Atlantas.Ai A Kwick Blocks product

Blogs · July 2026

Seven questions to ask before you let AI speak to your policyholders

A buyer's checklist for Heads of Claims and conduct risk leads. If a supplier cannot answer these seven, the demo does not matter.

The short answer: the demo will be impressive. It is designed to be. These seven questions are the ones that decide whether the thing survives contact with a real claims operation and a Consumer Duty review, and most of them have nothing to do with how good the voice sounds.

1. Where does it get its answers, and who approves them?

The only acceptable answer is: from your documents, and you approve them. Your policy wordings, your claims process, your complaints procedure.

If the agent is answering from a general model's knowledge of insurance, it will eventually tell a policyholder something confident, plausible and wrong about their cover. You will meet that answer again at the Ombudsman, and "the AI said it" is not a defence.

Ask to see the knowledge base. Ask who can change it and how a change is signed off.

2. What is it forbidden from doing?

Ask for the list. A supplier who has thought about insurance will have one ready. It should include deciding or valuing a claim, judging liability, and handling a vulnerable customer alone.

A supplier who says "it can handle anything" is telling you they have not thought about conduct risk, and you would be buying their inexperience.

3. What happens when it does not know?

Two acceptable behaviours: say so, and get a person. Anything else is a guess with a confident voice.

Ask them to show you a call where the agent did not know the answer. If they cannot produce one, they have not run this anywhere real.

4. How does it spot a customer who needs a human, and how fast?

This is the question your conduct risk lead will ask, so ask it first.

Bereavement, distress, an uninhabitable home, someone who cannot follow the process, someone already complaining. What are the escalation triggers, who do they go to, and does the customer have to repeat themselves when they get there?

The FCA has found this is thin across the industry: only 48% of firms offer dedicated telephone support to customers with characteristics of vulnerability. Widening the front door with an AI agent while the escalation route behind it stays thin just moves the problem.

5. Can we read back every conversation?

Every call, every email, every chat, on the record and reviewable.

You need this for complaints, for the Ombudsman, and for your own quality assurance. It is also the only way you will ever find out what your customers are actually asking for.

Worth knowing what "normal" looks like here: the FCA found 13% of firms carry out no quality assurance at all on the support channels they already offer. Do not add a channel you cannot inspect.

6. What actually happens to policyholder data?

Where does it sit, who can see it, what is it used for, and is it used to train anybody else's model. Get the answer in writing and put it past your DPO before the demo turns into a procurement.

7. How do we measure whether it worked?

Agree the numbers before you start, and take the baseline before anything is switched on.

The useful ones are already on your dashboard: calls abandoned, time to first response, claims open with no customer contact this week, service-related complaints, uphold rate on those complaints, retention on policies that claimed.

If a supplier will not let you measure the before and the after, they are not confident about the after.

Two more worth asking

Does it work with the systems we already run? It should sit alongside your telephony, CRM and claims platform, not demand you replace them. A project that starts with a replatform will not go live this year.

What does it cost when volume doubles? Storms and renewal peaks are when you need it most. If the price doubles exactly when your call volume does, you have bought an outsourcer with extra steps.

Why this matters more in insurance than most places

Because of who is on the other end of the phone.

In most industries a bad automated experience costs you a sale. In insurance it lands on somebody whose house is flooded, whose car is written off, or whose holiday has collapsed, and it lands at the moment they find out whether the last six years of premiums were worth anything.

The regulator is watching that moment specifically. Reviewing home and travel claims handling, the FCA found that "claims service and delivery" was the main root cause of complaints, and criticised firms for managing expectations on timescales poorly and not keeping customers updated on their claim's progress. In 2024/25 the Financial Ombudsman Service took 45,606 new insurance complaints and upheld 38%.

An AI agent can genuinely help with all of that, because the work being failed is repetitive work that nobody has time for. But only if it is scoped like a regulated conversation, which is what it is.

Frequently asked questions

Is there a question you would add?

"Show me a real conversation where it went wrong." A supplier with production history has one. A supplier without one is showing you a demo.

How long before it is live?

Ask for the honest answer, and ask what it depends on. If the plan needs an IT project before anything happens, it is not going live this year.

Should we start with the whole book?

No. Start with the window where the gap is widest and the harm is clearest, which for most insurers is out of hours and first notice of loss.

Sources

See Atlantas.Ai answer for your own book.

Every call, email, chat and message, grounded in your own policy wordings, with a person on hand the moment a case needs one.