Underwriting the Unprovable: Why the AI Insurance Market Needs a Signed Record

The question an underwriter actually asks

Sit across from a serious underwriter and you learn quickly that they are not interested in your demonstration. They have seen the demonstration. What they want to know is narrower and colder. When this system causes a loss, and at scale it will, what can you put in front of me to show what happened, when it happened, who or what decided, and whether anyone touched the record afterward. That is the whole conversation. Insurance is not a bet on your good intentions. It is a price on the evidence you can produce after the fact.

For most artificial intelligence (AI) systems shipping today, the honest answer to that question is a shrug. There are application logs that the operator can rewrite. There are model outputs nobody can reconstruct because the weights have since been updated. There is a vendor assurance, which is to say a marketing claim with a signature block. An underwriter cannot build a premium on a shrug. So they do one of two things. They decline the risk, or they price it as if the worst is both certain and undiscoverable. Neither outcome is good for the company seeking cover, and the second is quietly strangling deployment in exactly the high-value sectors that AI was supposed to transform. The pattern repeats across healthcare, lending, logistics, and the public sector, wherever a single bad automated decision can become a class of claims.

Insurance is a machine for pricing proof, not risk

We talk loosely about insurers pricing risk. What they actually price is the provability of risk. Those are different things, and the difference is the entire subject of this essay. A peril you can measure, attribute, and reconstruct after the event is an insurable peril, even when it is frequent and severe. Motor insurance covers millions of collisions a year precisely because each collision leaves a legible trail: a police report, a damage assessment, telematics, witnesses, a timestamped sequence. The actuary has data, the loss adjuster has a method, and the courts have a settled approach to fault. Frequency and severity are knowable, so the line is writable.

Now strip the trail away. Imagine collisions that left no skid marks, no witnesses, no agreed account of who had right of way, and dashboard cameras whose footage the at-fault driver could silently re-edit before the adjuster arrived. No serious carrier would write that book at any sane premium. The peril has not changed. The provability has collapsed, and with it the insurability. This is the precise condition of most AI deployment today. The harm is real and increasingly frequent. The trail is editable, partial, or absent. The actuary has nothing to stand on, and an actuary with nothing to stand on does not invent a number, they walk away.

Why AI breaks the loss-adjustment model in particular

Traditional software fails in ways that are tedious but tractable. A bug is deterministic. Give the same input, get the same crash, find it in the stack trace, attribute it. AI systems violate every one of those comforts at once. Outputs are probabilistic, so the same prompt can produce a harmful result on Tuesday and a benign one on Wednesday. Models are mutable, retrained and fine-tuned on a cadence that means the system that caused the loss may no longer exist by the time anyone investigates. Behaviour is emergent, arising from the interaction of model, prompt, retrieved context, tools, and other agents, so there is rarely a single line of code to point at.

For a loss adjuster, this is a nightmare with a name: non-reconstructability. You cannot adjust a loss you cannot reconstruct. And there is a second, sharper problem that keeps risk officers awake. Most AI logging is written after the fact and held by the operator, which means the party with the largest incentive to alter the record also has the technical ability to do so. The adjuster is asked to trust evidence produced and held by the claimant. In every other line of insurance that arrangement would be considered a moral hazard so severe as to be uninsurable. In AI it is simply the default architecture, accepted because nobody has stopped to notice how strange it is.

The regulatory clock makes this urgent, not theoretical

None of this is a distant worry. The European Union (EU) Artificial Intelligence Act brings its high-risk obligations into force from August 2026, and those obligations are, read plainly, a demand for provability. High-risk systems must keep records, enable traceability, support human oversight, and produce logs adequate for post-market monitoring and incident investigation. A regulator asking to see the logs after an incident is, functionally, a loss adjuster with subpoena power. Liability regimes are moving in the same direction, shifting the burden so that an operator who cannot evidence what their system did may be presumed to have caused the harm. Absence of proof is curdling from an inconvenience into a finding of fault.

