Skip to content

Working copy

Draft text, rebuilt from the manuscript on every change. Stubs and source passages are still here, and none of it is final. Back to the working copy contents

Drafted The chapter is written and still being revised.

The insurers: a pilot, then a standard

Chapter 3 asked what an insurer would need to measure before it could cover an estimating bot: a shortfall in margin across many tenders, against costs declared on bid day. This chapter turns to the insurers themselves. Some are already insuring AI, none is insuring construction prices, and the project has one thing to offer them that their own methods have to rebuild for every client.

The players on the field

Two organizations are already insuring AI, and both are a few years ahead of this project. Neither is working on construction pricing.

AIUC. The Artificial Intelligence Underwriting Company is a venture-funded firm founded in 2025. It wrote AIUC-1, a standard for AI agents with about fifty requirements across six areas: data, security, safety, reliability, accountability and society. Products are audited against it, certified for a year and retested every quarter, and the firm arranges insurance for certified products. Two of its accountability requirements come close to the estimate file: a named lead who approves each material change to the system, and a log of the system's inputs and outputs, which the standard suggests, but does not require, making tamper-evident.

Its strength is speed, which came from building on existing frameworks instead of starting from scratch. Within about seven months it had a standard built on NIST, the EU AI Act and MITRE, an accredited auditor and a first insured customer. Its weaknesses for this project are just as clear. It was written for chatbots and agents, so much of it, such as prompt injection and jailbreaks, does not apply to a weighted median of past prices. It sets no accuracy threshold and runs no statistical test. One company owns the text, the tests and the certificate, and the terms of use of its website forbid copying or adapting what is published there.

So the project should keep going on its own. AIUC-1 was written for a different kind of risk, such as a chatbot giving a wrong or harmful answer. Few AI products price and bid on someone's behalf, so nothing in it measures whether a price is fair. In this project the price is the product, and its accuracy is what drives adoption. That needs a fair point of comparison and an agreed range around it, which AIUC-1 does not have. Joining would mean writing under terms that rule out an open text, which is the main thing this project offers. The better path is a crosswalk that cites AIUC-1's requirement numbers, so a vendor could hold both certificates without doing the work twice.

Munich Re and Armilla. One of the world's largest reinsurers, Munich Re has insured the performance of AI models since about 2018. It assesses the model and how it has performed in the past, and sets the premium by the model's quality and stability. The guarantee is that the model performs to an agreed level, and when it falls short the policy pays the losses that follow. This is close to what insurance for bot procurement could look like. The agreed error would be the margin the bot promised, and the shortfall would be measured across all of the vendor's clients. The provider locks its predictions and checks them against what happens later, which is close to how this project plans to test its own tools on newer contracts. Armilla, a smaller firm in Toronto that writes policies through Lloyd's, works in a similar way. It assesses a model before insuring it, with a verification of its own that it describes as independent. Its main product covers liability for losses that an AI causes to third parties. That is a separate space, since most problems in a tender are between two parties, the buyer and the bidder.

Munich Re's strength is that it already pays claims on performance alone, with no need to prove anyone was negligent. Its method works on any kind of model, and its own material says no model can ever be perfect. The weakness of both firms is that each evaluation is custom. The metric is agreed client by client, the method is public only in outline, and two quotes from two insurers may not measure the same thing. Neither firm knows much about tendering law or the gap between a bid and a cost.

So the project is not a competitor to either firm. It does not sell insurance and has no financial interest in doing so. What it can offer is the common point of comparison that their processes have to rebuild for every client. A public way to measure the misses, a public test, and an agreed range for construction pricing would make each evaluation cheaper and each quote comparable, and it would give any insurer, not only one, something to write a policy against.

Many market issues are still open. There is no loss history yet that links a low bid to real losses. Many bidders using the same standard could all fail at once. The bidders keenest to buy coverage may be the riskiest ones, and an open range can be gamed. The market may also be too small for an insurer to build a product alone. The project is not meant to solve these. It exists to show that the conversation is worth having.

Not every bidder needs this. A small company can hold its whole picture in one or two heads: its equipment, its crews, its costs and how much risk it will take. A vanilla AI or a home-built tool already gives it a bid to review, and paying for software and insurance on top would turn profit into overhead. Medium and large companies are different. Their estimating departments have middle managers and span regions, and they bid on the largest jobs, where a losing bid wastes the most effort. For them, insured bot procurement is a way to standardize estimates and raise productivity while keeping risk under control. These are the bidders a pilot should start with.

