By Micky Irons, founder of Mickai.
There is a quiet moment in every insurance negotiation that decides everything, and it arrives before a single premium is quoted. The underwriter stops talking about your business and starts asking about your evidence. Not your intentions, not your policies on paper, but the actual record of what happened. In 2026, with artificial intelligence now embedded in decisions that move money, deny claims, hire people and steer machines, that moment has arrived for AI risk. The question on the table is no longer whether your model is clever. It is whether you can prove what it did.
I have watched this shift accelerate over the last eighteen months, and I want to set out plainly where it leads. The uninsurable risk in AI is not the rare catastrophic failure. It is the ordinary action you cannot reconstruct after the fact. An insurer can price a known hazard. What an insurer cannot price, and will therefore exclude or refuse, is a loss you cannot account for.
The market has stopped pretending AI is just software
For a long time, technology liability cover treated machine learning as a slightly exotic flavour of ordinary software. A bug was a bug. A defect was a defect. The policy language barely changed. That truce is over. Carriers have spent the last two years writing AI-specific exclusions, carving algorithmic decision-making, generative output, autonomous agents and model drift out of the general technology errors and omissions wordings that used to absorb them quietly.
The reason is simple. Traditional software fails in ways an adjuster can trace. You read the code, you read the logs, you find the line. A modern model fails in ways that are probabilistic, emergent and, crucially, often undocumented at the moment of action. When a claims handler cannot reconstruct the decision path, they cannot apportion fault, and an insurer who cannot apportion fault is writing a blank cheque. No serious underwriter writes blank cheques twice.
So the market is doing what markets do. It is repricing the unknown. And the single largest driver of that repricing is the quality of the evidence you can produce when something goes wrong. This is where AI liability insurance audit evidence stops being a compliance footnote and becomes the centre of the conversation.
Tyche steadies the wheel of fortune above the ledger, and every spoke is a sealed record rather than a guess.
What an underwriter actually wants to see
Strip away the language of risk committees and the request is concrete. When an insurer assesses AI exposure now, they want proof that you have controls and proof those controls were operating at the moment of the loss. The second half matters far more than the first. A control you cannot evidence is, for pricing purposes, a control that did not exist.
In the submissions I have seen and the conversations I have had, the demands cluster into a recognisable shape.
- A tamper-evident log of every consequential action the model took, with the inputs, the outputs and the model version that produced them
- Proof the log was not edited after the fact, which means cryptographic integrity, not a database an administrator can quietly rewrite
- The ability to reconstruct a specific decision on demand, months or years later, without depending on a vendor still being in business
- Evidence that a human review or override path existed and was followed where required
- Assurance the record itself will survive long enough to outlast the limitation period on any claim against you
Read that list again and notice what it is really asking. Not whether your AI is safe in the abstract. Whether you can stand in front of a judge, an adjuster or a regulator and reconstruct exactly what happened, with proof that nobody touched the account. That is an evidentiary standard, not an engineering one.
The action you cannot reconstruct is the action you cannot insure
I want to state my central claim without hedging. The uninsurable AI risk is the unrecorded action. Everything an underwriter fears about your model, bias, hallucination, drift, runaway agents, becomes manageable the moment it is reconstructable, and catastrophic the moment it is not.
Consider the asymmetry. If your system makes a flawed lending decision but you hold a sealed, verifiable record of the inputs, the model state and the output, you have a defensible position. You can show the decision was made in good faith, on the data available, within the parameters you declared. The claim may still cost you, but it is a known, boundable cost. The insurer can price it.
Now remove the record. The same flawed decision, with no trustworthy account of how it was reached, is an open-ended liability. You cannot prove the input data. You cannot prove which model version ran. You cannot prove a human signed off. Every gap in the record is filled, in litigation, by the assumption least favourable to you. That is not a risk an insurer prices. That is a risk an insurer excludes.
A loss you can reconstruct is a claim you can defend. A loss you cannot reconstruct is a claim you have already lost. The record is the difference between the two.
Why ordinary logs will not save you
At this point someone always says, we already log everything. I am sure you do. The problem is not whether you log. The problem is whether your log can be trusted by a party who assumes you might be lying.
An ordinary application log is written by a system the defendant controls. It sits in a database the defendant can edit. It is exactly the evidence opposing counsel will tear apart with a single question: how do we know you did not change this after the incident. If the honest answer is, you have to take our word for it, the log is worth very little when it matters most. Evidence you can alter is evidence the other side can dismiss.
This is the distinction between recording an action and proving it. A log records. A sealed, cryptographically integrity-protected record proves. The first tells your own story to yourself. The second withstands a hostile reading by someone whose job is to find the gap. Insurers, regulators and litigators are all hostile readers by design, and the AI liability insurance audit evidence they will accept must survive that hostility.
Value lands on the tablet where it can be counted, not into the shadow where it cannot be proven.
The post-quantum clock is already running
There is a second problem almost nobody pricing AI risk has fully absorbed yet, and it changes the maths. The limitation period on a serious liability claim can run for years. A decision your model makes today might be litigated in 2032 or later. The evidence has to survive that long and stay trustworthy the whole time.