Underwriters read regulation as a leading indicator of claims. When the law says you must be able to reconstruct and disclose, the law is also telling carriers where the litigation will land. A company that cannot satisfy the regulator's evidentiary demand is, to an insurer, a company whose defence will fail and whose claim will pay out in full. That company is either uninsurable or surcharged into uncompetitiveness. The compliance deadline and the underwriting question have become the same question, arriving on the same calendar. The firms treating the EU AI Act as a documentation chore are mispricing their own future premiums, and the carriers know it even where the firms do not.

What insurers will demand, stated as an underwriting checklist

It is worth being concrete about what closes this gap, because the requirements are not exotic. They are the conditions any actuary would impose if they thought the problem through from first principles. An AI record becomes underwritable when it has four properties working together, and the failure of any one of them tends to undo the other three.

• Completeness: every consequential action is recorded, not a sampled or summarised subset, so the adjuster sees the whole sequence rather than a curated highlight reel.
• Integrity: the record cannot be altered after the fact without detection, removing the moral hazard of operator-held evidence.
• Independent verifiability: a third party, whether regulator, adjuster, court, or counterparty, can confirm the record's authenticity without trusting the vendor's word or the vendor's servers.
• Durability: the record survives model retraining, vendor turnover, and the passage of time, because claims arrive years after the event that caused them.

Read that list again and notice what it is not. It is not more dashboards. It is not a better incident-response runbook. It is not a service-organisation control report that attests a vendor followed a process. Those things describe controls. The underwriter is asking for evidence, and the distinguishing feature of evidence is that it holds up when the party who produced it has every reason to wish it said something else. A control tells you a process existed. Evidence tells you what that process actually did on the day it mattered, and only the second thing settles a claim.

The actuarial precondition, named plainly

Put the pieces together and a clean principle falls out. Auditability is the actuarial precondition for the AI economy. Not a feature, not a compliance line item, but the thing that has to be true before a premium can be calculated at all. Where a tamper-evident, independently verifiable record exists, the actuary can attribute losses, build a credible frequency and severity model, set deductibles and limits that mean something, and write a line that prices the real risk rather than the fog around it. Where it does not exist, the only rational responses are refusal or a punitive load, and the punitive load tends toward the cost of assuming maximum undisclosed liability.

This reframes auditability from a grudging cost into a financial asset with a measurable return. The spread between the uninsurable-fog premium and the provable-record premium is the value of the audit trail, expressed in pounds per year. For a company deploying AI in regulated, high-consequence settings, that spread is not small, and it compounds, because a clean evidentiary history also lowers the load on every renewal. The firm that can prove its conduct does not merely get cover. It gets cheaper capital, because insurance is a form of capital, and capital is always cheaper for the borrower who can show their books. The same record that satisfies a regulator therefore shows up on the balance sheet as a lower cost of risk transfer, year after year.

Why most attempts at this quietly fail

The obvious objection is that everyone logs already, so the problem must be solved. It is not, and the reasons are instructive. Most logs are written after the action they describe, which means a system can act and then decide what to admit, the digital equivalent of an unwitnessed confession. Most logs live in storage the operator controls, so integrity rests on trusting the very party with the motive to edit. Most verification, where it exists at all, requires calling the vendor's interface and trusting the vendor's answer, which is no verification at all in any sense an adjuster would accept. And most retention strategies do not survive the multi-year gap between an AI decision and the claim it eventually generates.

There is also a quieter failure mode worth naming. A record that can only be checked using the vendor's own tooling, on the vendor's own servers, is not independent. If the company disappears, is acquired, or simply declines to cooperate during a dispute, the evidence evaporates with it. An underwriter pricing a ten-year tail cannot accept evidence with a dependency on the continued goodwill and solvency of the insured's software supplier. The verification has to work in a plain browser, offline, with no live trust in anyone. Anything less is a story, and we have already established that underwriters do not buy stories. The market has not yet punished this gap loudly, but it will, in the form of exclusions that quietly carve AI conduct out of policies that buyers assumed it was in.

