By Micky Irons, founder of Mickai.
An underwriter sits in front of a decision that will price a life, a building, or a fleet, and the model doing the heavy lifting was built by someone else. It was trained on data the insurer never inspected, tuned to objectives the insurer never set, and shipped through an interface that returns a score without a reason. This is the third-party model the insurer did not build, and in high-risk lines it has quietly become the thing on which everything else depends.
High-risk insurance is where the third-party model problem bites hardest. The premiums are large, the tails are fat, and the regulator is awake. When a vendor model declines a commercial property, surcharges a driver, or flags a claim as fraudulent, the insurer carries the legal and reputational weight of that call. The vendor carries almost none of it. That asymmetry is the real exposure, and it does not show up on any cession statement.
The score with no provenance
The oracle answers. It does not show its working. A score without provenance is a verdict an insurer cannot defend.
A third-party risk model returns a number. The number is confident, fast, and opaque. Ask why it landed where it did and the honest answer is usually a shrug dressed up as a feature, because the weights are the vendor's commercial secret and the training set left the building years ago. For a marketing recommendation that opacity is tolerable. For a refusal to insure, or a fraud flag that freezes a payout, it is a liability waiting for a complaint.
Three failure modes follow. First, proxy discrimination, where a postcode or a device fingerprint stands in for a protected characteristic and no one set out to make it so. Second, drift, where the world moves and the frozen vendor model keeps pricing last year's risk. Third, silent dependency, where a model the insurer cannot see becomes load-bearing across underwriting, claims, and reserving at once. None of these is exotic. All of them are findable, but only if every consequential decision leaves a record you can actually read.
Custody, not just compliance
The reflex response is a governance binder. Policies, model cards, a quarterly review committee. Useful, and entirely insufficient, because a binder describes intentions while the model makes thousands of live decisions a day. What the insurer needs is custody: the ability to run the model on infrastructure it controls, to capture every input and output, and to prove after the fact what was decided, on what evidence, and under which version. Compliance is what you say. Custody is what you can show.
Justice does not ask the model to be transparent about its soul. It asks for a record of what it actually did.
This is the tension Mickai is built to resolve. Mickai is a Sovereign Intelligence Operating System, a SIOS that runs fifty specialised brains (twenty-five domain and twenty-five operational) on the operator's own hardware, fully offline-capable. The insurer is not renting a black box across someone else's API. It is running the intelligence inside its own perimeter, where the inputs, the outputs, and the model version are all visible to the people who answer for them.
The Open Audit Record
Custody is only worth something if the record cannot be quietly rewritten after a bad outcome. In the Mickai SIOS every consequential action, a decline, a surcharge, a fraud flag, a reserve adjustment, is sealed into the Open Audit Record and signed with FIPS 204 ML-DSA-65, the published NIST post-quantum signature standard. Mickai did not invent that standard. It adopts it, which is the point: the cryptography is the regulator's, not a proprietary trick the insurer would later have to defend in front of an ombudsman.
The practical effect is a chain of accountability that survives the dispute. When a declined applicant complains, or a regulator opens a thematic review, the insurer does not reconstruct what probably happened from logs that could have been edited. It produces a signed record of exactly what the model received, what it returned, and which version was live at that moment. The vendor's opacity stops being the insurer's problem, because the insurer's own boundary is now the system of record.
Memory that cannot be revised after the fact. Every decision sealed at the moment it was made, not reconstructed when challenged.
Permanence the insurer can point to
A signed record inside the perimeter answers most questions. For the disputes that escalate, the insurer needs to prove that the record itself was not assembled yesterday to suit today's argument. That is what Pantheon provides. Pantheon is Mickai's own sovereign, Bitcoin-anchored Layer 1, with a native token, PAN, and a fixed supply of five billion. It anchors a hash commitment of the audit record to Bitcoin, so the existence and integrity of the record at a point in time can be checked against the most heavily defended timestamp in the world.
Two things this is not, because both matter to a risk officer. Pantheon does not move Bitcoin, and it is not a Bitcoin Layer 2. It commits a hash, a fingerprint of the record, and nothing else crosses the boundary. Anchoring is not spending. The insurer gets tamper-evidence rooted in Bitcoin's permanence without taking on custody of a cryptoasset, a distinction that keeps the audit story clean for the very people who would otherwise object to it.
Anchored to bedrock, not adrift. A hash commitment to Bitcoin fixes the record in time without ever moving a coin.
What this changes for the underwriter
The underwriter who started this article does not stop using third-party intelligence. High-risk lines need every edge they can get. What changes is custody of the decision. The model can still be sophisticated, even partly external, but it now runs inside a boundary the insurer controls, leaves a signed record of every call, and anchors that record where no one can quietly revise it. The vendor's cleverness stays. The vendor's opacity is no longer the insurer's silent liability.
That posture is also where the supervisory wind is blowing. Across the EU AI Act, the FCA's expectations on outsourcing and operational resilience, and the long-standing demand that pricing be explainable, the direction is the same: you may use the model you did not build, but you must be able to account for what it did. Mickai is the substrate that lets a high-risk insurer say yes to that without flinching.
The evidence under the claim is concrete, not rhetorical. Mickai rests on 101 filed UK patent applications, around 2,234 claims, owned by Mickai LTD, with named inventor Micky Irons. The portfolio is the proof of work, not the headline. The headline is simpler. In high-risk insurance, the model you did not build is only safe to use when the record of what it did is one you can hold, sign, and anchor yourself.
Written by Micky Irons. Originally published at https://mickai.co.uk/articles/high-risk-insurance-ai-and-the-third-party-model-the-insurer-did-not-build. More from Micky Irons and Mickai at mickai.co.uk.
