Four weeks embedded with a ~€300M specialty carrier: every leader interviewed, the platform reviewed end to end, a working fraud agent on the table.
4wksKickoff to costed roadmap
9Leadership working sessions
1.5daysTo a working fraud prototype
5/5Leaders aligned · CEO to data science
Dashed figures are placeholders pending confirmed engagement data.
01 · The problem
Every AI project paid the integration tax
The client writes roughly €300M in gross written premium across a deliberately fragmented portfolio: aviation, marine, and dozens of highly customised, often one-of-a-kind products. Low frequency, high severity claims. Small data volumes per line. The hardest possible statistical terrain for AI, and a business already investing seriously in it.
Like most organisations moving fast on AI, each new use case was being wired directly into source systems rather than into a governed data layer. Every project paid a full integration tax: slow to build, impossible to reuse, and disconnected from the data governance discipline the business had invested in. There was no standard way for agents, regardless of which model powered them, to discover and consume governed data products.
Leadership had a clear three-layer ambition: fix the foundations, build a governance and context layer, then scale business process automation on top. What they needed was an outside view on whether their architecture would survive that journey without a redesign.
Before the audit
Bespoke integrations per use case1
Reuse across AI projects0
Weeks to wire each new agent6+
Shared agent–data interfaceNone
02 · The audit
Four weeks, session by session
Week 1 · Listen
CEOKickoff: ambition and sequencingAnchored the audit to the journey the board already backs: foundations → governance → automation at scale.Roadmap
CIOEstate walkthrough: will it survive without a redesign?Surfaced the hard questions on catalog readiness, metadata richness, lineage, and API exposure.DatabricksLegacy PAS
DEPlatform deep dive: products, contracts, lineageHead of Data Engineering · mapped the onboarding path for governance metadata into Unity Catalog.DatabricksUnity Catalog
DIUse-case portfolio: where the integration tax bitesHead of Digital Innovation · claims fraud detection selected as the flagship use case.ClaimsPAS
Week 2 · Prove
DSModels & defensibilityHead of Data Science · ensembles plus generation-then-governance for decisions that hold up in court.LLMGBM
DHFraud prototype: half-day finish + QAA demoable agent in 1.5 days: every flag with its evidence written at decision time.Fraud agent
ALLPrototype demo: full leadership teamThe architecture debate turned concrete: a live agent the room could argue with.Fraud agent
Week 3 · Design
DHTarget architecture: Context & Control PlaneExtending the client’s existing governance discipline to models and agents rather than importing a foreign framework.Unity CatalogMCP
CIOWritten technical Q&ACatalog readiness, metadata, lineage, API exposure. Answered in writing, folded into the roadmap.Unity Catalog
Week 4 · Cost & hand off
DHPhasing and costingThree phases, each budgeted: a plan the client’s own team can execute against.Roadmap
Databricks lakehouseCore data platform: a solid foundation, ready to carry the context plane.Governed
Governance metadataData products, contracts, and lineage: mature discipline, currently outside the catalog.Governed
Policy admin systemsOne per specialty line, with every AI project wired in directly today. The integration tax starts here.Action needed
Claims systemFraud checks are manual today; the flagship use case moves them onto the governed layer.Action needed
Unity CatalogCatalog of record for the target state; the metadata onboarding path was designed in week 3.Target
MCP semantic layerOne governed interface for every agent, on any model. Phase two of the roadmap.Target
04 · The plan
A Context and Control Plane for AI
Not a foreign framework: an extension of the client’s existing governance discipline to models and agents, in three moves.
01Govern
Onboard metadata into Unity Catalog
Existing data products, lineage, and contracts brought into an open, standards-based catalog, and keeping the governance discipline the business already invested in.
ResultOne governed catalog of record
02Expose
A governed semantic layer via MCP
Any agent, on any model, discovers and consumes governed data products through one interface. The per-project integration tax ends here.
ResultOne interface for every agent
03Deliver
A flagship use case, done properly
Fraud detection taken from prototype to production on top of the governed layer, with generation-then-governance controls for auditability.
ResultProduction proof of the plane
05 · The prototype
A prototype that walks the claim graph
Built in a day and a half and demoed to the leadership team: the agent hops the claim graph entity by entity, ensembles LLM reasoning with traditional models, and writes its rationale on every decision.
A replay of the prototype demo. Nothing here runs in production; taking it there is step three of the plan.
06 · What four weeks bought
Alignment first, architecture second
5/5Leaders behind one plan
CEO, CIO, and the heads of data engineering, digital innovation, and data science signed off a single target architecture.
6Systems reviewed end to end
From lakehouse to legacy policy systems: every surface the agents will touch, assessed in writing.
1ifaceInterface for every agent
The MCP layer replaces per-project integrations with one governed way in, on any model.
3Phases, each costed
Govern, expose, deliver: a plan the client’s own team can execute, budget line by budget line.
0Redesigns needed
The target architecture extends the client’s existing governance investment. Nothing gets thrown away.
36hrsIdea to working prototype
The fraud agent turned an abstract architecture debate into a demo the leadership team could argue with.
Dashed figures are placeholders pending confirmed engagement data.
Why it matters
Invest once in the context layer. Every agent after that plugs in.
Most insurers are being sold AI use cases one integration at a time. This engagement shows the alternative: invest once in a governed context layer, and every subsequent agent (underwriting, claims, fraud, or pricing) plugs into the same governed foundation. Faster to ship, reusable by design, and auditable end to end. That is the architecture Data Hatch designs and builds for regulated industries.
