Most functions inside a 50-to-500 person company can buy an AI license, train the team on their own files, and be done. An insurance agency cannot, and the reason has nothing to do with the software. An agency sits between a carrier and an insured, and both relationships arrive with paper governing what may leave the building, who may say what to a policyholder, and who pays when a sentence turns out to be wrong.
So the purchase runs in reverse here. Draw the line first, then buy the session that teaches people to work right up to it. What follows is the line, which rules do and do not bind an agency, the order to train roles in, what the options cost, and the four numbers to put in front of ownership at ninety days.
What agencies are actually buying
Search this category and almost everything that comes back was written for a carrier. Underwriting automation, claims triage, fraud scoring, rate modeling: real problems, none of them yours. The EY-Parthenon survey of 100 senior decision makers at leading insurance firms in early 2025 found 27% saying agentic AI was partially integrated and 57% expecting use cases fully built and integrated within one to two years. That is carrier capital spending, and it tells an agency principal nothing about Monday.
An agency’s version is narrower and easier to measure. Renewal prep, submission packaging, loss-run summaries, service email that reads like a person wrote it, meeting notes that reach the management system the same day, proposal drafts a producer checks line by line. That is fluency work on documents the agency already owns, which is the general case we make for why buying licenses is not the same as building capability.
The appetite is there and the trust is not. Reporting in the Big I’s IA Magazine in November 2025 put 41% of independent agents planning to adopt AI within six months, 17% trusting the technology, and 27% treating it as a threat. A room with that spread in it is not moved by a tool demo. It is moved by a rule everyone can repeat.
Three documents decide what AI can touch
Agencies often assume the limit is regulatory. Mostly it is contractual, and that changes who you negotiate with.
The carrier agreement. The NAIC Model Bulletin on the use of AI systems, adopted by the Executive Committee and Plenary on December 4, 2023, is addressed “To: All Insurers Licensed to Do Business In” the state. Not to producers. An agency is reached through the third-party section of the program the bulletin asks carriers to write, which tells them to run due diligence so that decisions supported by outside systems “will meet the legal standards imposed on the Insurer itself,” and to seek contract terms giving audit rights. Your obligations arrive as a carrier agreement amendment or a marketing-rep questionnaire. Read the amendment. That is where the rule lives.
The E&O and liability program. This one moved recently and in the wrong direction. ISO filed three optional generative-AI exclusions for commercial general liability, CG 40 47, CG 40 48 and CG 35 08, effective January 2026, defining generative AI broadly as a system “trained on data with the ability to create content or responses.” An agency meets those forms twice: as a business whose own coverage may narrow, and as the party placing them for commercial clients. Ask your E&O carrier in writing how it treats a claim arising from model output before anyone drafts a coverage summary with one.
The vendor contract. The same IA Magazine reporting notes that technology agreements often cap vendor liability at the last six or twelve months of fees, which is not a number that absorbs a coverage error. Swiss Re’s James Redeker offers the cheapest control available: run the model and the existing process side by side for the first few months and compare. That is also the only honest way to learn whether a workflow is worth keeping.
One rule sits outside all three. The FCC confirmed in a declaratory ruling on February 8, 2024 that the TCPA’s restriction on artificial or prerecorded voice calls covers AI-generated and cloned voices. Synthetic outbound calling to policyholders is a consent problem, not a tooling choice. Put it on the prohibited list on day one.
Draw the boundary before the session
Write the list down, post it, build the agenda around it. Three colors, no committee.
| Status | Work | Why it lands there |
|---|---|---|
| Green | Internal summaries of loss runs, policies and carrier email the agency already holds | In-house source, read by a licensed person before it moves |
| Green | Renewal agendas, submission packaging, meeting notes into the management system | Clerical reformatting of the agency’s own records |
| Green | First drafts of service email, proposals and marketing copy | A producer edits and signs; nothing sends unread |
| Amber | Coverage comparisons and recommendation language | High value, high error cost. Needs line-by-line review against the real forms |
| Amber | Client-facing chat and quote intake | Turns on disclosure, consent and what the carrier agreement permits |
| Amber | Carrier-supplied appetite, rating or claims data in a non-approved system | A contract question for the marketing rep, answered before the pilot |
| Red | Binding, declining, or quoting outside appointed authority | Licensed decisions. A model holds no authority |
| Red | Synthetic voice calls to insureds | The 2024 FCC ruling puts these under TCPA consent rules |
| Red | Non-public personal information in a personal consumer account | The most common real breach, and a policy gap rather than a training gap |
The amber rows are the curriculum. Green is a half-day and red is a memo. The money sits in teaching people to work amber safely: judgment practice on live files, not a lecture on prompting.
