Skip to main content
DoneThat

AI Adoption GuideInsuranceService

AI voice agent for policy servicing

Voice AI handles inbound coverage questions, certificate requests, and payment tasks, routing complex calls with full context.

Insurance processQuoteUnderwriteBindIssueBillServiceRenewClaim

By Don, DoneThat’s AI coach · updated

What a spoken answer is allowed to contain

A voice agent on a policy servicing line may say only what the authenticated policy file supports, and it must name that support. A coverage sentence without a cite is not an answer. It is an uncited statement that happened to use a calm voice.

The allowed set is operational, not interpretive. The agent can confirm a policy is in force for a named insured on a stated effective date. It can read a limit, deductible, or form ID that appears on the declarations or schedule. It can say a certificate request matches a pre-approved template and has been queued for a person to review. It cannot decide whether a loss would be covered. It cannot invent a missing endorsement. It cannot treat a spoken "yes, you are additional insured" as a document a landlord can rely on.

Voice is the channel. Retrieval follows the same cite-or-stop pattern as the policy RAG coverage assistant: find the clause or schedule line, attach the cite, and stop when the file does not contain the fact the caller asked for.

Authenticate before you retrieve the file

Do not open the policy until the caller passes the authentication the carrier already uses for servicing. Typical checks are policy number plus a second factor the file already holds: date of birth or tax ID fragment for personal lines, named insured and a secret the account record stores for commercial. If the check fails, retrieval does not run. The agent says authentication did not complete and offers a live transfer. Empty stays empty.

Telephony and CRM stacks such as Genesys and Salesforce, and policy admin systems such as Guidewire and Duck Creek, already hold the identity and policy records this flow needs. Use them as systems of record. Do not keep a parallel coverage copy, and do not invent a lookup those systems did not return.

If the policy number authenticates to a different named insured than the caller claims, stop. Do not search for a similar account. A near-match is not a cite.

Payment questions that survive authentication still belong to the billing record. For a balance, due date, or payment-method call, apply the same rules as a billing inquiry voice agent rather than stretching a coverage prompt to guess a dollar amount.

Cite the policy or leave the field empty

After authentication, retrieve the object the question needs: declarations, form schedule, endorsement list, or billing snapshot. Every spoken fact carries a cite the agent can pass to the desktop of the person who may take the call next: document type, form number or schedule line, and as-of date.

If the retrieve returns nothing, the spoken field stays blank. The agent says it cannot confirm that item from the file, then offers a transfer with the transcript, the authentication result, and the empty field marked empty. It does not substitute a neighboring form. It does not paraphrase a brochure. It does not finish the sentence because the caller sounded urgent.

This rule is what blocks an invented coverage answer. Fluency is not evidence. When the schedule does not list CG 20 10, the agent does not say the landlord is an additional insured on most packages. The missing fact stays unspoken, and the call transfers.

Mid-term changes the caller believes already happened are not voice facts until the endorsement file shows them. If the caller insists a vehicle or location was added last week, retrieve the endorsement record. If it is not there, the field stays empty. The work belongs to a person, or to the intake described for a mid-term endorsement processing agent, not to a spoken guess.

Exceptions, certificates, and coverage opinions stay with a person

A service agent still owns three classes of work the voice path must not close.

Exceptions are anything that failed a rule, failed a retrieve, or arrived with incomplete authentication. The voice agent collects caller intent, auth result, cites, and blanks, then routes with that packet intact. It does not retry the same question with a looser prompt until it hears a yes.

A certificate of insurance is a document, not a conversation. Do not auto-issue a COI. If the request asks for wording the template library does not contain, names an additional interest the schedule does not show, or needs a signature a person must apply, the agent says the certificate was not issued and queues it for a licensed servicing person. A caller who hears "I can take care of that" will treat the call as an issued COI. That is a failure even when the tone was polite.

Whether a described incident, contract clause, or additional-insured demand is covered is a coverage opinion. The voice agent may read the form ID and the cited grant or exclusion. It may not apply those lines to a fact pattern. "Does this cover the forklift accident in the parking lot?" is a transfer with the question preserved, not a spoken yes or no.

Complaints that surface on the same inbound line (delay, denied certificate, alleged misquote) should leave servicing and enter the clocked path in complaint classification and deadline routing instead of sitting in a certificate queue.

Failure modes that still sound like a completed call

The dangerous calls are the ones that sound finished.

A spoken coverage line with no policy cite. The agent says the general liability limit is one million dollars because that is the product's usual limit. The declarations were never retrieved, or they were retrieved for a different term. The caller repeats the number to a landlord. No document ID sits on the wrap-up. Rule: no cite, no speak.

Treating the call as an issued COI. The agent confirms the named insured and says a certificate will go out, or that the caller is all set for the additional insured. No certificate record exists. No template matched. The caller leaves believing a certificate exists. The correct close is: request logged, not issued, a person will complete it or will call back if the request fails rules.

Inventing a coverage answer. Authentication timed out, or the form schedule was empty, and the model completed the thought: yes, landlords are additional insured on this product. That sentence is fluent and false for this file. Empty stays empty. Transfer with the blank marked blank.

Here is the pattern on a single inbound call. A facilities manager for a warehouse tenant calls two days before a landlord inspection. She reads a policy number, confirms the named insured, and passes the commercial authentication the account file already stores. She asks whether form CG 20 10 naming the landlord appears on the current GL schedule, and whether a certificate can be emailed to the property manager today.

The schedule cite shows CG 20 10 for that landlord on the current term. The agent speaks that fact and names the schedule and as-of date. The certificate request asks for a waiver of subrogation the template library does not allow and the schedule does not show. The agent does not issue the COI. It says the additional-insured form is on the file, with the cite, and that the certificate was not issued because the waiver request failed the template rules. It routes the exception to a service agent with the transcript, the schedule cite, and the failed certificate rule.

If authentication had failed, none of the schedule would have been read back. If the schedule retrieve had returned empty, the agent would not have said the landlord was additional insured. After hang-up, a supervisor should be able to point to the policy file for every sentence spoken, see blanks where auth or the file was missing, and confirm that the voice channel issued no COI and delivered no coverage opinion as a fact.

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first