The Signature Is the Product: Insurers Will Write the Rules of AI

Strip a signature down to first principles and it is a promise with something behind it. Your name means something only because of what can be taken from you if you are wrong. A license that can be revoked. A reputation that can be ruined. A balance sheet that can be sued.

Now follow that one step further. For almost every professional who signs anything that matters, the last thing standing behind the signature is an insurance policy. The doctor has malpractice cover. The lawyer and the accountant carry errors and omissions. The contractor has general liability and a bond. The license says you are allowed to do the work. The insurance says someone will pay if you do it wrong. Clients have always been buying both, even if they only noticed the first.

When I say the signature is the product, I mean something precise: the product is insurable accountability. Nobody can credibly sign for more than they can pay for.

Manufactured intelligence does not change this. It makes it the whole game. When a machine can do the analysis for the cost of the electricity, the analysis is no longer what the client is paying for. The client is paying for someone to be on the hook.

Who pays? Not the model maker. Its terms of service disclaim exactly this. The claim goes where claims always go: to the party with the relationship, the one who could have caught it, the one whose name the guest actually knows. The restaurant. Which means the restaurant’s insurer. Which means that at renewal, the restaurant is going to be asked some new questions.

That is the general rule, and I expect it to hold. Liability rolls downhill to the last name on the page: the deployer, the professional, the business that put the output in front of a customer. Model makers will take liability back only where doing so sells, as the largest labs already did in 2023 when they began indemnifying enterprise customers against copyright claims. Everything else stays with the signer.

Which is why the signer is about to become the most scrutinized party in the entire AI economy, and the scrutiny will arrive as an insurance form.

Insurance has met new risks before, and it always follows the same three steps. We are watching the first step happen with AI right now.

Cyber insurance ran this exact play in the last decade. First, policies were “silent” on cyber and insurers got surprised by claims they never priced. Then they excluded it and sold it back as its own product. Then came the conditions. Insurers began requiring multi-factor authentication, backups and incident plans as a condition of coverage, and a large share of the business world adopted those practices not because any law said to but because they could not get insured otherwise.

The same phrase is already in circulation for AI. The industry calls it “silent AI”: existing policies that never mentioned AI and may or may not respond when it causes a loss. Silent risk is the thing insurers hate most. They are ending it now.

Here is the part people outside insurance tend to miss. An insurer cannot price a risk by thinking hard about it. It prices from history. Every premium is a prediction, and the only thing that can train that prediction is losses that actually happened.

This is the Reality Equation operating at institutional scale. An insurer is an Actualizer like any other: it has its own Reality, the ratio of the claims that arrive to the claims it expected. And like any other Actualizer, it cannot reach in and set its own Expectation. No actuary can reason a correct AI premium into existence. No regulator can decree one. The only road runs through the record: claims happen, claims become history, and history slowly retrains the prediction.

Which explains the exclusions perfectly. With no record, the honest move is to refuse the bet. It also tells you what comes next. The first wave of priced AI coverage will be wrong, in both directions. Some lines will turn out to be far safer than feared, and those premiums will fall. Others will produce a shock, most likely the first time one model error hits thousands of policyholders in the same week, and that market will slam shut for a year or two before it reopens at a new price.

The practical translation: the AI insurance market will not be smooth. It will lurch, the way every market does when it is learning from its own losses in real time.

1. The renewal questionnaire becomes the AI rulebook

   Before any statute tells a small business how to use AI, its insurance renewal will. Do you use AI in anything a customer sees? Who reviews it before it goes out? Is there a record? Your answers set your premium, and in some cases whether you get a policy at all. Most businesses will change how they use AI to give better answers on that form, years before a law requires it.

   Watch for: AI questions appearing on routine small-business and professional liability renewals by 2028.

2. The log becomes the defense

   Insurable AI is AI with a record. What was asked, what came back, who looked at it, what they changed. In a claim, the business that can produce that record is defensible and the one that cannot is not. “If it isn’t logged, it wasn’t reviewed” becomes the working rule, and quietly keeping an audit trail of AI work becomes as ordinary as keeping receipts.

   Watch for: policies that require, or discount for, logged human review of AI output.

3. The certificate becomes the credential

   In my last piece I predicted a “human-accountable” label, the way “organic” works for food. I now think the label will be an insurance certificate. Contractors already hand over a certificate of insurance before they start a job. Professional firms will do the same for AI work: proof that the output is covered, by whom, and up to what amount. A low premium will quietly signal a good process, and clients will learn to ask.

   Watch for: “AI work covered” or its equivalent appearing in proposals, contracts and vendor requirements.

4. Model makers start acting like insurers

   The copyright indemnities of 2023 were the first move. The next is warranties on performance: a lab guaranteeing, for a specific task in a specific industry, that its output meets a standard, and paying if it does not. To do that at scale, the labs will need reinsurers behind them. The lab that is willing to stand behind its own output will win the regulated industries, because it takes weight off the signer.

   Watch for: task-specific performance guarantees from a major AI company, backed by a named insurer.

5. The flip: not using AI becomes the negligence

   This is the prediction I hold most strongly, and it runs against every instinct people have today. Right now insurers treat AI as a risk they add to your file. Once the loss record shows that AI-checked work produces fewer errors than unchecked human work, and in some fields I expect it will, the math reverses. First insurers discount for using the machine as a second reader. Then they surcharge for skipping it. Then the standard of care moves, and the professional who did not run the check is the one who cannot get covered. Nobody will pass a law saying you must use AI. The premium will say it for them.