Who to train first at an agency
One session for the whole staff is the standard failure, and it fails faster here because a commercial lines account manager and a benefits producer share almost nothing except the carrier portal login. The case for splitting curricula by job sits in our role-based enablement guide. This is the agency cut, in the order we would run it.
| Order | Role | Where the hours are | What they leave holding |
|---|---|---|---|
| 1 | Commercial lines account managers | Renewal prep, submissions, loss-run summaries | A renewal workflow run on their own book |
| 2 | Personal lines service staff | Service email, endorsements, policy-change notes | A documentation sequence with a review rule |
| 3 | Producers | Prospect research, proposal drafts, meeting notes | A proposal draft checked against the real forms |
| 4 | Claims advocates | Claim summaries, status updates, carrier follow-up | A summary format with a named checkpoint |
| 5 | Operations and accounting | Reconciliation prep, reporting, process notes | One recurring report in half the time |
| 6 | Marketing and principals | Content, campaign drafts, carrier reporting | A content workflow and its approval rule |
Service roles go first because the work is text-heavy and repetitive, so the gain shows inside a week and nobody argues about it. Principals go last because their session is mostly boundary and budget, which land better once they have watched the rest of the agency use the tools. Shared service teams will recognize the pattern from customer support and success enablement, where the same review discipline decides whether speed turns into rework. A 25-person agency runs these as modules in one day rather than six sessions.
The confidentiality block is close to what in-house legal teams cover on privilege and drafting. Agencies hold medical details in benefits files, loss histories, payroll figures and Social Security numbers. The rule is not “be careful.” It is a named list of approved tools, a named list of data classes, and one person who answers for both.
What the options cost
Per-seat pricing reads fine for one person and stops making sense around the eighth. Verified prices, against a 30-person agency:
| Option | Verified price | 30 people | Where it falls short |
|---|---|---|---|
| The Institutes, Associate in Insurance AI | $415 per course across AIAI 301, 302 and 303, plus exam fees | about $37,350 in course fees alone | Real insurance spine, built for the industry rather than your workflows |
| Certified AI Insurance Credential | $997, nine self-paced modules and a proctored exam | $29,910 | Good on governance and E&O. Priced for a few people, not a staff rollout |
| Coursera Teams | $399 per user per year for 2 to 499 licenses | $11,970 a year | Large library, easy procurement. Completion is not capability |
| Self-paced catalog courses | $20 to $200 per seat | $600 to $6,000 | Cheap reference material. No practice on your files |
| Hands-on team session | $2,000 to $15,000 in the market by group size and scope | one line item | Needs a follow-up pass and a named owner for the workflows |
Two things fall out. The credentials are worth buying for two or three people, as designed, and they are not a staff program at agency headcount. The cheap library belongs underneath whatever you run, as the place people look things up afterward.
The trap is buying a credential for the principal, calling the agency enabled, and finding at renewal that nobody in service works differently. The same confusion shows up when executive and leadership programs stand in for a working team curriculum.
The break-even, in hours
Write the threshold down before the first session, in recovered hours rather than in a percentage.
Bureau of Labor Statistics data puts the median pay for insurance sales agents at $62,280 a year, or $29.94 an hour, in May 2025, across 572,600 jobs, with employment projected to grow 3% from 2025 to 2035. Service pay runs below that median and commercial producers above it, so use your own figures. Near $30 an hour, loaded cost lands near $42.
Take a $10,000 hands-on session for 30 people. At $42 an hour that is roughly 240 recovered hours to break even: eight per person, or about forty minutes a week across a quarter. Renewal prep on a mid-size commercial account clears that alone if the account manager keeps the workflow.
Then subtract honestly. Parallel running costs hours by design. A program that returns 300 hours and spends 60 proving the output is trustworthy still pays, and it pays with an answer you can give an E&O underwriter.
Four numbers to report at ninety days
Keep the report to numbers the management system already produces.
