Two Carriers’ Assistants, Two Different Numbers: How Do You Reconcile Them in Front of the Client?

Updated Nov 1, 2026 — added the free two-numbers quick card, two-carrier reconciliation pack, and reconcile-it bot book downloads to this guide.

Short answer: you don’t average them, and you don’t hand the client the lower one. Two different chat-surfaced numbers on the same account usually aren’t a data error — they’re two different slices of the data, two different versions, or two different questions. Your job is to pull both numbers apart until you can see what each one is actually made of, go back to the facts in your own file, and put exactly one defensible number in front of the client — with the reconciliation documented behind it.

Since Verisk’s October 28, 2026 announcement, authorized insurer underwriters can pull ISO Loss Costs and PAAS content through Microsoft Copilot and Amazon Quick, and through Claude with property-level Benchmark weather history added. Brokers get no login — you hear these numbers secondhand, from the underwriter across the table. Sooner or later, two carriers working the same account will come back with two different figures, both delivered at chat speed, both sounding final. That moment is yours to referee. We broke down what Verisk actually announced, and who gets access, in our piece on Verisk’s data moving into Copilot, Claude and Amazon Quick. This piece is about what you do when the answers don’t match.

Why would two assistants give two different numbers on the same account?

Because “the Verisk number” isn’t one thing. The client will assume one of the carriers made a mistake — usually, neither did.

Different entitlements, different slices. Verisk was explicit in the announcement: existing content entitlements, permissions and governance controls continue to apply inside the AI tools. Two carriers may not be entitled to the same content, class detail, or Benchmark history — so their assistants are answering from different slices of the data.

Different versions. Loss costs move on effective dates. An assistant surfacing last cycle’s loss cost and one surfacing the current one will produce two clean, confident, different numbers. Neither looks stale in a chat window.

Different questions. A chat answer reflects the prompt behind it. One underwriter asked about the class; the other asked about the class in a specific territory, or about the audit classification under PAAS instead of the loss cost. Change the question and the answer changes — without anything being wrong.

Class number vs. account number. Loss costs are class-level advisory figures — pooled expected losses for a class and territory — not an answer about your client’s operation. If one carrier quoted the raw class figure and the other applied account-level judgment on top of it, you’re not looking at a disagreement about data. You’re looking at two different stages of the same pricing process.

What’s the first move when the numbers don’t match?

Slow it down, and treat the gap as information, not a problem to make disappear. The two failure modes are both tempting: split the difference, or lead with the lower number because the client will like it. Averaging two figures you don’t understand produces a third figure nobody can defend. Leading with the lower one turns a chat answer into a promise.

The professional move is boring and it works: acknowledge both numbers, tell the client (and yourself) that two different indications came back and you’re reconciling them before anyone relies on either, and buy the time that takes. A broker who says “I have two numbers and I’m not repeating one until I know what each is built on” sounds stronger than a broker who guesses fast — and it’s the version of events you want in the file if the final number lands higher than the first one the client heard.

How do you pull the two numbers apart?

Ask each underwriter the same four questions, and compare the answers side by side. In most cases the gap explains itself before you finish.

1. Which content, which version, which effective date? Loss cost or PAAS guidance — and as of when? A version difference is the cheapest explanation there is, and the easiest to fix.
2. Which class and which territory? If the two carriers classified the account differently — or one of their assistants did — you don’t have a pricing dispute, you have a classification question, and that’s a different conversation with different evidence.
3. What exposure base is under it? Payroll, sales, square footage — if the assistants were fed different exposure figures, the outputs differ for arithmetic reasons, not judgment reasons. Check the inputs before you argue about the outputs.
4. Is that the class figure or the account figure? Has the underwriter applied your client’s actual loss history, credits or debits yet — or is this the raw advisory number? Knowing which stage each number came from tells you how far apart they really are.

If you want a sense of how much of this the underwriter can already see from their side of the desk, our breakdown of what underwriters actually see when they look at your submissions is worth the read — mismatched classifications and stale exposures are visible to them now at the same chat speed.

What wins when the inputs check out and the numbers still disagree?

The file. Your client’s actual loss runs, actual exposures, actual operations — the account-specific facts that class-level data can’t contain. Loss costs are a starting point built from pooled data; the underwriter’s job is to apply judgment on top of them, and your job is to make sure the judgment is being applied to the real account.

So ask each underwriter to put their basis in writing: the class, the effective date, the exposure base, and what account-level adjustments were or weren’t applied. The number that survives being written down is usually the number that was real. The one that can’t be reconstructed outside the chat window gets treated as what it is — a research shortcut, not an indication. And keep your carriers’ answers in context: carriers are already answering to regulators for how they validate AI-surfaced data, the landscape we mapped in AI governance in insurance and the 2026 regulatory reckoning. Asking “show me the basis” is a normal professional request in that world, not an accusation.

What do you actually say to the client?

One number, a range with a reason, or a timeline — never two competing figures with no referee. If the reconciliation is quick, the client hears the final number and, briefly, why an earlier figure moved: “Two carriers priced it off two different classifications; we confirmed the right one, and here’s the number that holds.” If it isn’t quick, the client hears the honest version: “Indications are coming back with a spread because the carriers are reading the classification differently. I’m reconciling it before I quote you, because the wrong one would cost more to fix at audit than it saves today.”

What the client should never hear is the lower chat number framed as the price. The moment you repeat a figure, it’s yours — whoever’s assistant produced it. Presenting one defensible number, and being the person who checked it, is the whole value of having a broker in this workflow. The assistants made the data fast. They did not make it reconciled.

What goes in the file afterward?

Both numbers, and the trail between them. Record each figure you were given, which carrier gave it, the content and effective date behind it, the class and exposure base it rested on, the questions you asked, and the basis the final number stands on. That record is the difference between “the broker relied on a chat answer” and “the broker tested two indications against the file and documented the call” — and if the account is questioned, at audit or in a dispute, that difference matters. The policy check has always been the E&O front line in an agency; our piece on the policy-check agent and the E&O front line makes that case, and chat-surfaced numbers extend the same rule to indications: if you acted on it, the file should show what you checked.

FAQ

Does a different number mean one carrier’s AI made a mistake?
Usually not. Different entitlements, versions, classifications or exposure inputs produce different answers from the same underlying data. Assume an input or scope difference first; call it an error only after the inputs check out identical and the outputs still differ.

Should I just quote the client the lower number?
No. A chat-surfaced figure is a research output, not a quote. Repeating the lower number unreconciled makes it your promise — and makes you the one explaining the gap at binding or audit.

Can I ask the underwriter what they asked the assistant?
Yes, and you should when a number matters. Asking for the content, version, class and exposure base behind a figure is the same diligence brokers have always applied to a rate sheet — the tool being a chat window doesn’t lower the standard.

What if the client has already seen both numbers?
Don’t pretend there’s one. Explain that two carriers built the figure two different ways, say what you’re verifying, and give a time when you’ll come back with the reconciled number. Clients forgive a explained spread; they don’t forgive a number that quietly changes.

Will this happen on every account now?
On routine accounts, probably rarely — the mainstream classes are the mainstream classes. Expect it where classification is genuinely arguable, where loss costs recently moved, or where the account’s real exposures don’t sit neatly in the class. In other words, exactly the accounts where a broker earns the fee.

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 include, who is eligible, and the continuation of existing entitlements and governance controls are as stated in that release. The reconciliation workflow above is broker practice, not a Verisk product description.

Informational only. This article explains an industry development 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.