Your AI Won't Bind If Your Carrier Can't Read It

October 1 is not a fall problem. Brokers are collecting commercial cyber, errors and omissions, and general liability applications this week. If you run HVAC, plumbing, waste hauling, or any local service shop whose software now answers the phone, the 2026 packet is not last year's packet.

Carriers and wholesalers added generative-AI and automated-decision supplements. That did not come from a product webinar. It followed the NAIC model bulletin on insurers' use of AI systems, then a summer that already put failure modes on buyers' desks: Anthropic reported elevated errors across Claude models on August 5. The August 10 CEVA incident is already in their files. Underwriters read the same news. They are not asking whether your receptionist sounds human. They are asking whether they can price a decision your software made in your name.

The named insured is still you

Most AI-and-insurance coverage stays on the carrier side of the table: model bias in underwriting, unfair claims handling, the NAIC bulletin as a constraint on the insurer. You have the inverse problem.

Your stack quotes jobs, takes cards, texts vendors, and writes QuickBooks. The named insured is still the contractor. If the quote is wrong, the card data is stored badly, the vendor is sent to the wrong site, or the ledger entry cannot be reversed, the claim does not land on the model. It lands on your policy.

That is the part vendors skip in the demo. A voice that "sounds human" does not survive an exclusion for automated decisions. A SOC 2 letter does not reconstruct who authorized a $4,800 emergency quote at 9:14 p.m. A one-pager that says "we use encryption" does not tell Travelers or Hiscox which service account posted the invoice, what left the call recording, or how you unwind the books.

We already argued, in Your AI Runs on Your Infrastructure. Who Controls It?, that a customer-controlled server is not the same as a customer-controlled operation. The renewal form is that argument with a premium attached. Credentials, logs, approvals, and reversals are not a governance hobby. They are what the supplement is scoring.

Insurability is a product capability. If the vendor cannot produce an evidence pack your broker can forward without rewriting it in Word, you are not automating the back office. You are originating uninsured operational risk.

What the form is actually asking

The questionnaire looks bureaucratic. It is not. Each line is a proxy for a claim scenario.

• What decisions happen without a person? Quoting, dispatch, payment, customer create, ledger post. "It answers the phone" is not an answer. List every action that moves money, a crew, or the books.
• Who is the legal actor? The AI is not the named insured. If a human must approve above a threshold, state the threshold, name the role, and show the audit row. No threshold is an answer. Silence is how you buy an automated-decision exclusion.
• What data sits on the call? Recordings, transcripts, addresses, access codes, card PANs, gate codes. Underwriters want collection, retention, residency, and who can export a file after a dispute.
• Which identities does the agent use? A shared QuickBooks admin login, a Twilio number in the owner's name, a Stripe key in a vendor portal: those are how a cyber application gets referred.
• Can you reverse the write? Quote by SMS, vendor text, invoice in QuickBooks. If you cannot show the void, the compensating entry, and the customer notice, you cannot show a controlled system.

A finished prompt is not a finished job. That is the same gap we called out in The $2B AI Bet Misses the Workflow. A finished job that cannot be reconstructed is worse: it is an uncovered one.

The evidence pack you should forward this week

Do not send a marketing PDF. Send artifacts. Score each item 0 (missing), 1 (partial), or 2 (underwriter-ready). A serious vendor should clear 16 of 20 without a custom statement of work.

1. Decision inventory (2): One page of every action the agent can take, marked human required, human optional, or autonomous. Include quotes, card capture, vendor SMS, job create, and ledger posts.
2. Authority matrix (2): Dollar and scope limits, who can raise them, and how an override is recorded. Name the role, not "the team."
3. Identity and secrets map (2): Every service account, OAuth grant, and API key the agent uses for QuickBooks, telephony, SMS, payments, and calendar. Scope, rotation owner, no shared owner logins.
4. Data-class register (2): What is captured, stored, where it lives, how long it is kept, who can export it. Separate lines for card data, recordings, and access instructions.
5. Subprocessor and model list (2): Legal names, function, data that leaves your tenant, contract vehicle. "We use a leading LLM" scores a 0.
6. Reconstruction sample (2): One redacted job with transcript pointer, tool-call log, quote versions, dispatch message, QuickBooks transaction ID, and any human approver. Timestamped. Exportable in 15 minutes.
7. Reversal runbook (2): How you void the quote, recall the vendor, refund the card, and reverse the books, including who is on point after hours.
8. Incident notice to the carrier (2): How you tell the broker if the agent is compromised, sends a bad price, or writes a bad ledger batch. Time-to-notice, not a brand statement.
9. Retention and legal hold (2): Recording and log retention that matches the E&O and cyber applications, plus the ability to freeze one job file without taking the phones down.
10. Exclusion read-through (2): Written vendor note on which forms they reviewed and which of their actions they believe are uninsured unless you endorse.

Print the scores. If the vendor cannot sit with ops, IT, and your broker and fill this in, they cannot sit with an underwriter either.

Treat the packet as the RFP

Most buying committees still run the demo first and the insurance question last. Reverse it for any system that can quote, take money, move a crew, or touch QuickBooks.

This week, before anyone submits the October 1 packet:

1. Have the broker send the actual AI / automated-decision supplement, not a summary.
2. Walk the ten items above with the vendor on a shared screen. Record the score.
3. If an item is a 0, turn that action off before the effective date or get the exclusion in writing so you know what you are self-insuring.
4. Put the reconstruction sample in the submission. Underwriters believe artifacts. They do not believe adjectives.

You do not need a 40-page AI policy. You need a packet that lets a human at the carrier follow one job from the call to the books and see who was allowed to do what.

When we look at a voice-to-dispatch-to-books path, we treat that underwriter pack as a product requirement, not a legal appendix. If a decision cannot be reconstructed, it should not be automated.

Ask your vendor for the pack this week. If they send a one-pager, send them the scorecard back. Your broker can use either answer.