   Watch for: a premium discount for AI-assisted review in a professional line within the next decade.

6. Monoculture risk forces a government backstop

   A car crash is one claim. A hurricane is a hundred thousand claims on the same afternoon, and that correlation is what breaks insurers. AI risk looks much more like the hurricane. When millions of businesses run on a handful of models, one bad update can produce the same error everywhere at once. Reinsurers will cap that exposure, and when the private market will not carry it, the pattern is familiar: flood got the National Flood Insurance Program and terrorism got a federal backstop after 2001. I expect a serious proposal for a public backstop on catastrophic AI losses once the first correlated event happens.

   Watch for: the phrase “systemic AI event” in reinsurance reports, followed by a backstop bill.

7. Signatures consolidate, and the professionals form their own insurers

   If the signature is backed by insurance, then the firms with the biggest balance sheets and the best claims records can offer the most valuable signatures. That pulls work toward large firms and squeezes solo practitioners who cannot get AI-inclusive coverage at a price that works. But there is a precedent for what happens next. When malpractice insurance became unaffordable in the 1970s, doctors formed their own mutual insurance companies. I expect small professional firms to do the same for AI: pooled, member-owned cover built around shared review standards.

   Watch for: a professional association launching or sponsoring an AI liability pool for its members.

If you own a business that uses AI in front of customers, call your agent this month and ask one question: what does my policy say about AI? If the answer is “nothing,” you are holding silent AI risk, and the renewal is when it gets resolved, possibly not in your favour. Keep a record of what your AI tools tell customers. Put a real human review step on anything that could hurt someone, and put a name on it.

If you are a professional, stop thinking of your value as what you know. Your value is what you can sign, and what you can sign is bounded by what you can get insured for. Make your process insurable: reviewed, logged, explainable. That is the new craft.

If you are twenty-four, look hard at the jobs that sit next to the signature. Reviewer. Auditor. Underwriter. Claims analyst. Risk engineer. The people who decide what machine work is worth standing behind are going to be among the most important people in the economy, and very few of your classmates are looking there. Insurance is not a boring industry anymore. It is where the rules of AI are going to be written.

I hold myself to the same standard, and I have taken it one step further. Charleston AI is a service provider, and we make and keep promises in a particular order: we do the work, the customer receives it, inspects it, uses it, and only then pays. The customer pays after. That turns the usual risk dynamic around. Until the customer is satisfied, the risk that the work is wrong is ours, not theirs. It is the signature in its simplest form: I stand behind the work before I ask to be paid for it.

Here is the refinement I keep coming back to. In a signature economy, insurance gets more powerful, not less. It prices the one thing that is still scarce, it writes rules without asking anyone’s permission, and it holds the record of what actually went wrong. No coverage, no signature. No signature, no business.

But the function of insurance gaining power is not the same as today’s insurance companies keeping it. The power to price accountability is going to be fought over, and there are at least four contenders.

• The incumbent insurers

  They hold the capital, the licenses, the distribution through agents, and decades of experience turning losses into prices. Their weakness is that they learn late. They see the failures, months after the fact.

• The model makers

  They see every output as it happens, not just the ones that fail. Whoever holds the richer record can set the better price. Tesla already sells car insurance priced from its own driving data. An AI company can warrant and insure its own work the same way and step around the insurer entirely.

• The capital behind the capital

  Underwriting is itself thinking work, so it gets manufactured too. If judging a risk becomes cheap, the insurer’s expertise stops being special and what remains is the balance sheet. Power drifts toward whoever holds the most capital: the reinsurers and the large investors who stand behind them.

• The government

  When one model fails the same way everywhere at once, the private market cannot carry it, and the loss lands on the public. Whoever backstops the catastrophe gets a say in the rules. And even today, insurers are not free agents: state insurance commissioners had to approve those AI exclusions.

My prediction is that the winner is whoever holds the best record of what actually went wrong. That is the same principle that runs through this whole piece. Nobody can think their way to the right price for being wrong. It is learned from history, one claim at a time, and the party closest to that history will hold the pen.

If I had to bet, I would bet on a merger of the first two: insurers and model makers building the product together, one bringing the capital and the licenses, the other bringing the live record. The companies that refuse that partnership are the ones that lose the power the function is gaining.

Already true

Major carriers have moved to limit AI liability, standard general liability forms now include optional generative AI exclusions, specialist AI liability cover exists at Lloyd’s, and the largest labs indemnify enterprise customers against copyright claims. Insurers have shaped safety standards before, in fire, autos and cyber.

What has to happen

Claims have to accumulate. Every prediction here depends on insurers building a loss record for AI and then using it: first to price, then to set conditions. If the record grows, the questionnaire, the log, the certificate and eventually the flip all follow.

Where I am probably wrong

Courts could put liability on the model makers instead of the signers, through product liability law, and then the signature matters less and the labs carry the weight. Or legislation could move faster than I expect and set the standard of care directly, leaving insurers to follow rather than lead. I think both are less likely than the insurers getting there first, because insurers do not need anyone’s permission to ask a question on a form.