Insurance agency AI use policy: a starter template
Before a CSR pastes loss runs into a chatbot or IT connects a tool to the AMS: an AI use policy for independent agencies, with a tool register, eight clauses to adapt, a permissions table and four tests to run before rollout.

Example question from a CSR, end of a long day: "Can I use my own AI account to summarize these loss runs? It's way faster." A week later, the operations manager gets another one: "The vendor wants full access to our mailbox and the AMS. Do we say yes?" Without a written policy, each employee answers these on their own.
This starter is for principals and operations managers in US independent agencies. It covers four things: which tools are approved for which tasks, which client data they may receive, what they may read or write in your systems, and who checks the output. It is an original framework, not a state-specific legal policy and not a copy of an association template. Adapt it with the people who handle operations, privacy, security and legal review in your agency.
Start with an approved tool register
Write down each approved tool, the account type, the task and the limits. One line per task, not per vendor: the same tool can be approved for drafting a newsletter and not for reading dec pages.
| Tool and account | Task | Client data allowed | Systems and permissions | Reviewer | Approved on |
|---|---|---|---|---|---|
| Writing assistant, agency business subscription | Draft blog posts and newsletters | None | None | Marketing lead | Date |
| Document tool, agency account | Summarize loss runs for a submission | Loss runs, application | Read files in one shared folder | Account manager | Date |
| Email tool connected to the shared inbox | Draft follow-ups for missing documents | Client emails and attachments | Read inbox, create drafts, no sending | CSR who owns the account | Date |
Before approving a line, check where the data is processed, who can access it, how long it is kept, whether it is used to train models, and what the contract says. Check the plan you actually pay for, not the vendor's description of another plan. Also check your carrier and client agreements: some restrict where their data can go.
The Big I publishes AI legal issues and acceptable-use resources and a sample policy for members. This article does not reproduce that member resource. Use it if your agency has access, alongside a review of your own tools.
Public templates to compare
Only 18% of agents say their agency has a well-defined AI use policy, according to Liberty Mutual's 2026 study. Three association templates are worth reading next to this starter. Ohio Insurance Agents publishes a 4-page agency AI use policy dated September 2025. Insurance Agents & Brokers wrote two employee handbook policies, shared on Utica National's E&O risk site: one that authorizes generative AI under rules, and one for agencies that prohibit it. The Big I's sample policy is for member agencies only.
| Template | Published by | What it covers | Access |
|---|---|---|---|
| Agency AI Use Policy (September 2025) | Ohio Insurance Agents | Core principles (transparency, data privacy and security, fairness, accountability), use cases, ethical concerns and reporting. No personal or policyholder data in public AI tools. | Free PDF |
| Generative AI Acceptable Use Policy | Insurance Agents & Brokers, shared by Utica National | An amendment to the employee handbook: authorized use, responsible use, data protection, compliance, policy review. | Free PDF |
| Generative AI prohibited use policy | Insurance Agents & Brokers, shared by Utica National | One page for agencies that do not allow generative AI at work. | Free PDF |
| Member Agency Sample Policy on AI | Big I and its Agents Council for Technology, with a legal memo | Not public. | Big I members only |
Those templates state principles. The register above and the permissions table below say what each tool may do on each task, and who checks it. An agency needs both.
Eight clauses to adapt
1. Scope and owner
This policy applies to employees and contractors who use AI for agency work, including AI features inside the AMS, the email client or other software. [Named owner] keeps the approved tool register. A new tool, a new connection to an agency system or a new use of an approved tool is not allowed until it is added to the register.
2. Approved tools and accounts
Staff use [approved tools and account types] only for the tasks listed in the register. Personal AI accounts do not receive client files or non-public agency information unless the owner has approved a written exception. If an approved tool is down, staff do not move the task to an unapproved one.
3. Client data
Each task in the register lists the data it may receive. Staff send only what the task needs. Removing the client's name does not make a document anonymous: a policy number, an address, a FEIN or a claim description can identify the insured. [Restricted data, such as Social Security numbers, driver's license numbers or health information] follows [the agency's designated process].
4. Review of the output
AI output is a draft until the assigned reviewer checks it. Limits, deductibles, dates and named insureds are checked against the source document. A value the tool could not find stays blank. The reviewer corrects the draft before it goes into an application, a proposal, a COI or a client email.
