GLBA Safeguards for Insurance Agencies: An AI Tool Checklist

An insurance agency should add AI tools to its written information security programme the same way it adds any vendor: list each tool, classify the customer data it touches, run a vendor risk check, set access and logging rules, extend the incident plan, and then confirm the detail against your state's insurance data security law.

The wrinkle for insurance is who writes the rule. Most search results for GLBA safeguards point to the FTC's Safeguards Rule, but insurance agencies are usually answerable to their state insurance regulator under the same law, and many states have adopted rules modelled on the NAIC Model Law. So the FTC text is a strong benchmark, and your state's text is the standard you must meet. The checklist and register entry below fit either, with a note at each step on what to confirm locally.

Follow me on Instagram@sagnikteaches

Which rulebook applies, and why the FTC pages are only half the answer

The Gramm-Leach-Bliley Act requires financial institutions to protect customers' nonpublic personal information. Different regulators enforce it for different kinds of business. The FTC's Safeguards Rule (16 CFR Part 314) covers non-bank financial institutions within its jurisdiction. The Act's enforcement section gives "any person engaged in providing insurance" to the state insurance authority of the state where the business is domiciled, and leaves the FTC with institutions no other regulator covers. That is why many agency owners find the FTC guidance detailed but not quite their rule.

Connect on LinkedInSagnik Bhattacharya

The NAIC's Insurance Data Security Model Law (#668) is the template many states used. Its text covers an information security programme, a risk assessment, oversight of third-party service providers (due diligence when choosing them, and a requirement that they protect the information they hold), an incident response plan, and notice to the commissioner no later than 72 hours after determining that a cybersecurity event has occurred, where the event meets the model's tests, such as the state being the agency's home state or 250 or more of the state's consumers being involved. It also has a small-licensee carve-out: in the model, a licensee with fewer than ten employees, including independent contractors, is exempt from the programme section. The NAIC's cybersecurity page put the number of adopting states at 21 when checked in November 2026, and states vary on how they changed the text and what the exemptions are. Do not assume your agency is exempt on size without reading your state's version.

Subscribe on YouTube@sagnikteaches

Action for this week: find your state's insurance data security law or regulation, note its section on the written programme, third-party providers and notification, and keep the citation in the programme document. Everything below is built to fit either rulebook, because the controls overlap heavily.

The benchmark controls, and where AI tools touch each one

The FTC rule's list is a handy checklist because it is specific. Several of its items map directly to AI tool decisions.

One more fact from the FTC rule is worth knowing even if it is not your regulator: section 314.6 switches off four requirements for institutions that hold customer information on fewer than 5,000 consumers. They are the written risk assessment, continuous monitoring or annual penetration testing, the written incident response plan and the Qualified Individual's annual written report. The rest, including MFA, encryption and service-provider oversight, still apply. State rules use different thresholds, so treat these as a map to check, not a loophole to claim.

Step one: build the AI register

You cannot protect what you have not listed. Start with a plain spreadsheet that holds every AI feature in use, including those hidden inside tools you already pay for. Ask each team member, because the list will include things the owner has not heard of: the transcription feature in the phone system, the summary button in the CRM, a browser extension, a personal chatbot account used for emails.

• Tool and feature: name, plan and the account type (business or personal).
• Purpose: one line, such as "draft renewal emails".
• Data it sees: none, business data only, or customer nonpublic personal information, split by type: contact details, policy details, claims descriptions, health or financial data.
• Vendor terms checked: training use of your content, retention, deletion, where data is processed, sub-processors, breach notice terms.
• Security features: MFA, single sign-on, audit logs, admin controls.
• Owner and approver: the staff member responsible, and the date the Qualified Individual approved.
• Review date: at least annually, and whenever the vendor changes terms.

Worked example: a seven-person agency's register

Say a seven-person personal lines agency finds four AI features. A chatbot on a business plan for drafting emails; a call transcription add-on in the phone system; an AI summary feature in the agency management system; and a free chatbot account two staff use on their own phones. The register classifies them: the business chatbot is approved for general drafting with no customer identifiers; the call transcription holds customer voices and policy details, so it needs the full vendor check; the management system feature is covered by the existing vendor contract but needs the new terms read; the personal accounts are banned for any customer data and the staff are told, with the date recorded. Four rows, one afternoon, and the agency can now answer the first question any examiner or auditor asks: what tools do you use and what do they see?

Step two: the vendor check for each AI tool

For each tool that touches customer data, ask the vendor questions and keep the answers. The vendor's security and privacy pages, data processing terms and trust centre are the usual sources. If an answer is missing, ask in writing.

1. Is customer content used to train models by default, and can it be switched off at the account level? When checked in November 2026, OpenAI said ChatGPT Business and Enterprise data is not used for training by default, Anthropic said the same of its commercial products, and Google said Workspace does not train on customer data without permission. Personal and free accounts are a different matter, so verify the terms for the exact plan.
2. How long are prompts and outputs retained, and can you delete them or set retention?
3. Where is data processed and stored, and which sub-processors are used?
4. What encryption is used in transit and at rest?
5. Does the plan offer MFA or single sign-on, and audit logs of who used it?
6. What is the vendor's breach notification commitment, and to whom?
7. What happens to your data if you cancel?
8. Are there independent security reports, such as a SOC 2 report, available under non-disclosure?