The sureties

Sureties are what the rest of this book calls bonding companies. Surety is the industry's own word, and this section uses it because it is written for them. They have stood behind construction bids longer than anyone, though what they guarantee is the builder, not the price. On most public work in North America a bidder needs a surety's bond to bid and another to sign the contract, and if the builder fails, the surety pays to finish the job.

A surety is closer to a lender than to an insurer. The contractor, and usually its owners, sign an indemnity agreement promising to repay any loss, so the surety pays only when the firm cannot. It underwrites the firm rather than the price: its character, its capacity to do the work and its capital. The method works. In 2023 American sureties paid out about 22 cents in losses for each dollar of premium they earned, and the line has made more than $2 billion a year in underwriting profit since 2022.

Sureties already handle prices at two moments. Right after the bids open, they look at the spread. One large surety asks its contractor for a letter whenever the next bid is more than 10 percent higher, with five questions: are you comfortable with the bid, was anything left out, were there arithmetic errors, what profit do you expect and why, and is there an engineer's estimate. On Texas highway work that letter would go out on about one job in three (chapter 3). Over time, they track a contractor's bids for patterns, such as a firm that keeps underbidding, and when a pattern shows they ask for more frequent reports on the work in progress.

A check before sealing adds little to the first moment. Estimating software already does the arithmetic, bid-levelling tools look for gaps in scope, bid-history services give the comparison, and the spread is the market's own free second opinion. Nor does it touch the surety's biggest losses. When the surety industry reviewed 86 of its claims, unrealistic growth led the causes at 37 percent, including big jumps in the size of single jobs; estimating sat inside accounting problems, at 29 percent. A better price does not give a firm the supervisors, managers and accountants it needs to run a bigger job, and sureties know it.

Where the check could matter is the public backstop. The US Small Business Administration guarantees most of a surety's loss, 80 to 90 percent, on bonds for small contractors, on contracts up to $9 million, or $14 million on some federal work. In fiscal 2025 it backed $10.6 billion of contract value for more than 2,200 small businesses. Here the buyer of a check would be the agency, not the surety. It could ask for a recorded check with each guaranteed bid, or charge less for bids that passed one, and so back more small firms for the same risk. How much that would save depends on how many of the program's claims begin with an underpriced bid, and the program does not publish its claims.

One correction to the way this project first imagined sureties. In a fight over whether a low bid was a slip or a bet, the surety stands with its contractor: when the bidder proves a slip, the surety is released from the bid bond too. A record that tells slips from bets protects the buyer. For the surety its use comes earlier, in knowing quickly whether to back the bidder's claim or to get ready for a losing job.

The ask for sureties is narrow. What would a bid file need to show for a reading like this one to count in your review of a bid?

What we need now: the ranges an off-the-shelf policy would require

The project would like input from underwriters on the ranges and variance an off-the-shelf insurance policy for estimating bots would require. How far off may a bot's price be, how often, and how steady must that be before an insurer could sell a standard policy rather than a custom one?

What we need later: a pilot, then a standard

The ask will change as the tools improve, roughly in three steps.

  1. Once the reference Estimator and Model are good enough, the next step is a pilot with a real insurance policy. One insurer would write a small, capped policy on the reference tools, used by a few willing bidders on real tenders. The pilot would produce real claims or near misses, and policy terms that have been tested in practice. A lower-stakes start could be pricing change orders after a contract is awarded, where no bid is at risk. A draft term sheet for such a pilot is in Appendix E, for an underwriter to mark up.
  2. The pilot then feeds the standard. Large-scale adoption needs firmer limits on liability than any single policy can give, and an industry standard is where those limits get written. The policy terms tested in the pilot would become its first requirements.
  3. With the standard in place, insurers could cover any product that passes the same public test. Private companies and programmers could then walk through the door, if they find the market attractive enough. They can build much better estimating software than this project, and the reference tools would stay only as the floor everyone is measured against.

The reference tools do not need to be the best. They need to be good enough to run the pilot, build the standard, and build the market.

What would an insurer gain from the pilot? Not a proprietary product. Whatever the pilot teaches would go into an open standard, and that may cool some appetite for the experiment. What the insurer would gain is first-hand experience with the technology and the standard before anyone else, which is a real first-mover advantage. It would also help build a field that bidders are willing to enter, because they could finally put a value on the insurance. A proprietary product would be worth more to one firm, but a field is far more likely to grow into a market, to the benefit of the first mover, with latecomers trying to keep up.