AI for UAE Insurers & TPAs
Classify claims, route documents, and surface answers from policy wording — without the manual sorting that slows every claim down.
- Auto-classify and route incoming claims and correspondence at 95% accuracy
- Cut manual document sorting by 10x and assign cases 3x faster
- RAG over policy wordings and regulations with cited answers
- On-premise deployment keeps policyholder data inside your network
- Aligned with CBUAE insurance supervision and UAE PDPL
How We Help
Claims Triage & Routing
- Classifies incoming claims and documents by type and urgency
- Routes each case to the right handler automatically
- Confidence scoring with human review for edge cases
- Integrates with your claims and policy-admin systems
Policy & Document Q&A
- RAG over policy wordings, endorsements, and regulations
- Handlers get cited answers instead of reading full policies
- Works with PDF, scanned documents, and legacy exports
- Runs on-premise — no policyholder data leaves your network
Document Data Extraction
- Extracts structured fields from forms, invoices, and reports
- Validates against policy and regulatory rules
- Flags missing or inconsistent information before processing
- Feeds clean data straight into your core systems
Results
document classification accuracy (my legal/insurance deployment)
reduction in manual document sorting time
faster case assignment
The Monday after go-live
It's Monday and the weekend's backlog has already sorted itself. Every claim and document that came in is classified, scored, and routed to the right handler — the straightforward ones queued for fast-track, the complex ones flagged with the policy clause that matters. A handler opens a disputed claim and asks the assistant what the wording says about it; the answer comes back cited to the exact endorsement. The pile of unsorted correspondence that used to define Monday mornings is simply gone.
Where insurers actually lose money: rejections vs. denials
Walk into any UAE health insurer or TPA and the leak is in the same place: the gap between what's submitted and what's actually paid, on time. Two different failures live in that gap, and conflating them is where a lot of automation projects go wrong.
A rejection is a pre-adjudication failure. The claim is bounced at eClaimLink or Shafafiya for a technical, format, or eligibility reason before anyone weighs whether it's payable. A denial is post-adjudication: the claim was assessed and found non-payable on the merits. Industry estimates put rejections in roughly the 12–18% band, and under PD-05-2025 the resubmission clock matters — 30 days to resubmit, against the 45-day and 141-day windows that frame the cycle. Catch a rejection late and you've burned a resubmission cycle you didn't need to.
Here's how the pieces map. Classification and validation go after the format, eligibility, and coding rejections — the pre-bill scrub. The cited Q&A serves the denial and appeal path, where a handler has to defend a position. Extraction prevents the upstream errors that cause both. What none of it does: adjudicate, decide payability, or replace the examiner. The tool gets the claim clean and to the right desk. A person still makes the call.
The 2026 regulatory perimeter: CBUAE, not the Insurance Authority
If your mental model still has the Insurance Authority as the regulator, update it. Federal Decree-Law No. 6 of 2025 — issued 8 September 2025, in the Gazette on 15 September, effective 16 September — repealed both the 2018 Central Bank law and the 2023 insurance law and pulled insurers, reinsurers, takaful operators, and insurance-related professions under CBUAE. TPAs are named directly, under Articles 78–106. The transitional deadline to reconcile your operations with the new law is 16 September 2026.
The stakes moved too. The administrative fine ceiling went from AED 200m to AED 1bn. When the downside scales like that, the conservative architecture — on-premise, or air-gapped where the data warrants it — stops being the cautious option and becomes the obvious one.
Takaful carriers are in scope on their own terms. CBUAE's External Shari'ah Audit standard landed in December 2024, with the HSA as the apex body, so a takaful operator reading this should see itself here, not a generic "insurer." One thing we don't do: give legal advice or tell you what the law means for your book. We build to your compliance team's interpretation of it. That line holds across every engagement.
How we keep a probabilistic model safe for claims
A model that's right 95% of the time is wrong 1 in 20. For a regulated line, that number on its own is useless — what matters is what happens on the other 5%. So the design fails to a human, never to a silent wrong answer.
Routing runs on a confidence threshold. Above it, the claim auto-routes. Below it, the claim holds for human review with the model's reason attached. The Q&A is extractive and citation-anchored, and it abstains at the edge of what it knows rather than filling the gap with something plausible. Underneath sit groundedness checks, a measured human-review and escalation rate, and an immutable audit trail.
