Somewhere in your office this quarter, someone will evaluate an AI coding seat for the staff who build your automations, your submission tooling, your renewal spreadsheets. The seat most likely to be tried is Meta’s Muse Code, because its frontier model is strong and its plans are cheap. Before anyone pastes client work into it, read what Meta’s own documents say about the setting that seat ships with — because the papers a brokerage handles every day are precisely the class of material both Meta and the Big “I” tell staff to keep out of a training arrangement.
What a submission is, on paper
A commercial submission is not an abstraction. A wholesaler’s published applications checklist asks the retail agent to assemble ACORD 125 and 126 for general liability, ACORD 131, 140, and 146 when umbrella, property, and inland marine are needed, copies of underlying policies, a supplemental application with detailed descriptions of operations, five years of currently valued loss runs, and prior carrier history. The recommended items add company financials, resumes of the principals, and detailed descriptions of any losses.
The claim side is the same class of paper: date of loss, claim details, open or closed status, payout amount, whether repairs are complete. That is the file your producers and account managers open, summarize, compare, and draft from — and it is the file that ends up pasted into whatever tool the office is trying this month.
What the seat does with it, by default
Meta’s help page states the arrangement in one paragraph: “Muse Code subscriptions default to Contributor mode. In Contributor mode, you get more usage from your plan in exchange for letting Meta use your content — the prompts, code, files, and context you send, plus the model responses you get back — to train future Meta models.” The Terms count documents among Inputs, including content a user authorizes the service to access as context or background.
Meta’s instruction to the user of that seat is printed on the same page, and it does not hedge: “You must not submit sensitive, confidential, or personal information to the Discounted Services. If your use case involves processing personal information, please use Standard Services.” The Terms add the contract case in so many words: if you intend, or are required by contract, to keep information confidential, you must not submit that information to the Contributor Services.
A loss run is personal and confidential information by any reading a broker would apply to their own file. A submission carries client financials. The default setting of the seat is the opposite of the handling rule the documents themselves prescribe.
Two traps in the off-switch
The fix is real: open /settings in Muse Code and turn Contributor mode off; the plan stays the same and the usage amount changes. But the switch has two edges a brokerage needs to know before it relies on it.
First, it does not reach backward. Meta’s Terms make the change prospective only: the settings change “would not apply to Content previously submitted to the Contributor Services,” and Meta’s rights in that content survive the change. A seat that ran on the default for a month has a month of submissions under the training grant, and flipping the toggle in week five does not pull them back.
Second, context reclassifies. “Content previously submitted to the Standard Services becomes Content submitted to the Contributor Services if it is included in any request to the Contributor Services, whether directly or as context or background.” A clean session’s material, quoted or attached into a later session on a seat still running the default, changes tier with it. In an office, seats and sessions multiply faster than settings get checked.
What guidance actually exists
The association document that speaks to agency staff and AI tools is the Independent Insurance Agents & Brokers of America’s October 2024 legal-issues memo on artificial intelligence, with its sample acceptable-use policy. Its examples are this office’s examples — an agency using AI “to create client proposals that incorporate personally identifiable information” — and its warnings are plain: employees “should not be permitted to input confidential, proprietary or sensitive information, such as PII, without adequate protections in place.” The sample policy’s rule: “Do not upload confidential, sensitive or proprietary information to AI platforms.”
The same memo flags the contract layer: the Big “I”‘s general counsel had reviewed an agent appointment agreement that would “prohibit the use of AI in connection with carrier data without express prior permission,” and agents are told to check their appointment contracts. A carrier’s loss runs, processed through a seat whose vendor trains on the input, is a fact pattern worth checking against your own appointments before it happens, not after.
Be precise about what does not exist: no regulator bulletin retrieved for this piece tells a brokerage what its staff may paste into an AI seat. The NAIC’s model bulletin on AI systems is addressed to insurers and governs systems that make or support regulated insurance decisions; it is not a producer rule for office tooling. The operative documents for this decision are Meta’s terms and help pages, and your own acceptable-use policy. That gap is exactly why the policy has to be written by the brokerage rather than looked up.
The rule to write down
Treat the seat like any other system that touches the client file, and put four lines in the acceptable-use policy. One: every AI seat gets its data setting checked at setup, by a named person, before client work — the default is a decision, and it should be yours. Two: submissions, loss runs, financials, and claim papers never go into a seat or tier whose terms grant training rights over the input. Three: client-shaped work runs on standard, no-training arrangements — on Meta’s platform that means the API in Standard mode, or the seat with Contributor mode off, verified. Four: re-check the setting when seats, plans, or staff change, because the prospective-only switch means a lapse is not cured by discovering it.
None of this is an argument against the tool. It is an argument that a $50 seat is a data-handling decision wearing a productivity purchase’s clothes. The brokers who get this right will be the ones who read the vendor’s own warning label — “You must not submit sensitive, confidential, or personal information” — and wrote it into their office rules before the first submission was ever pasted.
The same launch, read for the operator who buys the seat, is on Tygart Media: Wonder Full Stays in the Seat. The standing reference for the cover this touches: Errors and Omissions Insurance — What E&O Covers, Who Needs It, and Where It Surprises People, and this desk’s earlier look at where the policy check went: The Policy Check Was the E&O Front Line. Now It’s an Agent’s Job.