Atlantas.Ai A Kwick Blocks product

Blogs · July 2026

Is it safe to let AI answer policyholder calls?

The risk is not the technology. It is the scope. Here is what an AI agent must never be allowed to do in insurance, and who is accountable when it goes wrong.

The short answer: yes, if you are strict about scope. An AI agent is safe doing the work that is high volume, repetitive and documented: answering the phone, taking the first notice of loss, explaining what the policy says, keeping people updated. It is not safe deciding claims, judging liability, or handling a customer in crisis on its own. The firms that get this wrong are not the ones that used AI. They are the ones that let it do a job it was never supposed to have.

What is the actual risk?

Not that the agent sounds robotic. That the agent oversteps.

Insurance is a regulated conversation with someone who is often having a bad day. Under the Consumer Duty, you have to avoid foreseeable harm, and the FCA has already been specific about where the harm shows up in insurance: claims-handling processes that cause unreasonable delays, and firms that cannot show they are able to identify vulnerable customers who need extra support.

So the question is not "can the AI talk". It is: when a distressed customer rings, does the system know that this is not a routine call, and does it get a person on it fast?

The FCA's review of consumer support found the oversight is not always there to answer that. 13% of firms said they carry out no quality assurance at all on the support channels they offer. If you are adding a channel that talks to customers, that number should worry you more than the technology does.

What Atlantas.Ai will not do

The safe version of this is a short list of hard limits, and they are worth stating plainly because a supplier who will not state them is a supplier to walk away from.

Atlantas.Ai will not:

  • Decide a claim. It does not accept, reject, or value anything. It takes the notification and it tells the customer what the process is. Handlers and loss adjusters decide.
  • Judge liability or fault.
  • Invent cover. It answers from your policy wordings. If the wording does not say, it says it does not know and gets a person, rather than producing a confident, plausible, wrong answer that you will meet again at the Ombudsman.
  • Handle a vulnerable customer alone. A bereavement, a fire, someone in distress, someone who cannot follow the process. It hands to a human.
  • Argue with a complaint. The moment a customer complains, it routes into your complaints procedure. It does not defend the firm.
  • Pretend to be a person. It says what it is.

What it does instead

The work that is genuinely repetitive and genuinely documented, which is most of the volume.

  • Answers the phone at 2am and takes the first notice of loss properly.
  • Answers "what does my policy cover", from your own wording, and shows the source.
  • Tells people where their claim is, which is the update nobody has time to send.
  • Chases the document you are waiting for.
  • Answers the same routine questions across voice, email, chat and social from one knowledge base, so the answer does not change with the channel. That consistency is itself a Consumer Duty point: an answer that changes depending on which channel the customer used is the definition of a communication that is not clear.

Handing over to a person is the feature, not the failure

This is the part suppliers undersell, and it is the part your conduct risk lead will ask about first.

A handover is not the system breaking. It is the system working. The measure of a good AI agent in insurance is not how many calls it handles alone, it is how reliably it spots the call it should not be handling and gets it to a human with the context attached, so the customer does not have to start again.

The FCA data suggests the industry has room here: only 48% of firms offer dedicated telephone support to customers with characteristics of vulnerability, and 66% have specialist vulnerability training. If an AI agent widens the front door but the escalation route behind it is thin, you have moved the problem, not fixed it.

Who is accountable?

You are. This is not a technology answer, it is a governance one.

The agent works from your documented process, and everything it says is on the record and reviewable. You approve the knowledge base. You set the escalation rules. You can read back every conversation. If it says something wrong, that is a fault in your source material or your scope, and both are things you control and can fix.

That is why the scope list above matters more than any demo. A supplier who will happily let their agent quote coverage it cannot support is handing you a Consumer Duty problem with a friendly voice on it.

Frequently asked questions

Can it give a policyholder incorrect coverage information?

It answers from your wordings and shows its source. Where the wording is silent or ambiguous, it escalates instead of guessing. That behaviour is the point.

What about a customer who is clearly distressed?

It hands to a person. Distress, bereavement, emergency and vulnerability are escalation triggers, not conversations to optimise.

Can we see what it said?

Every conversation, on the record, reviewable. If a supplier cannot show you that, do not buy from them.

Does it work in the middle of a storm surge?

Yes. It does not queue. That is the point of it.

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.