UAE · MOEI
One AI platform.
Three challenges.
A multi-agent ecosystem on a single Agent Builder — agents that research, debate in councils, and hand off to humans, working live across every channel.
Agent BuilderMaster Orchestrator
Council of AIHuman-in-the-loopSelf-learning
Housing · partnerships · citizen engagement — built at 42 Abu Dhabi · live at moei.bimco.ae
The problem
Three real ministry pains, one platform
Today
Housing-loan arrears
Manual case-by-case rescheduling · officers buried · slow, inconsistent decisions.
→ Tawasul: minutes, with human sign-off
Challenge 1
Today
Country briefings
Analyst-heavy, days to prepare · energy & infrastructure delegations need fast, grounded intel.
→ Tawasul: a grounded briefing in under a minute
Challenge 2
Today
Citizen channels
Web, WhatsApp, phone, officer desk — disconnected, no unified memory.
→ Tawasul: one brain across every channel
Challenge 3
Three real MOEI workloads — slow, manual, fragmented. One Agent-Builder platform fixes all three.
01
Challenge 01
Housing-Loan Arrears Rescheduling
Live trace
Case Intake✓
Policy✓
Risk & Hardship✓
3 agents in paralleldone
Parallel intake · reads case + attachments
Reschedule calc
InstallmentAED 2,180/mo
20% deduction ceiling
within original term
20% rule ✓
term ✓
Compliant plan · Programme rules
Council debate
Compliance 20% ceiling holds — plan is compliant.
Advocate Approve — affordable for a family of 4.
Risk re: add one condition to de-risk.
Chair verdict Approve with conditions · 86%
Advisors debate → chair reaches a verdict
Decision
● Auto-approved
Clean & within policy — no officer needed.
● Officer review
ApproveOverride
● Refer to a human
Unstable income — officer decides.
Tiered human sign-off
Builder agents run the whole case — parallel intake → compliant plan → council debate → tiered human sign-off. Cached, with openable attachments.
02
Challenge 02
Country Intelligence Briefings
Live research · Brazil
World Bank · GDP$2.17T
Renewables48%
CEPA statusin talks
CapitalBrasília
Live web + World Bank + connectors
Strategic Council
Economic Strong CEPA upside — pursue.
Energy Hydrogen + ports fit our strategy.
Devil's Advocate re: dependency risk — hedge it.
Chair verdict Pursue with conditions · 78%
Advisors debate + Devil's Advocate → verdict
Executive briefing
82%confidence
Talking point · hydrogen MoU
Talking point · ports & logistics
Decision-ready briefing · grounded to today
AI slide deck
Brazil
Prepared for UAE leadership
Snapshot
Opportunities
Masdar · hydrogen · ADNOC
1 / 9
Generated HTML deck · presentation-ready
Four agents gather live data → a Strategic Council debates → one decision-ready briefing, grounded to today, with an AI-generated slide deck.
03
Challenge 03
Omnichannel AI Engagement
Master Orchestrator
Web
WhatsApp
Phone
in →
Web “What documents do I need?”
→ Housing Eligibility
WhatsApp “Reschedule my arrears”
→ Arrears Rescheduling
Phone “I want to file a complaint”
→ Hand off to officer
Routes every channel · or hands off
Tawasul · Citizen
How do I apply for housing support?
You'll need your Family Book & Emirates ID.
And my arrears?
● handled by Housing agent
Citizen Assistant · every channel + memory
Officer Co-Pilot
Ahmed Al Mansoori high risk
Next best action
Confirm income docs, offer a reschedule.
Confirm income docs, offer a reschedule.
“Thank you, Ahmed — you qualify for a reschedule…”
⚠ escalate to a human officer
Decision-ready brief + reply + handoff
Command Center
1,240cases
82%CSAT
54%auto
Auto-resolved vs handed off
“Build me a risk view” → HTML · illustrative figures
AI-generated leadership dashboards
One master orchestrator powers all three surfaces — routing, the officer Co-Pilot, and AI-generated leadership dashboards — learning from every resolved case.
The Agent Builder
Compose anything — no code
KnowledgeVector RAG
Ingest any source
PDFWordExcelCSV
TXTMDImageVideo URL
Web linkOpen Data
Attached to the agent
✓Housing policyMD
✓Tariff scheduleXLSX
✓Bayanat datasetOPEN DATA
Docs, media, web & datasets — chunked & embedded into pgvector for semantic search (RAG)
Agent31 built
Persona
“You are the Housing Eligibility Agent — assess eligibility, cite the rule…”
Model
Google Vertex AI
Tools granted
policy_searchapplicant_lookupweb_search
Bound service
Housing Assistance
Persona + model + knowledge + tools, bound to a service
Connector7 wired
Kind
HTTP APIMCPOpen DataBayanat
Endpoint
api.worldbank.org/v2/country/{ISO3}/indicator…
World BankHTTP
DeepWikiMCP
MOEI Open DataCKAN
REST APIs, MCP servers & open-data portals — granted to agents as tools
Workflow6 built
Ordered stages
Intakeparallel ×3
Decisionsingle
Councildebate ×4
Synthesissingle
Chain stages that run agents in parallel, alone, or as a debating council
Author knowledge, agents, connectors and workflows in the Builder — parallel, single or council stages. Every surface and the orchestrator use them live.
Under the hood
One architecture, end to end
Inputs
Citizen request
web · WhatsApp · phoneOfficer desk
case queueLeadership ask
“build me a view”Live data
connectors · open dataDocuments
files → knowledgeAgent Builder
knowledge · agents · connectors · workflowsMaster
Orchestrator
Orchestrator
Execution
parallel · council · human-in-loopCitizen answer
grounded replyOfficer brief
+ suggested replyDecision / verdict
council-backedAI dashboard
& slide decksHuman handoff
escalationOutputs
↻ self-learning · every outcome → memory
ReactFastAPITurborepo
Google Vertex AIpgvectorRedis
Channels in · the master orchestrator draws on the Builder & data · agents execute · surfaces out — and every outcome feeds back.
Trust & sovereignty
Built for government — sovereign by design
Deployment
Sovereign cloud or on-prem
Runs on UAE sovereign cloud or fully on-premise · no citizen data leaves the perimeter · swappable model adapter.
UAE data stays in the UAE
Data sovereignty
Human-in-the-loop
Auto for clean cases
Officer review for borderline
Complaints & sensitive → human
No consequential decision without a human
Audit & explainability
Every decision logged & attributable
Reversible & appealable
Answers cite the policy clause
Reviewable AI decisions
Validated learning
Only human-validated outcomes feed the self-learning loop — never learning from bad or biased outcomes.
Human gate on what is learned
Safe self-learning
PII encrypted at rest · UAE Pass / officer SSO + role-based access for the pilot · the demo is password-gated to protect AI credits.
Impact
From days to minutes
0% less manual review*
0sec per country briefing*
0% queries auto-resolved*
0live channels, one brain
∞self-learning loop
Arrears decision: days → minutes
Country briefing: analyst-days → < 1 min
Officer load → focused on the hard cases
Adopt
1-day training
officers live after a single day
No-code
Any employee builds agents
no coding — point & configure
Scales
Every MOEI service
one platform extends to any service
* Illustrative / projected from the pilot — not final KPIs.
See it live
Open the Tawasul platform
Open the live platform →
Scan or click · moei.bimco.ae