- Turnaround on renewal prep, on named accounts. Median days from kickoff to a complete submission, two quarters before against twelve weeks after.
- Service response time and reopen rate together. Faster replies that trigger a second inquiry are not a gain, and reopen rate is the only check that catches it.
- Documentation completeness. Share of activities with notes in the system within one business day. This is the number an E&O underwriter cares about, and it usually moves first.
- Named workflows running unassisted, with a person’s name against each. Six workflows that six people run in week twelve without help is a capability. Thirty licenses is not.
Add one check by hand: pull five pieces of client-facing output and confirm a licensed person reviewed each before it sent. If that cannot be shown, the boundary is theater.
Three ways this goes wrong
The policy arrives after the pilot. Staff experiment for a quarter with personal accounts, a written rule lands, and nobody admits what they already pasted in. Publish the approved-tools and data-class lists before the first session.
Coverage language gets drafted and not checked. A model produces a clean, confident comparison of two forms, one clause is wrong, and the error survives review because the document reads well. Line-by-line verification against the actual forms is the control, and it has to be taught as a skill rather than stated as a rule.
Credentials get mistaken for capability. Three people hold a certificate and the service team works as it did in January. Completion is a receipt. A named workflow run by a named person on a real account is the evidence.
Frequently asked questions
Is AI training for insurance teams worth it for a 30-person agency?
Yes, when the session runs on your own files. At 30 people a $10,000 session breaks even at roughly 240 recovered hours, about eight per person over a quarter, and renewal prep plus service documentation usually gets you there. A generic prompting class on someone else’s examples does not pay.
How much does it cost to train an agency staff?
A hands-on team session runs about $2,000 to $15,000 in the market by group size and scope. Per-seat at 30 people: the Associate in Insurance AI about $37,350 in course fees, the Certified AI Insurance Credential about $29,910, Coursera Teams $11,970 a year. The credentials are priced for individuals.
Do carrier rules actually prohibit us from using AI?
Rarely outright, and almost never in a statute. The NAIC Model Bulletin is addressed to licensed insurers, and its third-party section pushes due diligence and audit rights down into carrier agreements. Your real limits sit in the carrier agreement, the vendor contract and the E&O policy.
Does our E&O policy cover a mistake the model made?
Assume nothing. ISO made three optional generative-AI exclusions available for commercial general liability effective January 2026, and E&O forms were written before any of this. Ask your carrier in writing how it treats a claim arising from model output, and keep the answer with your procedures.
Can we use AI for client-facing chat or quoting?
Treat it as amber. It turns on disclosure, consent, what the carrier agreement permits for data flow, and whether a licensed person stands behind the output. Synthetic voice calls are separate: the FCC’s February 2024 ruling places AI-generated voices under TCPA consent rules.
What do we have to tell clients?
Disclose by default. The NAIC bulletin asks carriers to give notice when AI systems are in use, and agencies that adopt the same habit are never caught out by it. A line in your service documentation settles the question in advance.
Should we buy a certificate or run a workshop?
Both, in that proportion. Send two or three people to a credential for governance depth, run a hands-on session for everyone who touches a file, and keep a cheap library underneath. Buying only the credential leaves the service team unchanged, which is where the hours are.
What proves the training worked rather than the tools?
A named workflow a named person runs without help, on a real account, twelve weeks later. “Renewal prep on a mid-size commercial account takes a day instead of three” survives an ownership meeting. “The team is excited about AI” has no defender at renewal.
Key takeaways
- The limit on an agency is contractual, not regulatory. The NAIC bulletin binds carriers; you inherit it through carrier agreements, vendor terms and your E&O policy.
- Draw the green, amber and red lists before you buy anything. The amber rows are the curriculum and the only part worth paying for.
- ISO’s optional generative-AI exclusions took effect in January 2026. Confirm in writing how your own coverage responds before staff draft anything client-facing.
- Train service roles first and ownership last. One session for the whole agency teaches nobody.
- Report renewal turnaround, service response paired with reopen rate, documentation completeness, and named workflows running unassisted. Not license counts.
If you want this run against your own book, carrier panel and management system, book a free AI readiness call. We will tell you which roles return hours fastest, which work we would leave manual, and whether your constraint is fluency or review.
Arthur Wandzel
