The Assistant’s Number and the File Disagree — Which One Wins?

Short answer: the file. The verified, dated, account-specific record — the loss runs, the signed application, the inspection, the exposure figures you can point to — governs what you present to a client, what you recommend, and what you stand behind. A number an underwriter surfaced out of a chat window is class-level data delivered at chat speed: an input to a conversation, not a fact about your account.

But “the file wins” is not a slogan you get to say and move on. It means you trace what the number was built from, check your own file with the same suspicion, work the disagreement to a decision, and write the note that proves you did. Because two years from now, if this account is sitting in an E&O look-back, nobody will ask what the assistant said. They’ll ask what your file shows.

Why is this disagreement suddenly an everyday event?

Because the number now arrives mid-conversation, with total confidence, from the other side of the table. Since Verisk’s October 28, 2026 announcement, authorized insurer staff can pull ISO Loss Costs and PAAS content through Microsoft Copilot, Claude and Amazon Quick — our breakdown of the announcement walks through what was put where. Brokers got no login. You hear a figure quoted by an underwriter who may have gotten it thirty seconds earlier; you never see the query behind it.

And the two sides were never measuring the same thing. A loss cost is pooled class data, built from thousands of accounts that are not yours; PAAS content is classification and audit guidance, not a ruling on this insured. Your file holds the opposite: this account’s actual losses, this building, this payroll. Neither cancels the other — but the collision is now a weekly event, so it needs a standard handling, not improvisation.

Why does the file win?

Because the file is the only side of the argument you can verify and produce. Every item in it has a source and a date: loss runs issued by a carrier, an application the insured signed, an inspection somebody performed, payroll records the client provided. Challenged by an underwriter, an auditor, a plaintiff’s attorney — you can lay it on the table.

A chat-surfaced figure can’t do that. It came out of a content entitlement, under one user’s permissions, in conversational language, possibly built on inputs you never saw. It might be right. It might be the class number quoted as if it were the account number. You can’t tell from where you sit — and a number you can’t trace is a number you can’t defend. Repeat it to a client and it’s wrong, and the repetition is yours: “the carrier’s assistant said so” is a defense nobody accepts.

One honest qualification: the file wins as the governing record, not because your paperwork is infallible — verified facts govern. Which is why the next two steps exist.

Before you rule for the file, what do you check on the number?

Trace the build. Most disagreements die here — most turn out to be one of four ordinary errors:

1. Which classification did it use? If the assistant worked off a different class than the file carries, you’re not disagreeing about price — you’re disagreeing about what the business is. Classification disputes get settled with evidence about operations, not by averaging two numbers.
2. Which version, which effective date? Loss costs move on effective dates. A figure built on last cycle’s loss cost and a figure built on the current one will “disagree” while both are internally consistent. Ask which one sits behind the quote before you argue with it.
3. What inputs did it run on? Payroll, sales, square footage, territory — do they match the submission, or an older version of it? Assistants answer from what they’re fed, and the feed is often the first draft of the account.
4. Is that the class number or the account number? The assistant surfaces class-level content. Credits, debits and account-specific judgment are applied by the underwriter afterward. If the layers got blended, separate them out loud and ask which one you’re looking at.

Ask the underwriter to walk the build with you — a good one can. If nobody can say what a number is made of, that is itself the finding, and it goes in the note.

And what do you check on your own file?

Files go stale quietly: loss runs that stop a year short, operations that changed at the last renewal and never made it into the application, payroll from two cycles ago. Our piece on loss runs, COPE data and large-account underwriting covers how much weight underwriters put on current, complete loss history — and how fast a thin file loses an argument.

Sometimes the assistant is right and your file is the stale side. Then fix the file, note what was wrong and what replaced it, and move on. Credibility with underwriters is built exactly this way — the broker whose file corrects itself on the evidence is the broker whose file wins the next argument.

How do you document the call?

