AI for Life Insurance: Where It Pays, and Where the Rules Bite

The short answer first. In a life insurance firm, AI earns its keep in six places. Rank them by value and you get one order. Rank them by risk and you get a new one.

Underwriting is worth the most. It also has the most rules. Agent practice is worth nearly as much to a firm with a field force. And almost no consumer law touches it. The prospect on a practice call is a machine. Lead response, policy service, sales call review and claims sit in between.

So start in the top-left corner of the map above. The value there is high and the rules are light. Write the AI program the states now ask for while you build. Then take on underwriting once that program can carry it.

How is AI used in life insurance?

Here are the six uses, ranked by what each tends to be worth. Each has a rule that shapes it.

The NAIC is the group of state insurance regulators. Its own survey is a good check. It heard from 161 life carriers. Of those, 58% said they use AI or machine learning. Or they plan to use it, or plan to explore it.

The uses it lists are all about who to cover, and at what price. Issue cover faster. Approve or decline. Put each person in a risk class. That is the price band they land in. That is where the money is. It is also just where the states are looking.

Underwriting: the biggest prize, under the heaviest rules

Underwriting is how a carrier decides who to cover, and at what price. In 2024, Munich Re asked 27 carriers how many cases they fast-track. That means the carrier issues cover with no human underwriter. The average was 11%. The same carriers expect 49% by 2030.

Most of that gap is paper. There are reports from the person's doctors. There are lab results and drug records. Then there are the questions an underwriter asks of each one.

The safe pattern is simple. AI reads, and rules and people decide. Triage means sorting each case by what it needs. The model pulls the values out of a long medical file. Each value is stored with its source. That means the file, the page and the sentence. Then it sorts the case. Clean. Needs more proof. Needs an underwriter.

Rules your staff can read set the price. A person owns each case the rules can't settle. A value with no page number is not proof. It does not matter how neat it looks.

Then come the rules. On 4 December 2023, the NAIC adopted a model bulletin on AI. A model bulletin is a template. Each state can adopt it as its own. This one asks each carrier for a written AI Systems program.

The program says who answers for each system. It sets risk controls. It covers the vendors whose tools the carrier uses. And it spans the whole firm: marketing, sales and distribution, pricing, underwriting and service.

On the NAIC's map dated 8 October 2026, 29 jurisdictions have adopted it. Most of them are states. Four more have rules of their own.

Two of the four spell it out. Colorado's SB21-169 makes carriers answer for testing their big data systems. They test for unfair discrimination. The test covers outside data on consumers. It covers algorithms and predictive models too. Its amended Regulation 10-1-1 covers life carriers.

New York's Circular Letter No. 7 (July 2024) goes further on vendors. A carrier "may not rely solely on a third-party's claim of non-discrimination." Put simply, the vendor's say-so is not enough. It must also say when it uses AI to decide cover or price. The people who apply have to be told.

Read side by side, they say one thing. The vendor's model is your model. Say a third party's score feeds your choice of who to cover. Then you need its tests and its data lineage. That is the record of where each piece of data came from. You also need the right to audit it. All three go in the contract, before go-live. For the wider view, see how to write an AI governance policy.

Agent practice: the use most lists skip

Most "AI for life insurance" pages are about carriers. The field force gets a line about lead scoring. Yet no part of financial services keeps its people worse.

LIMRA is the life insurance trade's research group. It counts agents and advisers as "financial professionals." In 2020 it found that only 15% were still with the firm that hired them. That was four years on. The figure is for agency systems that mostly hire people new to the work. Most of those who left went in their first two years.

LIMRA's 2025 session on sales teams is blunter. It says that, on average, no more than 20% have stayed. That has held over the years.

New agents quit for many reasons. One of them can be fixed. They meet their first fifty objections on real prospects. They handle them badly. No one had time to practice with them. A sales manager can run a handful of practice calls a week. A new agent needs dozens.

A voice practice coach closes that gap. The agent picks an objection to work on. "I already have coverage through work." "Send me something to read." "It's too expensive." Then they talk it through, out loud. An AI plays the prospect.

The coach holds the firm's own sales method. It is built from the firm's training calls. It checks each turn against a few rules from the playbook. These are called guardrails. Say the agent tells instead of asks. Or they plant an objection the prospect never raised. A flag shows on that sentence. A safer line sits next to it. When the call ends, the agent gets a debrief.

We built one of these for a life insurance sales firm. Agents rehearse objections out loud. The voice coach holds the firm's own method. Break a guardrail and the flag shows mid-call, on that sentence. The debrief is ready within 60 seconds of hanging up. It gives a summary, strengths, friction points and the next focus.

Voice latency is the pause before the AI speaks. Here it stays under 400 milliseconds. That is quick enough that a pause sounds like the prospect thinking.

We estimate a coach like this is worth $300K to $1.6M+ a year. That is for a sales firm with about 20 agents. Most of it is the coaching it takes over. Here is the math. Take 20 agents, with 4 coached sessions each a month. Each session costs $750. Over 12 months, that is $720,000.

The rest comes from new agents getting up to speed sooner. And more of them stay. That part is real, but harder to count. So count it last, on its own line. (Our AI ROI guide shows why the hard number has to carry the case alone.)

