Standard Cyber Policies Don't Cover Your Chatbot: The AI Coverage Gap Nobody Read

A recruitment firm in the Midlands added a chatbot to its website last spring. Nothing dramatic — it answered questions about job listings, screened a few applicants, pointed people to the right form. Six months in, the bot told a candidate it had rejected her application "due to a gap in employment following maternity leave." It hadn't, technically. No human had made that decision. The model had strung words together that looked like a reason. But the screenshot went round, a solicitor got involved, and the firm faced a discrimination complaint it had never consciously made.

The owner assumed the cyber insurance would help. The claim came back declined. Not because of a technicality buried in the small print, but because the policy did exactly what it was written to do: it covered losses from attacks on the business, not losses the business caused itself through the software it chose to run.

That distinction is catching out more UK SMEs than anyone is talking about, and it's worth understanding before you switch on your next AI tool.

What cyber insurance actually covers

Most SME cyber policies were designed around a fairly clear picture of harm. Someone breaks in. Ransomware locks your files. A phishing email tricks an employee into wiring money to a fraudster. A supplier gets breached and your data spills out with theirs. The insurer pays for the response — forensics, legal advice, notification costs, business interruption, sometimes the ransom itself.

The common thread is an external cause. A third party did something to you. The whole model of cyber cover assumes a threat coming from outside your walls.

An AI tool making a bad decision doesn't fit that shape at all. When your chatbot invents a refund policy, gives a customer wrong medical or legal-adjacent advice, or leaks another customer's details because someone crafted a clever prompt, there's no intruder. The loss came from a system you deployed, working roughly as it was built to, producing an output nobody checked. Insurers see that as an operational choice, not a security event.

The exclusions are already there

Here's the part that surprises people. You don't need to go looking for a special "AI exclusion" clause. In most policies the gap already exists in the ordinary wording.

Professional advice and professional services exclusions are common in cyber cover, on the logic that giving advice is what professional indemnity insurance is for. A chatbot that answers customer questions is, functionally, giving advice. If it gets that advice wrong and a customer relies on it, your cyber policy points at your PI policy, and your PI policy may point back — because it never contemplated advice being dispensed by an algorithm at 2am with no human sign-off.

Bodily injury and property damage exclusions sit in nearly every cyber policy too. Fine, until a copilot tool in a maintenance business generates a work instruction that's plainly unsafe and someone follows it.

Then there's the newer wave. Several insurers have started adding explicit language about generative AI, machine learning outputs, and "autonomous systems." Some exclude losses arising from the use of these tools outright. Others require you to declare that you use them, and quietly void cover if you didn't. A client of ours found this line in a renewal document last year: cover was conditional on "appropriate human oversight of any automated decision-making system." Nobody in the business could have told you what that meant, whether they were compliant, or who was responsible for proving it if a claim landed.

Data leakage is the quiet one

The wrong-output problem is easy to picture. The data problem is sneakier and, frankly, more likely to bite.

When your staff paste content into an AI tool — a customer email, a contract, a spreadsheet of names — that data leaves your control in a way most people don't register. Some tools use inputs to train future models. Some retain prompts on servers outside the UK. Some copilots draw on your whole document store and can be coaxed into surfacing information a particular user shouldn't see.

If that leads to a personal data breach, you might expect the cyber policy to step in, because breaches are its home turf. But if the breach happened because you fed personal data into a third-party tool without a lawful basis or a proper data processing agreement, the insurer can argue you created the exposure through your own configuration. It's the difference between a burglar taking your data and you leaving it on a train. One is covered. The other is negligence.

Why this lands on the MSP's desk

The uncomfortable truth is that AI tools are being switched on by people who don't read insurance policies, and insurance policies are being written by people who don't watch AI tools being switched on. The two conversations never meet — until a claim, when it's far too late to change anything.

This is exactly the kind of gap a managed IT provider should be catching. Not because we sell insurance, we don't, but because we're the ones who see the tools going in. We know when someone connects a copilot to SharePoint, when a chatbot gets bolted onto the website, when a team starts using an AI note-taker in client calls. We're in a position to raise a hand before deployment and ask the questions nobody else is asking.

At Cloudworks we've started treating this as a proper piece of work rather than a passing comment. An AI-deployment pre-check, done before the tool goes live, covers a handful of things that matter:

• What data will this tool touch? Personal data, commercial data, regulated data — and where does it physically go once it leaves your systems?
• What decisions or outputs will reach customers or staff without a human checking them first? That's your liability exposure in plain terms.
• What does your current insurance actually say? We read the cyber and PI wording alongside you and flag where AI use sits outside cover, so you can go to your broker with specific questions rather than a vague worry.
• What controls close the gap? Human review on customer-facing outputs, prompt logging, data-loss prevention rules, a written acceptable-use policy for AI, and configuration that keeps sensitive data out of the model in the first place.
• What do you need to declare? If your policy requires you to disclose AI use or maintain oversight, we help you document it so a claim can't be waved away.

Make it a decision, not an accident

None of this is an argument against using AI. These tools save real time and, deployed with care, they're worth having. The point is that most businesses are adopting them by accident — one team here, one plugin there — and the insurance nobody re-read is built for a world that no longer matches how the business actually operates.

A pre-check turns that into a deliberate decision. You get to weigh the benefit against a clearly understood risk, put controls in place, and know where you stand if something goes wrong. It costs a fraction of a denied claim, and it's a good deal easier to arrange before the chatbot says something it shouldn't.

If you're running AI tools now, or planning to, the sensible first step is to find out what your policy really covers. We're happy to sit down and read it with you. It's a short conversation that has saved our clients from some very long ones.