An AI voice system prompt insurance teams can defend is a short boundary, not a novel. The product stores it as optional advanced instructions for the agent's behavior. Use it to list allowed actions, banned phrases, and the tone you already chose. Allowed means greet, ask the saved questions, book, take a message, disqualify, or warm-transfer with the transcript. Banned means enroll, quote, bind, recommend a product, invent a benefit, or promise an outcome. Tone means the words stay inside Professional or Empathetic, whichever you set, and do not slide into a pitch. The greeting and the question cards remain the script. If the prompt and the greeting disagree, fix both before the number is live.
Boundaries the prompt does not replace
Put allowed actions in the AI voice system prompt insurance reviewers will read
Write the allowed list as verbs a supervisor can check against a transcript. The agent may say the greeting, ask the questions on the card, save answers, offer a transfer to the named queue, offer a booking, take a message, or end the call with the not-qualified sentence. It may repeat a fact a tool returned, such as the office hours you stored. It may not add a product recommendation on the way to that fact. Keep the list shorter than the question card. A prompt that restates every sales objection is a second script, and the two scripts will diverge the first time someone edits only one of them. Point the prompt at the card. Do not paste the card into the prompt and then edit them apart.
Name the handoff. When the rules say qualified, the prompt should tell the voice to speak the outcome message and transfer, not to keep the caller for 'one more question' the card does not contain. When the rules say not qualified, the prompt should end. When the caller asks for a premium, a plan name as advice, or an enrollment, the prompt should refuse and either transfer to a licensed person or end, according to the outcome you configured. Refusal is a sentence you write, not a vibe. Example shape: I cannot quote a price or recommend a plan. I can connect you with a licensed agent or book a time. Then do only that. Do not add a soft pitch after the refusal.
Match the list to the line of business without turning the prompt into legal advice. Medicare marketing still follows CMS Medicare communications and marketing guidance. A prompt for that number should say the voice will not describe plan benefits as a reason to switch and will not complete an enrollment. A life or final-expense prompt should say the voice will not bind coverage or interpret a health question as an approval. HHS HIPAA privacy rule is the privacy reference when the call collects health details: do not ask for more medical detail than the question card already requires, and do not read a full medication list back for no routing purpose. The prompt reminds the model. Your process still has to be lawful.
Banned phrases are a list, not a mood
Write the banned lines in the words producers are already forbidden to say. No 'you qualify for extra benefits.' No star rating used as a reason to switch. No 'we can get you a lower premium' unless a licensed person is on the call and your process allows that sentence from them, which means the AI still does not say it. No carrier name used as a recommendation. No confirmation that an application was submitted. No threat and no fake urgency about a deadline you have not put in the question card. Add the phrases your last compliance review actually flagged. A generic 'be helpful and compliant' line does not catch the sentence that got you in trouble last fall.
Banned also covers invention. The prompt should say: if you do not have the fact from a question, a tool result, or the greeting, do not guess. Office hours may be spoken if a tool returned them. A rate, a subsidy, a death benefit, and an approval may not be guessed. Silence or a transfer is better than a fluent error. Tell the model to use the fallback you configured when it cannot map the caller, instead of filling the gap with a sales sentence. Temperature is a separate setting and defaults to 0.7. Do not turn it up to make the voice 'more natural' and then hope the banned list holds. Natural is how wrong sentences get smooth.
Tone instructions that cannot outrun the script
One line of tone is enough. If the profile tone is Empathetic, the prompt can say to acknowledge the caller's situation in a short clause and return to the next question. If the tone is Professional, say to skip small talk. Do not write a paragraph of personality. Personality is how a prompt grows a second greeting that marketing never approved. Do not instruct the model to be persuasive, to overcome objections, or to 'get to the close.' Those instructions fight the allowed-action list. The close belongs to the licensed person after the transfer. The AI's tone is how it asks state and intent without sounding rude, not how it wins.
Seniors, shoppers, and Spanish callers do not get three personalities inside one prompt. They get the tone you set and, if the language differs, a different profile. You may say: use short sentences and offer to repeat. You may not say: this voice converts better with older callers. You have no basis for that claim in the prompt, and you should not ask the model to behave as if you do. Hostile remains a tone on the list. The prompt for a live campaign should say never to use sarcasm, threat, or pressure. Read the prompt aloud. If it sounds like a coaching memo for a closer, it is the wrong document for this profile.
How you test a prompt change
Change one thing, then call the number three times. First call, give clean answers and confirm you are transferred or booked and that nobody quoted a price. Second call, ask 'what plan should I pick' and confirm the refusal sentence and the handoff you wrote. Third call, stay silent and confirm the no-input path, not a new pitch. Save the three transcripts next to the prompt version. If you cannot tell which sentence came from the greeting and which came from the prompt, the prompt is too long. Cut it. The question card should still be the thing a new supervisor edits when a disclosure changes. The prompt should rarely change, and never in the same hour as a live AEP push unless counsel is in the edit.
After the test calls, set the length limits so a wandering prompt cannot talk for ten minutes. Silence, retries, and max length cap the call. Fallback when the caller is unclear is what happens when the prompt's refusal is not enough and the model still does not know what to do. A prompt is not a compliance program. 47 U.S.C. § 227 and your marketing rules still sit outside it. The $500 figure is a statutory damage amount, not a score the prompt can waive. Write the cap next to the prompt version so a later edit does not quietly raise it to the 600-second maximum.
How AgentTech keeps the prompt beside the script
AI voice agents stores the system prompt under advanced instructions, separate from the opening greeting and the question cards. AI lead qualification and AI receptionist still decide the outcomes: transfer, book, message, or end. The prompt does not replace those buttons. compliance scoring scores the recording so you can see whether a banned phrase showed up anyway. Fix the prompt and the greeting together when it does. The $50 seat unlocks the profile. Minutes bill from the wallet. Nothing in the prompt field submits an enrollment or pulls a rate.
Use AI mock calls if you want producers to practice the human half of the same boundary. The Training agent is not the place you paste this prompt and hope it teaches the model to sell. Live profiles get a short allowed list and a banned list. Training profiles get a prospect persona. Keep them apart. When counsel asks what the AI is allowed to say, hand them the prompt, the greeting, and the question card. If those three documents disagree, you are not ready to take the queue live.
Frequently Asked Questions
Does the system prompt replace the greeting and the questions?
No. The greeting is what the caller hears first. The questions are the fields you save. The prompt is the boundary: allowed actions, banned phrases, and tone. If they disagree, edit them until they match, then test with three calls.
Can the prompt tell the AI to recommend a plan if the caller asks?
No. Recommend, quote, bind, and enroll stay off the AI leg. The prompt should refuse and transfer or book, using the sentence you wrote. A licensed person handles the sales conversation.
Will a good prompt make outbound calls TCPA compliant?
No. A prompt does not create prior express written consent. Artificial-voice telemarketing generally needs that consent. See the FCC guide to robocalls and robotexts guide and 47 U.S.C. § 227. Damages are $500 per violation, up to $1,500 if willful.
Key Takeaways for Agency Operators
- Allowed verbs, then stop. Greet, ask, book, message, transfer, end.
- Banned lines are specific. No quotes, no plan advice, no invented benefits.
- Tone is one line. It does not get to be persuasive.
- Test three calls. Clean answers, a request for advice, and silence.
- The prompt is not consent. Outbound still needs its own legal basis.
Cap the call with silence, retries, and max length and set what happens when the caller is unclear.