Save the answers with the date. If a tool fails the check, you have three choices: restrict it to no customer data, switch plan or vendor, or retire it. Write the decision in the register. For costs of the whole exercise in a small brokerage, the year-one cost tutorial puts a figure on the time.

Step three: data rules staff can follow

A security plan that nobody can apply in the middle of a busy morning will not hold. Reduce it to three data classes and one rule for each:

• Green, no customer data: templates, marketing copy, generic questions. Any approved tool.
• Amber, limited customer data: first name, product type, renewal month. Approved business-plan tools only, and no identifiers such as policy numbers, dates of birth, addresses or account numbers.
• Red, sensitive data: claims details, health information, financial account data, driving or criminal history, identity documents. Never into a general-purpose AI tool; only into tools the Qualified Individual has approved for that class of data.

The submission and claims workflows are the usual pressure points. The AI submission intake tutorial and the claims enquiries tutorial describe those workflows, and their data would sit mostly in amber and red.

Before and after: a prompt rewritten for the data rule

Before: "Write a renewal reminder for [client's full name], policy [full policy number], [home address], whose home policy renews on 3 March and who had a water damage claim last year." After: "Write a renewal reminder for [first name], whose home policy renews in March. Do not mention any claims. I will add the details." The second version keeps the identifier and the claim out of the tool, and the agent merges the details afterwards in the agency system. It takes ten extra seconds and removes the plan's biggest exposure.

Step four: access, logging and accounts

• One business account per user, with MFA on. No shared logins, which hide who did what.
• Remove accounts when staff leave, on the same day. Put AI tools on the leaver checklist.
• Switch on admin controls: disable features you do not need, restrict file uploads where the plan allows and limit connectors or apps that link the tool to email or storage.
• Turn on audit logs if the plan has them, and note how long they are kept.
• Ban personal accounts for work, in writing, and give staff an approved alternative so the ban is realistic.

Note on connected apps: ChatGPT's connectors are now called connected apps (formerly connectors), managed from the Plugins section of Settings, and many assistants can link to mailboxes and drives. Each link widens what the tool can see. Approve links one at a time and record them in the register.

Step five: add AI scenarios to the incident response plan

Your written incident response plan should cover the usual events. Add three that are specific to AI.

1. A staff member pastes red data into an unapproved tool. Delete the content if the tool allows, screenshot the deletion, record what was entered, assess whether customer data was exposed and decide whether it is a reportable event under your state rule. Retrain the person; do not just blame them.
2. An AI vendor reports a breach. Establish what data of yours the vendor held, from the register. If customer information was involved, follow your normal notification steps and the 72-hour style deadline in your state's rule, where it applies.
3. An AI tool produces and sends something wrong to a customer. Not always a security event, but a record and a correction process. Cover it in your quality procedures.

Rehearse one of these in a short tabletop session each year. Ten minutes with a printed scenario shows where the plan says "notify the appropriate parties" without naming anyone.

A realistic mistake and how it showed up

Imagine an agency approves a business-plan chatbot and writes a policy. Months later a customer service rep, in a hurry, uploads a scanned claim form to a free summariser on her own phone, because the approved tool could not read the scan. Nobody notices until an audit sample asks for the list of tools that touched claims data, and the register does not match what the team does. The cause was not carelessness. The approved tool did not fit the job, and the policy offered no alternative. The fix: choose an approved tool that can handle the scan, or give a safe manual route, and ask staff in the quarterly check what is awkward about the rules.

The security plan entry itself: wording you can adapt

Most agencies already have a written programme, so the job is an addition, not a rewrite. Add a short section along these lines, then ask your counsel or compliance contact to adjust it to your state's rule.

Artificial intelligence tools. The agency maintains a register of every AI tool and AI feature in use. No tool may be used with customer nonpublic personal information until the Qualified Individual has approved it in writing, recorded the data classes it may receive, and completed a vendor risk check. Personal AI accounts must not be used for agency work. Approved tools require individual accounts with multi-factor authentication. The register and vendor checks are reviewed at least annually and whenever a vendor changes its terms or adds features. Incidents involving AI tools are handled under the incident response plan.

That is six sentences and covers the points an examiner will look for: a list, an approval, a vendor check, an access rule, a review cycle and a link to incident handling. Keep the policy short and put the detail in the register, where it can change without reissuing the plan.

A filled-in vendor check: a call transcription add-on

Say the seven-person agency wants to keep the transcription add-on in its phone system. The Qualified Individual works through the vendor questions and records the answers, labelled illustrative. Training on customer content: the vendor's terms say content is not used to train models, confirmed in the data processing addendum dated this month. Retention: transcripts kept 90 days by default, adjustable to 30. Location and sub-processors: listed on the vendor's trust page, two sub-processors. Encryption: stated in transit and at rest. MFA and logs: both available on the business plan, logs kept 12 months. Breach notice: the contract says the vendor notifies customers without undue delay, with no number of hours stated. Decision: approved for red data with retention set to 30 days, and a request sent to the vendor to commit to a specific notice period. The open request goes on the register with a review date.