What a provable record looks like when it is built right

I run Mickai, a Sovereign Intelligence Operating System that is built and live, and I will describe how we approached this, not as a pitch but because it is the clearest illustration I have of the four properties made real. The mechanism is the Open Audit Record. Every AI action across the system is signed before it executes, not logged afterward. That ordering matters more than any other single design choice, because a signature that precedes the action removes the window in which a system could act first and account for itself later. The commitment is made when there is still nothing to hide, and the record is what happened, not what was reconstructed about what happened.

Those signed entries are hash-chained and append-only, so the sequence is tamper-evident: alter any link and every subsequent link breaks visibly. The signatures use a post-quantum scheme, the United States National Institute of Standards and Technology (NIST) standard Federal Information Processing Standard (FIPS) 204, also known as Module-Lattice Digital Signature Algorithm at security category three (ML-DSA-65), so a record written today still verifies decades from now, after the cryptography that protects ordinary systems has aged out, which is exactly the multi-year horizon an insurance tail demands. The whole thing verifies offline, in an ordinary browser, with no trust placed in Mickai as the vendor. A regulator, an adjuster, or a court can confirm the record themselves. The fifty brains that do the work, twenty-five domain and twenty-five operational, run on our own Poseidon silicon substrate, and the audit root is anchored externally through Pantheon, a sovereign Layer 1 settlement chain that anchors to Bitcoin, so the proof does not depend on us still being here to vouch for it.

The honest caveats

A signed, verifiable record is necessary, but it is not a cure for every risk, and I would distrust anyone who told you otherwise. It proves what the system did. It does not, on its own, prove that what the system did was wise, fair, or lawful. Those are separate judgments that auditability makes possible rather than makes automatic. A perfect record of a harmful decision still documents a harm. What it changes is that the harm becomes attributable, the conduct becomes assessable, and the dispute becomes adjudicable on evidence rather than on competing assertions. That is precisely the transformation an insurance market needs, and precisely no more than that.

There are also real costs, and I would rather state them than pretend the property is free. Signing before execution adds latency and engineering discipline. Post-quantum signatures are larger than their classical predecessors, and the storage adds up across billions of actions. Building verification that works offline and independently of the vendor is harder than shipping a dashboard. These are genuine trade-offs. But set them against the alternative, which is being uninsurable or surcharged as if you were the worst case in your sector, and the economics resolve quickly. The cost of provability is small next to the price of the fog, and it is the kind of cost that falls over time as the engineering matures, while the price of the fog only rises as regulators and courts tighten the burden of proof.

Where this lands

The AI insurance market is going to sort companies into two populations, and the sorting has already started in the quiet conversations underwriters have with their reinsurers. In the first population are firms whose AI conduct is a black box: plausible in a demonstration, indefensible in a dispute, and therefore either uninsurable or priced at the cost of assumed maximum liability. In the second are firms that can hand an adjuster a complete, tamper-evident, independently verifiable record and say, here is exactly what happened, check it yourself, you do not have to trust us. The second population will deploy AI in the places that matter, because they can be covered there. The first will be confined to the places where nothing much is at stake.

My thesis is simple and I will state it without ornament. In the AI economy, the signed, offline-verifiable record is not paperwork that follows the deployment. It is the precondition that decides whether the deployment can be insured, and therefore whether it can happen at all in any serious sector. Insurers cannot price what cannot be proven. Auditability is the actuarial precondition, and the companies that understood this early will find that the thing they built to satisfy a regulator turns out to be the thing that made them cheap to insure, credible in court, and trusted by the counterparties who would otherwise have walked. The record you cannot edit is, it turns out, the most valuable thing you can own. We hold one hundred and one filed United Kingdom patent applications, about two thousand two hundred and thirty-four claims, owned by Mickai LTD, because we believe that point hard enough to have written it down where it cannot be quietly revised.