AI Business Transformation Expert
I lead the AI transformation of commercial underwriting, frontier capability turned into systems underwriters trust, boards back, and regulators can inspect. Strategy through to production, and accountable for what ships.
A glass box, not a black one. Every layer inspectable, every decision replayable.
Most AI in insurance dies in the gap between the people who can build it and the people who carry the risk of using it. I sit in that gap on purpose.
On paper I'm the business-side deployer: I own the use cases, the obligations and the outcomes, and I'm the single channel into the engineering programme that ships them. Strategy on one side, production systems on the other, one person accountable for the translation between them.
In practice that means writing the harness before reaching for the model, designing the sign-off before the automation, and putting nothing in front of an underwriter that can't explain how it got there. That discipline is what carries a system past review and into daily use and keeps it there long after the pilot.
The unit of work isn't a use case. It's the whole path a submission takes and at every step, an explicit decision about what the machine prepares and what the human owns.
Broker email, schedules, questionnaires and attachments turned into structured, checkable fields with a privacy boundary that keeps sensitive material inside.
Appetite fit, risk banding and routing to the right desk, so the underwriter's first look is at something already worth looking at.
The account assembled from external and internal sources: operations, footprint, loss history, peers. Sourced, so every line can be traced back.
Exposure characterised against the questions underwriters actually ask, including the adversarial view of how a claim would be argued.
A layered architecture: the intelligence advises, a deterministic rules layer computes, and nothing overwrites a confirmed input silently.
Clause selection and deviation checking against standards, so what's granted is what was intended and any drift is visible before it's signed.
The quote leaves with its reasoning attached: what was assumed, what was checked, who decided. The audit trail is a by-product, not a project.
Stage maturity is stated honestly live, in build, designed. The interesting engineering is rarely the model; it's the handover between two stages, and the moment a human takes the pen.
Turn an underwriting workflow into something a machine can help run without flattening the judgement that made it worth running. The hard part is deciding what stays human.
Working software, not slideware. Agentic pipelines, retrieval, workbenches and dashboards that go into production, get used on a Tuesday morning, and survive the second week.
Deployer obligations, decision rationale, shadow deployment, graduated kill switches. The unglamorous layer that quietly decides whether any of the rest is allowed to exist.
Across all three Board & executive communication Training & enablement Operating-model design English / German
A web application pairing account benchmarking with deep, sourced research, in daily use across a 100+ underwriter base. Built to be boring in exactly the right places: every claim traceable to where it came from, nothing asserted that can't be shown, no confident nonsense dressed up as analysis.
Separates what a model may suggest from what the business will actually stand behind. A confirmed-input gate before anything computes, a sourcing waterfall that always prefers the most authoritative field available, and a rules layer no model is permitted to overwrite. The intelligence advises; the arithmetic stays answerable.
Simulates the shape of European collective redress against an insured, grounded in the Representative Actions Directive and the revised Product Liability Directive. Agents take genuinely opposing roles and are scored on the strength of the case they can build, the disagreement between them is the signal, not a defect to be smoothed away.
A set of connected primitives rather than a product: glass-box decision rationale, a decision twin that shadows the human long before it replaces a step, staged shadow deployment, a graduated kill switch instead of one red button, and a compliance view that reads obligations as live telemetry rather than an annual PDF.
Walks rapidly-built applications into enterprise compliance: identity, deterministic security scanning that doesn't depend on a model's opinion, and one hard architectural rule, a model never audits its own output.
Rationale written after the fact is memory, and memory is not evidence. Capturing it inside the flow of work turns the audit trail from an obligation into a by-product and makes the reasoning of a good underwriter reusable by everyone else.
Thirty years of pattern recognition normally leaves the building with the person carrying it. This turns that tacit judgement into something searchable, teachable and importantly reviewable, so it can be corrected rather than merely inherited.
A documentation standard for commercial liability data that removes the ambiguity models otherwise guess their way through. Unglamorous, upstream, and worth more than most prompt engineering. Written in English and German.
A layered separation so that external processing never sees what it has no business seeing. Designed alongside a structured evaluation practice an edge-case library and a scoring framework because "it demoed well" is not a procurement decision.
Taking the whole path apart deliberately: where the handovers leak, where a confident output would go unchallenged, where an obligation has no owner. Delivered as a scored review with a remediation path rather than a list of misgivings, the point of a red team is that someone can act on it.
Described by capability, not by system name. Internal designations, architecture, data and commercial detail stay where they belong, happy to go deeper in a conversation under the right cover.
Quality lives in the harness, retrieval, tools, guardrails, evaluation, escalation paths. Swap the model and a good harness barely notices. Swap the harness and nothing survives.
Deployer duties under the EU AI Act aren't a compliance chore bolted on at the end. Read at design time they're a surprisingly good requirements document and a very effective argument for doing it properly.
A model asked to interpret a badly named field will interpret it, confidently. Fix the vocabulary before blaming the output.
End to end doesn't mean end of the underwriter. It means the assembly is done by the time they arrive, and their attention lands on the part that actually needs a person.
Repeatable, teachable frameworks over folklore passed between colleagues. If a prompt can't be reviewed like code, it isn't in production, it's in circulation.
Every decision leaves a record. That's the whole design.
If a system can't say what it used, what it assumed and who signed, it isn't finished, regardless of how well it performs.
The academic grounding and the habit of asking what a number is actually evidence of.
Data and process work close to the factory floor, where a broken process announces itself immediately.
Machine learning research, learning where models genuinely earn their keep, and where they merely look impressive.
Learning the business from its data first, where the numbers come from, what they mean, and where they quietly disagree with each other.
Long Tail Underwriting, Liability, Cyber and Motor. Leading the transformation of the process itself, and owning what the systems do once they're live.
Four disciplines, one throughline: make the reasoning visible.
The route through manufacturing data, ML research and insurance analytics wasn't planned as preparation for underwriting AI. It turns out to have been exactly that. Underwriting is judgement under uncertainty, documented, which is the one problem every stop along the way had in common.
Automate the preparation.
Be present for the people.
I'm open to speaking, panels and advisory work on AI in insurance, end-to-end process design, harness engineering, deployer governance, and what it actually takes to get an AI system past a board and into daily use.