Write the note the same day, while the disagreement is still warm. Six items, in the file, in your own words:

1. The number as quoted — the figure, who quoted it, the date, and what it was said to come from (which platform, which data).
2. The build, as described to you — classification, loss-cost version and effective date, the inputs used. If the build couldn’t be described, write that.
3. The file fact it conflicted with — and that fact’s source and date. “Loss runs issued by the carrier, dated March 2026” is a fact. “Our records show” is not.
4. What you did to verify — the checks above, who you asked, what they answered.
5. The decision — which figure you presented to the client, why, and who made the call.
6. What the client was told, and when — including any correction of a number they’d already heard.

The test for the note: a stranger — an E&O adjuster, your successor on the account, you in two years — can reconstruct the decision without phoning anyone. This is the same discipline the policy check has always demanded — our piece on the policy check as the E&O front line makes that case, and our errors and omissions insurance explainer covers what the look-back examines. The narrower point here: that standard now applies at the indication stage too, because the indication is where the AI number enters the story.

When do you escalate to the underwriter?

Escalate — in writing, with the file evidence attached — when any of these is true:

• The gap is material. It changes the premium in a way the client will feel, or it changes what you’d recommend.
• You’ve traced the build and the file still contradicts it. That’s a genuine dispute, not a misunderstanding — it deserves a named decision-maker.
• The classification is what’s really in dispute. A class call follows the account past this renewal, into audits and future placements — don’t let one get set by an unexamined chat answer.
• Nobody will put the number in writing. If the figure can’t survive being written down by the person who quoted it, it doesn’t get carried to your client.
• The client has already heard the number. Then any correction has a clock on it.

State the question specifically. Not “this seems off,” but: “Your indication uses a loss cost effective [date] under class [X]. The file shows [Y], sourced from [document, date]. Which governs, and can you confirm in writing?” The answer, whichever way it lands, goes in the file.

What do you tell the client while it’s unresolved?

The truth, plainly: “Your carrier’s system produced one figure and your file says another. I’m reconciling them before I bring you a number I’ll stand behind.” Clients understand that instantly. What damages trust isn’t a delay; it’s a confident number that quietly gets replaced later.

And if you already repeated the chat number: correct it promptly, plainly, in writing, and document the correction the same way. A fast, clean correction is a professionalism story. A slow, silent one is an E&O story.

FAQ

Does the file always win?
No — verified facts win. If your file is stale and the number’s build checks out, the file gets corrected and the correction gets documented. What never wins is an untraced number: a figure nobody can say was built from what, as of when.

What if the underwriter insists the assistant’s figure is right?
Ask for the build in writing: classification, loss-cost version and effective date, inputs. If it checks out against the file, update the file and note why. If it can’t be produced, escalate: you’re being asked to rely on something nobody will stand behind in writing.

Is saving a screenshot of the chat enough documentation?
It’s a start, not the note. The screenshot captures what was said; the note has to capture what you did — the verification steps, the decision, and what the client was told.

What if the client already acted on the chat number?
Correct it fast, in writing, and document the correction the same day. Where liability ultimately sits is a live question — our piece on who pays when an AI agent makes the mistake works through it — but speed and a clean record are what’s inside your control.

Does this only apply to Verisk data?
No — any figure out of any carrier’s AI assistant gets the same treatment: trace the build, check the file, decide, document. The tools will keep changing. The discipline doesn’t.

Source: Verisk Analytics, Inc., “Verisk Brings Its Proprietary Insurance Intelligence to Amazon and Microsoft AI Platforms, Expands Offerings within Claude,” GlobeNewswire, October 28, 2026. What the integrations provide, and to whom, is as stated in that release; the documentation and verification practice in this article is the desk’s, not Verisk’s.

Informational only. This article explains broker documentation and verification practice for insurance professionals; it is not legal, compliance or coverage advice, and it does not describe any specific carrier’s systems or any specific account’s pricing.

Updated Nov 1, 2026 — added the free file wins quick card, file wins disagreement pack, and file-wins bot book downloads to this guide.