One split matters more than any feature. A practice coach talks to your own agents. A "live whisper" tool listens to real prospects on real calls. That second kind of tool has its own consent duties. It belongs further right on the map.

Lead response: the TCPA decides the design

Speed still wins leads. The Harvard Business Review audited 2,241 US firms. Some tried to reach a web lead within an hour. They were nearly seven times as likely to qualify it as firms that tried an hour later. The average reply took 42 hours.

So an AI that calls back in a minute is tempting. But the rules shape it before the code does. The TCPA is the US law on sales calls and texts. In February 2024 the FCC ruled on AI voices. Under the TCPA, it said, they are "artificial."

So these calls need the same consent as any robocall. For marketing, that is prior express written consent. It means a signed yes, in writing, given before the call. A design that works:

1. Check consent before the first dial, per channel. Store the proof with the lead.
2. Tell the person they are talking to an AI.
3. Hand off to a licensed agent before any advice on what to buy.
4. Honor a stop request on each channel at once.

Recording adds a second layer. About 11 states make each party agree before a call is recorded. For calls across state lines, assume the stricter law applies. California's wiretap law sets damages of $5,000 per violation. The sum is fixed in the law itself.

Javier v. Assurance IQ (9th Cir. 2022) was a case about an insurance quote site. The court held that consent given after the fact does not count.

Policy service, call review and claims

Policy service. AI can answer billing, address and status questions at any hour. First it checks who is asking. Then it shows only the fields the caller needs. A change to who gets the payout goes to a person. So does any request that moves money.

The cheap win is the book of policies you have now. Say a premium is missed. That can trigger a draft note for the agent who sold it. It beats a lapse notice weeks later.

Sales call review. Staff used to check only a few percent of calls. They listened for the required disclosures. AI can read all of them. It sends the gaps to a person.

Do you sell annuities? Then look at the NAIC's best interest model, now in 49 jurisdictions. It sets duties a person can check. They are care, disclosure, conflicts of interest and records.

FINRA is the body that polices brokers. Say your reps are registered and sell variable products. Then its Notice 24-09 is plain. Its rules on supervision and on messages to the public still apply. It does not matter if a person or a tool wrote the words.

Claims. Taking the claim and chasing the papers, yes. The decision, no. A death claim is the hardest call in the business. The person on the other end should never wonder if a machine decided it.

What AI costs to run in life insurance

Run cost is mostly two bills and one person's time. Our build library is the record we keep of each build. It puts the tools for a voice practice coach at $474 to $2,404 a month. Voice is the costly part. Live speech is billed by use.

Text systems cost less. Some of our live systems run 1,000+ AI interactions a day. Their all-in model costs tend to fall between $200 and $1,200 a month. Two habits help. Send easy work to smaller models. And reuse the parts of each request that repeat.

The time is the cost buyers forget. Someone owns the AI program. That means the list of systems and the testing. It means the drift checks the NAIC expects, and the vendor files. In a carrier, this is part of a compliance role. In an agency, it tends to be the ops lead. It takes a few hours a month. For the full math, see what it costs to run an AI system.

What goes wrong

• The vendor's score can't be explained. New York has said it. A vendor's word that its model is fair is not enough. If the vendor can't show its testing, the risk is yours.
• A proxy does the discriminating. A proxy is data that stands in for something else. Some data has an old bias in it. Some stands in for a protected class. Either can discriminate with no one choosing to. Test outcomes, not just inputs.
• A bot states a policy term wrong. In Moffatt v. Air Canada (2024), a tribunal held the airline liable for its chatbot. The bot had given a wrong answer about a fare. A carrier's chatbot is no different. Answer from the contract, and cite it. Send the rest to a person.
• Recordings without the right consent. In the strictest states, transcripts and call analytics count as recordings.
• The model drifts. Drift is a model slipping as the world it learned from changes. The NAIC defines drift for a reason. A model tuned on last year's applicants slides as the mix shifts.
• Practice built on someone else's script. A coach built on stock sales advice makes your agents sound like the rest. The value is in your own method.

Questions to ask an AI vendor in life insurance

1. Which of our rules does this system touch? Where in our AI program is that written down?
2. What backs each answer, or each value it pulls out? Can I see the source for any one?
3. How do you test for unfair discrimination? Will you share the results and give us audit rights?
4. Who consents to what, on which channel? Where is the proof stored?
5. What does the system do when it is unsure? Who gets that case?
6. Where do recordings and transcripts live? Who can read them, and when are they deleted?
7. Whose accounts do the model and hosting bills run on?
8. Is each model call traced? Can our compliance team read the trace without asking you?

Where AI doesn't fit yet

A small team. For a practice coach, our build library's line is about 10 agents. Below that, taped role play with a manager's notes is the better buy.

No written method. If your best manager coaches by instinct, a guardrail has nothing to enforce. Write the playbook first.

Underwriting without a program. Say no one owns the list of models, the testing and the vendor checks. Then a model that sets prices is a state exam waiting to happen. Build the program on a lower-risk use first.

AI that advises. Telling someone how term and whole life differ is information. Telling a named person what to buy is the licensed agent's job. It should stay there.