Now layer in the cryptographic horizon. A record sealed today with conventional signatures may be challengeable tomorrow if the cryptography that protected it has been undermined by advances in computing. An adversary, or simply opposing counsel with an expert witness, only needs to argue the seal could have been forged for its evidentiary value to collapse. A record that was sound when you made it but questionable when you need it is no record at all.
This is why I am convinced the durable answer has to be post-quantum from the start. If you are sealing evidence that must hold up across a decade or more, you cannot seal it with cryptography that has a known expiry. You have to use signatures designed to resist the machines that are coming, not just the machines we have. Building that in later is not an upgrade. It is a re-do of every record you ever made, and you cannot re-seal the past.
Where Mickai stands on this
I run Mickai, and our entire architecture is built around a conviction that predates the insurance market catching up to it. Mickai is a Sovereign Intelligence Operating System, the SIOS, and its founding premise is that an operator should own their intelligence, their data and, just as importantly, the record of what that intelligence did. We did not bolt the audit trail on as a compliance feature. We built the system so the record is the spine.
In the Mickai SIOS, fifty specialised brains run on the operator's own hardware, fully offline-capable, and every consequential action is sealed into a post-quantum Open Audit Record. We use FIPS 204 ML-DSA-65, the standardised lattice signature scheme, precisely because the seal has to outlast the threat. The record lives on your hardware, under your control, verifiable by anyone you choose to show it to without depending on us, on a cloud provider, or on any third party still existing in a decade. That is what sovereignty means in practice. You hold your own evidence.
I am not claiming this makes AI risk disappear. Nothing does. What it does is convert the worst class of risk, the unrecorded and unprovable action, into the manageable class, the recorded and verifiable one. It moves you from the side of the ledger an insurer excludes to the side an insurer can price.
The actuarial input nobody listed yet
Here is the part I want underwriters specifically to hear. A sealed, verifiable audit record is not a nicety that makes a risk officer feel tidy. It is an actuarial input. It changes the expected cost of a claim, because it changes the probability a claim is defensible. Two firms running identical models with identical exposure are not identical risks if one can reconstruct every decision with cryptographic proof and the other cannot. They should not be priced the same, and increasingly they will not be.
The firms that understand this early will find their AI cover available and affordable. The firms that do not will find the market quoting them exclusions where they expected protection, or declining them altogether. The discount for being provable is going to widen into a chasm.
The wheel turns, but each spoke is a fixed seal, so fortune is recorded even as it moves.
The objections, met head on
Whenever I make this argument, three objections come back, and they deserve straight answers rather than dismissal.
The first is cost. Sealing every consequential action sounds expensive. In practice the expensive thing is the denied claim, the excluded peril and the open-ended liability you carry because you cannot prove your own good conduct. A record that turns a potential catastrophic loss into a boundable one pays for itself the first time you need it, and quietly every renewal you do not.
The second is performance. People assume sealing slows the system. With the right design it does not sit in the critical path in any way that matters, and the comparison that counts is not seal versus no seal. It is seal versus the cost of being uninsurable. That is not a close call.
The third is trust in the seal itself. Why should an insurer believe your record over your ordinary log. Because the whole point of a post-quantum sealed Open Audit Record is that it does not ask for belief. It offers verification. A third party can check the integrity of the record independently and confirm it has not been altered since it was sealed. The evidence does not rest on your reputation. It rests on mathematics. That is the entire difference.
The pantheon holds the line.
What I would do this quarter if I carried AI exposure
If I sat on the board of a firm running AI in any consequential decision, regulated or not, I would treat this as a near-term operational priority and not a someday item. The insurance market is moving faster than the procurement cycle, and the gap between the two is exactly where firms get caught underinsured at the worst possible moment.
I would start by asking my own team one uncomfortable question. If our most important AI decision from six months ago were challenged in court tomorrow, could we reconstruct it, and could we prove the reconstruction had not been touched. If the answer involves any hesitation, you have found your exposure. The fix is not more logging. It is a record built to be proven, sealed against the cryptographic horizon, owned by you rather than rented from a vendor who may not be there when the claim lands.
The wider point is about where the burden of proof now sits. For a generation, software firms got to treat their internal records as a private matter, surfaced only when convenient. AI has ended that. The action your model takes is consequential, the loss it can cause is real, and the party who pays, the insurer, has every right to demand proof and is now exercising it. The age of taking it on faith is closing.
The record is the asset
I have spent years arguing that sovereignty over your intelligence is the defining commercial advantage of this decade. The insurance market is now making that argument for me, in the bluntest possible language, the language of price. Own your model, own your data, and above all own the verifiable record of what your model did, because that record is no longer paperwork. It is the asset that decides whether your claim is paid.
Tyche balances the wheel above the ledger for a reason. Fortune turns, and it always will. The only question that has ever mattered is whether, when the wheel comes round, your coins land on the stone where they can be counted, or fall into the shadow where no one can prove they were ever yours. Your insurer will ask to see the record. Make sure you have one worth showing.
Written by Micky Irons. Originally published at https://mickai.co.uk/articles/your-insurer-will-ask-to-see-the-record. More from Micky Irons and Mickai at mickai.co.uk.
