B2B and B2G invoices and credit notes as PINT AE e-invoices through an Accredited Service Provider; businesses below that go live on 1 July 2027. An invoice that is missing a buyer TRN for a VAT-registered customer, or leaves a mandatory field empty, is rejected in validation and has not been issued. Cabinet Resolution No. 106 of 2025 fines each e-invoice not issued or sent on time AED 100, capped at AED 5,000 a month. The check runs on every invoice before it goes out: buyer TRN present and validated, a VAT category on every line, the right treatment for a designated-zone or export sale. The person sees only the invoices that fail, with the field that failed. The record shows what was checked and what was fixed. We have no Sam role for invoicing; in SalesOrderOS the same check runs when the order is taken. One limit matters here: an invoice that is well formed but wrong, the wrong customer or a price nobody agreed, can pass validation. A rule catches what is missing. A person catches what is wrong. See what happens when a UAE e-invoice is rejected.

Two kinds of benefit

Less risk. Fewer missed steps, because no check depends on someone remembering it. And a full audit trail, because every check, every approval and every hand-off is written down as it happens, not reconstructed when the auditor asks.

The upside of compliance itself. Compliance is usually costed as overhead. Done on every item, it pays back. Claims accepted the first time are paid sooner. Jawda figures improve when patients were measured on time, without anyone treating differently. Invoices accepted by the provider count as issued, and the cash behind them moves.

What UAE health regulators expect of AI here

An AI worker that touches patient or claims data in the UAE is itself under rules. None of them was written for an AI worker, but read together they describe the same pattern: a person supervises, can override, and the work is disclosed and logged. Our Papers No. 11 and No. 12 read the documents in full; in short:

• Dubai: the DHA policy on AI in healthcare (DHA/HISHD/PP-09, dated 1 July 2021). AI functions must be disclosed, along with the role the healthcare professional plays in the final decision. The AI must be supervised and monitored by its users, and where it can affect people’s lives significantly, designated users must be able to override or reverse it. Failures must be traceable, and incidents are reported to DHA. Its listed revision date, 1 July 2026, has passed; check for a newer version.
• Abu Dhabi: the DoH Responsible AI Standard (DoH/ST/DDGO/RAI/V1/2025, October 2025). It applies to every DoH-licensed entity that develops, buys or deploys AI, for clinical, financial and administrative work alike. On our reading, claims and compliance work is in scope. AI systems must work under human oversight with clear escalation, and clinicians keep the final decision. Patients and clinicians must be told when AI contributes to a clinical or operational decision.
• Disclosure on calls and messages. Dubai’s AI Ethics Guidelines (Smart Dubai, January 2019, non-binding) say an AI that could pass for a person should first say it is AI. Sam does, at the start of every call and message.

Assist before autonomy answers the supervision duty in the work itself: staff approve each action as it happens rather than reviewing a sample later, and the log shows who approved what.

What to ask before you let an AI worker near compliance work

1. Where are the rules it checks written down, and who can change them? You should be able to read the role and audit against it.
2. Does it show its evidence for every item, exceptions first? A verdict with no working is something you have to trust, not approve.
3. Which actions does it take alone, who promoted them, and what sends them back to Assist?
4. Is every check, approval and hand-off logged in a record you own, that nobody can quietly edit?
5. Can a named person override or reverse anything it did?
6. Where do the model, the logs and the backups run, and is your data used to train the model? For UAE health data, the answer has to be in the UAE, in the contract.

Where Xamun fits

Sam is our AI worker; every Xami worker is called Sam, and the role is what changes. Sam works inside the system its role belongs to, such as ClinicOS for a clinic or InsureOS for claims, starts every action in Assist, and writes each check and approval to records the client owns. Sam is in early access; we start with one role, defined with the person who owns it.

Confirm before you rely on this. This is Xamun’s reading of the documents listed below, not legal or tax advice, and not endorsed by any regulator. Confirm your own obligations with DHA, DoH or MOHAP as relevant, the Ministry of Finance or Federal Tax Authority for e-invoicing, a lawyer, your tax adviser or your insurer.

---

Sources

1. Dubai Health Authority, Artificial Intelligence in the Healthcare, Policy DHA/HISHD/PP-09, version 1, dated 1 July 2021 (revision date 1 July 2026): disclosure (4.3.2), supervision and override (4.4.1, 4.4.2), traceability (4.3.1), incidents (4.8.1). dha.gov.ae. Read and summarised in our Position Paper No. 11.
2. Department of Health – Abu Dhabi, Responsible Artificial Intelligence (AI) Standard, DoH/ST/DDGO/RAI/V1/2025, published and effective October 2025 (revision date October 2026): scope (1.2), human oversight and the clinician’s final decision (3.1.1.2, 3.1.7.2.2), telling patients and clinicians (3.1.6.1). doh.gov.ae. As read in our Position Paper No. 12.
3. Smart Dubai, AI Ethics Principles & Guidelines, January 2019, non-binding: guideline 1.3.2.2. digitaldubai.ae.
4. Department of Health – Abu Dhabi, Outpatient Rehabilitation Service Jawda Guidance, Version 1, issued May 2026, effective Q4 2026: baseline, day-42 reassessment, missed reassessments in the denominator. doh.gov.ae.
5. Health Authority – Abu Dhabi (now DoH), HAAD Health Insurance Claims Adjudication Standard, HSF/CA/1.0, December 2010: denial codes on remittance, the 45-day payment period. doh.gov.ae.
6. Department of Health – Abu Dhabi, Shafafiya dictionary, CommonTypes data definitions: authorisation type Extension, authorised quantity and dates, the activity-clinician rule for rehabilitation therapy. shafafiyaportal.doh.gov.ae. Undated online schema.
7. Al Tamimi & Company, “Dubai issues comprehensive overhaul of health insurance claims management framework”, 2025. Law-firm summary of the Dubai Health Insurance Corporation’s Claims Management Policy Directive, effective 16 November 2025: payment and resubmission periods. tamimi.com.
8. UAE Ministry of Finance, “Ministry of Finance announces targeted amendments to eInvoicing system decisions”, 10 May 2026: Ministerial Decision No. 244 of 2025 amended, ASP appointment for businesses above AED 50 million moved to 30 October 2026, implementation no later than 1 January 2027. mof.gov.ae.
9. UAE Ministry of Finance, announcement of Cabinet Resolution No. 106 of 2025 on administrative fines for the electronic invoicing system, 8 December 2025: AED 100 per e-invoice or credit note not issued or sent on time, capped at AED 5,000 a month. mof.gov.ae.