The point is the record. If the vendor later changes its terms, you can show what you checked, when, and what you decided.

A second mistake: the AI feature nobody turned on

Another pattern: the agency management system adds an AI summary button in an update. Nobody chose it, but it is on by default and sends policy notes to the vendor's model provider. The register missed it because it was not a tool anyone had bought. The remedy is a standing question in each quarterly review: has any software we use added an AI feature? Check release notes and admin settings, then record and either approve or switch the feature off.

The checklist in one place

1. Find your state's insurance data security rule; cite it in the programme.
2. Name the Qualified Individual (or the equivalent role your state uses).
3. Build the AI register, including AI features inside existing tools.
4. Classify the data each tool sees: green, amber or red.
5. Complete and file the vendor check for every tool that touches customer data.
6. Enable MFA and individual accounts; ban personal accounts for work.
7. Write the data rules and train staff with examples.
8. Add the three AI scenarios to the incident response plan.
9. Add AI vendors to your third-party service provider list and contract review.
10. Include the register in the annual report to the owner or board.
11. Set review dates and put them on the compliance calendar.

For the calendar side, the compliance calendar tutorial shows how to track these alongside licences and certificates.

How to check the plan works

Every quarter, test three things. Pick a tool from the register and ask the person using it what data they put in; compare with the green, amber and red rules. Check that every account on the vendor's admin page belongs to a current employee. And ask whether any new AI feature appeared in your software since the last review, because vendors add them without announcement. If any answer is wrong, update the register and note the date.

Keep the evidence together: the register, the vendor answers, the approvals, the training record and the incident rehearsal notes. If your regulator asks, you should be able to hand over one folder. Read your state's rules on what the regulator may ask for, and ask counsel if you are unsure. If you would like help listing the AI tools in your agency and building the register entry, that is what my AI implementation consultation covers.

Questions agency owners ask about AI and data security rules

Does the FTC Safeguards Rule apply to my insurance agency?

Usually not directly. The Gramm-Leach-Bliley Act gives enforcement for anyone providing insurance to the state insurance authority, not the FTC. State laws and regulations, many modelled on the NAIC Insurance Data Security Model Law, set the detail. The FTC rule is still a useful benchmark. Confirm your state's rules with your regulator or counsel.

Is an AI chatbot a 'third-party service provider'?

If it processes, stores or has access to your customers' nonpublic personal information, treat it as one for risk assessment and oversight. That means due diligence before use, contract terms on security and data use, and a way to check it afterwards. Confirm how your state defines the term.

Can staff use free AI tools if they paste no names?

Remove obvious identifiers but remember that policy numbers, claims descriptions and combinations of details can still identify a person. Set a written rule on which tools are approved and which data types may enter them, and apply it to free tools too.

What if an AI vendor has a breach?

Treat it as a possible cybersecurity event for your agency. Your incident response plan should say who is told, how fast, and how you establish whether your customers' data was affected. Under the NAIC model, notices to the regulator have a 72-hour clock from determination, so check your state's version.

Further reads

• Can AI Answer Calls and Qualify Leads for an Insurance Agency? — Call and lead tools raise recording and consent questions too.
• Does Your Business Insurance Cover AI Mistakes? — Check whether your own policies respond to AI-related incidents.
• How Brokers Use AI to Compare Policy Wordings at Renewal — Policy wording comparison, a lower-risk use of AI.
• Where AI Saves Time in a Small Insurance Brokerage — Where the time savings are, to weigh against the controls.
• How to Automate Insurance Renewals With AI and Keep More Clients — Renewal automation and the customer data it handles.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — A readiness check for regulated professional firms.
• Best AI Tools for Insurance Agents: Seven for Quotes, Renewals and Calls — Seven AI tools for independent insurance agencies, with costs, test scripts, a six-person agency sum and the E&O checks before you buy.
• Zoho CRM for Insurance Agents: Policies, Renewals and Zia — A step-by-step Policies module and renewal rules in Zoho CRM, the edition limits that matter, and what Zia adds for an insurance agency.
• Hiring an AI Consultant for an Insurance Agency: Scope and E&O — A pre-hire question list with E&O and carrier checks, three first-project options sized against the saving, and a handover checklist.
• WISP for Tax Preparers: Add Your AI Tools Without Gaps — One added section for your existing WISP that covers AI tools: register, vendor due diligence, data rules, MFA and incident steps, with sample wording.
• Does Agency E&O Cover AI Mistakes? What to Ask Your Carrier — A per-tool list of questions for your E&O carrier, the exclusion wording to look for, and the written AI policy that gives you evidence. Not legal advice.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the FTC Safeguards Rule (16 CFR 314.4 and 314.6), the Gramm-Leach-Bliley Act's safeguards and enforcement sections (title 15, sections 6801 and 6805, as published by Cornell's Legal Information Institute), the NAIC Insurance Data Security Model Law (#668) text and the NAIC cybersecurity page, and the business data terms of OpenAI, Anthropic and Google, all checked November 2026. State requirements differ, so verify your own.