Group-scale AI assessment

Sequencing the data foundation before anyone bought an AI agent

A financial group wanted AI scoring across five affiliates. I proposed one affiliate, eighteen weeks, and a data readiness report before a single model — because the alternative is an agent trained on data nobody has reconciled.

One affiliate, one funded phase, one go/no-go
18 weeksOne affiliate, one funded phase, one go/no-go
Data readiness report gates everything downstream
Week 5Data readiness report gates everything downstream
What the client must send before we commit to numbers
4 asksWhat the client must send before we commit to numbers

The client

The South-East Asian arm of a large East Asian financial group with dozens of subsidiaries. Several hundred thousand policyholders, five affiliate data estates, a regulator pushing explainability and consumer protection, and an ambition to scale group-wide.

The engagement

An eighteen-week phase one on a single affiliate, structured as seven partly parallel phases with a go/no-go at the end and a group expansion roadmap as the final deliverable.

The problem

The group wanted customer scoring, retention and recommendation intelligence across five affiliates in a market where alternative data is abundant and consumer-protection regulation is tightening. The default vendor response is to propose the whole platform at once. That fails in a predictable way: identity resolution turns out to be impossible across affiliates, the historical depth is thinner than assumed, and by the time anyone discovers it the programme is already halfway funded.

What I did

I scoped phase one to a single affiliate and put a data readiness report at week five, before any model work, so the first thing the client buys is the truth about their own data. The architecture was designed for the group but validated on one entity — deliberately, and stated as such. Six assumptions were written into the proposal in plain language, each one a thing that would change the plan if false: access to raw sources and a metadata inventory, sufficient historical depth, technically feasible identity resolution, availability of digital engagement signals, an agreed personal-data regime, and the single-affiliate boundary itself. Governance and explainability were built as a layer of the platform rather than a documentation exercise, because in that regulatory environment an unexplainable model is not a model. And instead of closing with a price, the proposal closed with four things we needed from the client before quoting: regulator pre-consultation status, an anonymised data sample or catalogue, a written definition of what would count as success, and the name of a technical counterpart on the ground.

What was built

A medallion data foundation — raw, cleaned with personal-data masking and identity resolution, then business-ready customer and risk features — with an AI enablement layer of ten specialised agents over an orchestration runtime, and explainability, monitoring and audit trails carried as a first-class layer rather than a compliance afterthought.

On the table at the end

  • Forty-two-slide phase-one proposal with target architecture
  • Data readiness report as the week-five gate
  • Agent and use-case model across four business domains
  • Eighteen-week plan with seven phases and a go/no-go
  • Six named assumptions and a four-item 'send us this first' list

What it changed

Reframed a group-wide AI ambition into one funded, provable phase with an explicit exit — protecting the client from a multi-affiliate programme built on unverified data, and giving the group a scale decision based on evidence rather than enthusiasm.

How it ran

  1. 01

    Narrow the phase, keep the architecture wide

    One affiliate for phase one, an architecture designed for five — with the difference stated openly instead of blurred into a roadmap slide.

  2. 02

    Put data readiness on the critical path

    A readiness report at week five as a gate: what data exists, what quality it is, whether customers can be resolved to one identity at all.

  3. 03

    Layer the platform honestly

    Raw, cleaned with masking and identity resolution, then business-ready features — each layer with its own tooling, owner and definition of done.

  4. 04

    Make governance a layer, not an appendix

    Explainability, model monitoring, lineage and audit logging designed in from the start, because the regulator's question is the first one the board will ask.

  5. 05

    Close on questions, not on price

    Regulator status, a data sample, a written success definition and a named technical counterpart — requested before committing to a number.

Something similar on your plate?

Thirty minutes, no deck. I will tell you whether it is worth doing at all.