How brokers can use AI without risking client data

Brokers handle applications, claims details, policy documents, driver data, business information, and financial context. That information should not be pasted into random AI tools without a clear policy. Start by defining which data can be used, which data must be removed, and which tools are approved for agency work.

Controlled information has a home. The AMS or CRM is usually that home. A writing assistant, a document extractor, or a reporting workflow may receive a subset of fields for a specific job. It should not receive the whole file “just in case.” Least data is easier to explain to a client, a carrier, or your own errors-and-omissions conversation.

Personal devices and personal accounts are a common leak. If a producer forwards a complete application to a personal email so they can drop it into a consumer chatbot, the brokerage has lost the control story. Give people an approved path that is fast enough to use on a phone.

Write the control story in language a new CSR can follow on day two. Which tools are approved. Which fields may enter a drafting workflow. Which files stay in the AMS. Who to ask when a document does not fit the list. If that story only lives in the owner’s head, brokers will keep using whatever is open on the laptop during a rush.

AI is safest when it helps with summaries, drafts, reminders, routing, and checklists. It should not replace licensed advice, coverage recommendations, or final client communication review. A practical rule is simple: AI can prepare the work, but the agency team owns the decision.

Brokers feel pressure to answer quickly. That pressure is exactly when people ask a model to “tell me if this risk is okay.” That is judgment. Route the question to a licensed reviewer instead. The model can assemble the facts that are already in the file: vehicle list received, loss runs missing, garage ZIP on the form.

Keep generated language administrative until a person adds advice. “We received your documents and a producer will review” is workflow. “Your prior claims should not affect this” is a statement the brokerage has to stand behind. Do not let a template cross that line on its own.

Train producers to move questions, not answers, into the tool. “Summarize which documents are in the file and which are still missing” is a workflow prompt. “Tell me how to explain this exclusion” is a judgment prompt. The second belongs in a conversation with a licensed colleague, not in a model window, even when the model sounds confident.

Useful workflows should have clear access controls, stored records, and owner visibility. If nobody can explain where the data goes, who can see it, or what the workflow changes, it is not ready for production. This matters most when automations touch client records, policy files, email, CRM notes, or renewal activity.

Role-based access is not only an IT slogan. CSRs may need to see outstanding document lists. They may not need to export the whole book into a spreadsheet for a prompt. Producers may need drafts on their own accounts. They should not have a back door into every other producer’s files without a reason.

Auditability is how you reconstruct a send. Store the source record, the draft, the reviewer, and the disposition. You do not need a courtroom log for every internal summary. You do need to know who approved a client email that mentioned a missing MVR or a renewal document request.

Permissions should follow the same roles you already use in the AMS. A summer intern helping with filing should not export loss runs into a drafting tool. A producer should not need a shared “admin” login to approve a reminder. Shared super-user credentials make every audit trail a guess. Named logins are part of using AI without risking client data.

Redaction is a daily habit, not a special project. Internal summaries often need a name, a product, a date, and a missing-item list. They rarely need full government IDs, complete medical narratives, unredacted claim notes, or full account numbers. Teach the team to strip those before any tool sees the text.

Document AI should extract fields into a review screen, not dump the entire PDF into a general chat. Extraction with a human confirm step keeps more of the file in the system of record. It also reduces the chance that a prompt history retains pages the brokerage did not mean to store in a third-party log.

When in doubt, summarize instead of attaching. “Loss runs for three years are in the file; page count and carrier names need a human check” is enough for a routing note. Attaching the packet to an unapproved tool is how controlled data becomes uncontrolled.

Practice redaction on a sample file during onboarding, not after a scare. Show a CSR how to replace a driver’s license number with “ID on file,” and how to keep medical or claim narratives in the AMS. People copy the habit they saw, not the paragraph in a handbook. Brokers who skip the demonstration will see the full packet in a chat history later.

Maintain an approved-tool list with a named internal owner. For each tool, record the workflow, the data classes allowed, retention, and whether the vendor trains on customer content. If those answers are not in writing, the tool is still an experiment.

Consumer accounts are not a vendor strategy. A free personal login has different retention, support, and access realities than a contracted workspace the owner can turn off. Brokers should assume personal tools are out of bounds for client files unless the firm has explicitly approved a workspace with controls.

Revisit the list when a workflow changes. A tool that was fine for internal meeting notes may not be fine for ACORD packets. Approval is per use case, not a lifetime badge because the logo is familiar.

Ask vendors whether they train on customer content, where files live, how deletion works, and who at the brokerage can turn access off. If those answers are not in writing, keep the tool in a non-client sandbox. Convenience is not an approval. A broker still has to explain the stack to a client, a carrier, or their own E&O conversation.

Know how long drafts, prompts, uploads, and logs live in each tool. If the vendor cannot delete a file you uploaded for a one-time extraction, that is part of the risk. Align retention with how the brokerage already treats email and AMS attachments where you can.

Do not keep “just in case” copies of sensitive packets in extra drives, chat threads, or export folders created for an AI pilot. The pilot should write reviewed fields back to the AMS or CRM, then drop the extra copy. Extra copies are how retention policies fail in practice.

Give someone the job of offboarding. When a producer leaves, AI workspace access should leave with them. Shared passwords and forwarded magic links are a data-risk story waiting for an exit interview.

Align extra copies with how you already treat email attachments. If the AMS is the file of record, the drafting tool should not keep a second packet after the reviewed fields are written back. Set a reminder to purge pilot uploads. Retention failures are usually leftover folders, not a missing encryption checkbox.

Loss runs, MVRs, ACORD packets, medical statements, and financials are high-sensitivity by default. They are also the documents people most want to “just summarize.” Put those document types on an explicit allow or deny list per tool. A general writing assistant is usually the wrong place for them.

Certificates and evidence of insurance create a different trap: speed. Teams under pressure will ask a model to fill a certificate from a blurry photo. Extraction can help a reviewer. The reviewer still confirms named insured, limits, and holders before anything is issued according to the agency’s procedure.

Email threads that mix family details, driver histories, and unpaid-balance talk should not be dumped wholesale into a prompt. Pull the operating facts you need. Leave the rest in the system of record.

Certificates look harmless because they are routine, which is why they get rushed into the wrong tool. Extraction can help a reviewer read a holder name. Issuance still follows the agency’s procedure. The same discipline applies to MVRs and loss runs: the model may flag missing years or unreadable pages. It should not characterize the driving or claims history in a client-facing sentence.