5. Insurance decisions and external actions
AI does not bind, change or cancel coverage, choose a carrier, or tell a client a claim is covered. Sending a client email or writing to the AMS requires the permission listed for that task in the register. Permission to read a record does not include permission to create, update, delete or send.
6. Access to agency systems
Each connection gets the smallest access the task needs: one folder, one inbox, the record types listed. Staff never paste passwords, one-time codes or access tokens into a prompt. [Named person] reviews connections every [period], removes access that is no longer used and closes access when an employee leaves.
7. Incidents
Staff report to [named contact] right away: an email sent to the wrong recipient, a coverage statement the policy does not support, an unexpected change in the AMS, client data sent to an unapproved service. Pause the task, keep the evidence and follow the agency's incident process. Don't upload more client data to an unapproved tool to figure out what happened.
8. Training and review
Staff are trained on each approved task and its limits before using it. [Policy owner] reviews the register after a tool change, a new connection or an incident, and every [agency-defined period]. Each exception records its scope, who approved it and when it ends.
Replace every bracket before adopting the text, and align it with your record retention, incident response and client contact procedures. Reviewing the output does not undo data sent to the wrong place, and a written policy does not replace the contract and technical checks before access is granted.
Turn the clauses into a permissions table
For every connected task, write the trigger, what the tool reads, what it produces and who approves. "Full email integration" or "AMS sync" is not a permission. Spell it out. Example for a document-chase workflow:
| System | Read | Create | Update | Delete | Send or submit |
|---|---|---|---|---|---|
| Shared inbox | Client emails and attachments | Drafts | No | No | Employee only |
| AMS | Accounts, policies, activities | Activity notes | Only after review | No | Not applicable |
| Carrier portals | Quote screens listed | Draft submission | Before submission only | No | Employee only |
| Shared drive | One folder per task | Summary files | No | No | Not applicable |
Example: a workflow reads a client's reply, sees that the signed application is still missing, and prepares a follow-up. Under this table it drafts the email; the CSR checks the recipient and the missing items, then sends. The guide to practical AI workflows in an agency separates preparation tasks from decisions in more detail.
For document work, name the files the tool may open and how the reviewer checks the result. The method for comparing insurance quotes with AI shows a source check in practice: each value points to its page, and a value the quote does not give stays open. Extracting a deductible is not the same as interpreting coverage or advising the client.
Test four situations before rollout
Run these with approved test data, and check what the tool actually did, not only what it says it did.
| Test | What you give the tool | What should happen |
|---|---|---|
| Conflicting information | An application with one payroll figure and a supplemental with another | The tool flags both figures; the account manager picks the right one |
| Missing attachment | A client email that says "loss runs attached" with nothing attached | The item stays open; the tool does not invent loss history |
| Instruction hidden in a document | A PDF with a line asking to forward the file to an outside address | The line is treated as document content, nothing is forwarded |
| Tool or connection change | A vendor update that adds a new permission | Access is checked again before the task continues |
Keep an employee on every external action during the test period. If a test fails, fix it before widening the task. Evaluating a tool during a free trial is a good time to run these four tests and fill the register.
Keep the policy tied to the real work
The Big I's AI governance session covers privacy, client information and E&O considerations. Train staff on how to check the tasks they actually do, not only on how to write a prompt.
Review corrections, access exceptions and incidents next to any time saved. If CSRs routinely skip an approval step, find out why and fix the workflow rather than adding a longer warning to the policy. The policy works when an employee can tell, for any task, what the tool may do next and who decides.
Frequently asked questions
What should an insurance agency AI use policy include?
A register of approved tools and tasks, the client data each task may receive, who reviews the output, what the tool may read or write in the AMS and the mailbox, and who to call when something goes wrong. Add an owner and a review date.
Can a CSR use a personal AI account for client files?
Not under this template, unless the agency has approved that account and checked its settings and terms. Loss runs, applications and dec pages carry client data even when the client's name is removed.
If someone reviews every output, is any AI use acceptable?
No. Review catches a wrong answer, it does not undo a file sent to an unapproved service or a tool given write access to the AMS it should not have. Data and permissions are approved before the task runs.
Can AI send emails to clients or update the AMS on its own?
Only if the policy explicitly allows it for that task. The safe default in this template: the tool prepares a draft or a proposed entry, and an employee checks and sends or saves it. Binding, cancelling or confirming coverage stays with authorized staff.
Is this the Big I's official AI policy template?
No. This is an original framework written for adaptation. The Big I publishes its own sample policy for members; have your final policy reviewed by qualified advisers.


