eduba Prepared by Eduba for Al Ghurair Group
For Phil Chan, SVP of AI, Al Ghurair

Deterministic AI, agentic architecture, five sectors. One methodology that travels.

A short read from Eduba written in language you already use. We saw the headline. We are not going to explain why determinism beats LLM-only architectures. We are going to talk about how to operationalize that read across the Group.

Group-level AI mandate MLOps & productionisation Methodology that travels Five operating sectors
The setup

SVP of AI at a 60-year-old conglomerate. The blank page is the work.

Construction & Manufacturing, Real Estate, Ventures, Facilities Solutions, Investments. Twenty-eight thousand people. Five sectors with very different digital baselines. The blank page in front of an SVP-of-AI in seat one quarter is not "what does AI do" — you have already shipped MLOps platforms from scratch at scale. The blank page is "where does the function start, what is the architecture pattern that compounds, and how does the team get built without burning a year on the org-design phase."

Cross-cutting the whole Group: AGCEW's AED-136M Premier Inn Dragon City contract for Nakheel, the Al Jaddaf masterplan with Pelli Clarke, the BurJuman Mall and Reef Mall retail surface, the Kabi and Zed mobility plays in Ventures, the ongoing FM repositioning, and the dividend stream from Mashreq Bank and National Cement on the Investments side. Most of the day-to-day coordination across all of it runs on documents, project managers, and weekly status calls. The question is not whether AI can help. The question is which layer each piece of the work actually lives on, in each sector.

Shared language

You call it deterministic vs agentic. We call it 60 / 30 / 10. Same idea.

Sixty percent of most enterprise work is database and code that has existed for thirty years. Thirty percent is deterministic, rule-based logic that an LLM is bad at and a small set of tested rules is great at. Ten percent is a genuine agentic problem worth a probabilistic system. Treating those three layers as one undifferentiated “AI strategy” is what wastes budget. Treating them as three separate architectures is what compounds.

This is not a frame we want to convince you of. It is the frame you already work in. The leverage is in operationalising it across five sectors that have never had a shared architectural language before.

60%
Code & database

Payroll, scheduling, vendor management, equipment tracking, ERP, FM ticketing. Postgres and tested rules outperform anything dressed up as AI. Most existing pipelines already live here.

30%
Deterministic / rule-based

RFI routing, submittal review, compliance certificate generation, BIM coordination memos, SLA-compliance reporting, asset-data normalisation. Deterministic systems with thin agentic assist. Predictable, debuggable, auditable.

10%
Genuine agentic work

Cross-sector orchestration. Risk signals across heterogeneous document streams. Translating between trade languages and operating-business vocabularies. This is where the team's bandwidth, the MLOps platform, and the LLM cost actually pay back.

Where the 10% lives in each sector

The methodology is sector-agnostic. The applications are not.

Same 60/30/10 read, applied across each operating sector. The point is that one methodology gives every cluster a shared language. Phase One proves it inside one sector. Each subsequent phase adapts the same spine to the next.

Construction & Manufacturing

Cross-sub-company orchestration

Engineering, switchgear, aluminium fabrication, façades, contracting. AGCEW wins a contract and the work ripples across three or four sub-companies. Today that coordination is in PMs' heads and weekly calls. The high-leverage AI work is surfacing risk signals from cross-project document streams and translating between trade languages: engineering, finance, regulatory.

Real Estate

Tenant operations and asset-data unification

BurJuman Mall, Reef Mall, the Al Jaddaf masterplan, the Aires Mateus residential commission. Different tenant mixes, different data systems, different reporting cadences. The high-leverage AI work is unifying asset-level data across the portfolio so the Real Estate team is making weekly calls on consistent numbers, not reconciling spreadsheets.

Ventures (Kabi, Zed, retail, education, travel)

Operator-grade decision support

Multiple operating businesses with different unit economics under one Ventures roof. The high-leverage AI work is giving operators tools that compress the loop from raw data to weekly operating decisions, without each business reinventing its own dashboard layer.

Facilities Solutions

Service request triage and predictive maintenance

Integrated FM at scale runs on tickets, vendor coordination, and SLA compliance. The high-leverage AI work is service-request triage, predictive maintenance signals from equipment telemetry, and automated SLA-compliance reporting. The current FM repositioning is a natural second sector for the methodology to enter.

Investments

Stake-monitoring and reporting compression

Significant minority stakes in Mashreq Bank and National Cement. The work here is not "AI for banks" — it is collapsing the time between earnings releases, market signals, and the family principals' read on each holding. The 10% AI layer is summarisation across heterogeneous source streams. Smaller surface area, lower priority, but the methodology applies.

What Eduba would do

A five-phase Group AI architecture program. One sector at a time.

Roughly eighteen months end to end. Methodology-first. The Group's team is trained alongside the work, not after it. Each phase produces a working artefact. Each artefact feeds the next. By Phase Five, the Group owns the methodology, the documentation, and the team.

Phase 1 · Foundation · 90 days · one sector

Pick the sector with the most cross-sub-company coordination cost today. Deliver a working orchestration pattern across the boundary that hurts most. Document the methodology spine so it ports cleanly. Train the Group's internal team on the pattern as it gets built. Sector choice is a conversation, not a default.

Phases 2–5 · Sector Adaptations · one sector per phase

Same 60/30/10 framing applied to each remaining sector. Each phase is shorter than Phase 1 because the methodology spine is already built. Each one is scoped individually against the sector's actual coordination cost. Each one produces its own working artefact. Each one feeds internal documentation that the Group owns.

Methodology Retainer · in parallel from Phase 2 onward

Architecture office hours. ICM stewardship. Internal team enablement. Design review on builds the Group's own team is shipping. This is the layer that makes the Group self-sufficient by Phase 5.

Methodology spine: Interpretable Context Methodology (ICM), published in ACM TiiS. MIT-licensed. A 52-member practitioner community is using it in production already. The reason ICM matters here is that it gives the Group something defensible internally: a published framework, not vendor lock-in.

CTO-buyer engagement

Feeld — CTO Andrew Santus

Scoped sprint with a CTO who already had a strong technical worldview and wanted operating leverage, not a strategy deck. Deliverables: workshop, advisory, Organizational Context Architecture, Strategic Operations Framework. The relevance: same buyer profile as an SVP of AI — technically credible, hands-on, low tolerance for vendor theatre.

Methodology engagement

KPMG UK, one of the Big Four

Forty-plus executives trained on the same architectural question of where AI actually fits inside a regulated, document-heavy professional-services workflow. The artefact is the spine of the engagement we would adapt for Al Ghurair Group.

Adjacent agentic build

VigilOre — multi-agent compliance platform

One hundred and sixty-plus hours of compliance work collapsed to under five minutes. The pattern is the same one that maps onto Construction Cluster project documents, FM SLA reporting, and Real Estate compliance certificates: deterministic shell around a narrow agentic kernel, productionised through an MLOps spine.

Methodology paper: Interpretable Context Methodology · submitted to ACM TiiS · MIT license

See you on the call.

This is a short pre-read for our scheduled time together. Bring one architectural decision in front of you right now where you'd want a peer to push back. We'll spend the conversation there.

Matt Creamer, CRO, Eduba · thecro@eduba.io