Be honest about what the numbers mean, because vague claims don't survive contact with a regulator. The 95% is classification accuracy on that document mix. The 10x and 3x are throughput and cycle time, not a claim about adjudication quality. And on a new book of business, accuracy is established on the client's own documents during the pilot — before anything routes autonomously, not after.
Built for UAE health-claims rails (and life-underwriting AML)
The integration chips on this page point at real rails, so here they are by name. Dubai runs eClaimLink and the DHPO; Abu Dhabi runs DoH Shafafiya. Claims move as XML transactions carrying CPT, ICD, and DRG codes, plus the Dubai Drug Code, against standardized denial-reason formats. One distinction that trips up vendors who haven't worked these rails: NABIDH, Malaffi, and Riayati are health-information exchanges, not claims-submission channels. We don't route claims through them.
PD-05-2025 sets the cycle the system has to respect — 45, 30, and 141 days, electronic-only, twice-resubmittable. AML and CFT scope down to life and investment-linked products under Federal Decree-Law 10/2025, with goAML as the FIU channel for STR and SAR filing; we don't bolt AML logic onto health claims where it doesn't belong. And IFRS 17 connects honestly as a data-quality and auditability story — traceable contract data feeding the GMM, PAA, and VFA measurement models and the disclosures around them. We aren't performing actuarial measurement, calculating CSM, or setting reserves. We make the underlying data clean and traceable. Your actuaries do the measurement.
Integrations
Regulatory Awareness
Intelligent Document Routing
Manual document sorting created bottlenecks in case resolution, with junior staff spending hours on classification tasks.
Deliverables
- BERT-based intent classification system
- Automated routing to appropriate case handlers
- Confidence scoring with human review escalation
- Integration with existing document management system
Results
- 95% classification accuracy
- 10x reduction in manual sorting time
- 3x faster case assignment
Frequently asked questions
Does policyholder data ever leave our network?
No. The default deployment is on-premise, or air-gapped where the data calls for it, which means no policyholder data goes out to an external LLM API. That matters under PDPL — Federal Decree-Law 45/2021 — because a call to a cloud LLM is itself a cross-border transfer event, the kind that pulls in transfer mechanisms like adequacy, SCCs, BCRs, explicit consent, or contractual necessity. The on-premise design sidesteps that whole question by never making the transfer. With health-claims medical data in the mix, the case for keeping it in-network only gets stronger. We're not asserting the PDPL Executive Regulations are in force; we build to the law as it stands.
How accurate is it, really, and what happens when it's unsure?
The honest version: it routes on confidence, sends anything below the threshold to a human with the reason attached, and answers policy questions with citations or not at all. The 95% figure is classification accuracy on a legal and insurance document set — and it gets re-measured on your documents during the pilot before anything routes on its own. You see the real number on your data, not a demo number.
Does it integrate with our existing claims and policy-admin platform?
Yes. It connects to claims-management and policy-admin systems, eClaimLink and Shafafiya formats, your document-management system, and the WhatsApp Business API. The Q&A is grounded on your wordings and endorsements, not a generic schema someone else's policies trained. Nothing gets ripped out and replaced.
Does it handle Arabic claims correspondence and bilingual wordings?
Yes — classification, extraction, and Q&A all run in English and Arabic. In the UAE that's a requirement, not a nice-to-have, and it's built in from the start rather than bolted on.
How long to a working pilot, and what does a pilot look like?
A scoped POC on a sample of your real documents, AED 15,000–25,000 over about 2 weeks. The point is to establish accuracy and the human-review rate on your data — not a curated demo set — before anything routes autonomously. A production rollout runs AED 40,000–80,000 over about 6 weeks.
Is this compliant with CBUAE and PDPL?
It's designed for it: on-premise architecture, cited and traceable answers, confidence-scored routing, and a full audit trail built to be defensible to CBUAE and to your DPO under PDPL. The honest caveat we put on every engagement: we build to your compliance team's interpretation. We don't certify compliance and we don't give legal advice.
Why not just use a generic cloud AI copilot?
Because a generic copilot does the three things a regulated line can't absorb. It sends data off-network — a PDPL transfer event. It's grounded in a public corpus, not your wordings and endorsements. And it gives uncited answers a handler can't defend to an examiner. On-premise, grounded on your documents, cited, and logged is the opposite of all three, and it's what CBUAE accountability actually asks for.