Aipedia: IT consultants and developers

Building AI chatbots for clients: liability and PI

Aipedia, by the Apex broking team · Last reviewed 2026-10-27

When a chatbot you built gives a wrong answer or leaks data, your client usually bears the first claim, then looks to you if the build was negligent. Developers therefore need clear contracts on scope, testing and who controls content and data.

Part of: AI in IT consultants and developers

In short

Customer-facing chatbots built on large language models can hallucinate, be manipulated through prompt injection and expose personal data. The client deploying the chatbot is usually responsible to its customers for what it says. If the client then alleges the build was negligent, it may pursue the developer. The ICO's Snap "My AI" investigation shows that data protection risk must be assessed before a generative AI chatbot goes live. Technology PI usually responds to negligent build work subject to the wording, while data breach losses may sit partly with cyber cover.

How IT firms build chatbots for clients

A typical project connects a third-party large language model to the client's website or app, adds a system prompt, and gives the bot access to a knowledge base of policies, product details or account data. Some bots only answer questions; others can book, refund or change records. The developer often designs the prompts, retrieval and guardrails, while the client supplies the content and decides what the bot may do.

The duties that apply

The AI Cyber Security Code of Practice asks developers to design for security (principle 2), keep humans accountable (principle 4), document data, models and prompts (principle 8) and test before release (principle 9). Under principle 10, system operators should tell end users in an accessible way where and how their data will be used. If you also host or run the bot, you take on those operator duties.

The ICO concluded its investigation into Snap's "My AI" chatbot in May 2024 after issuing a preliminary enforcement notice for failing to assess the risks properly, and called it a "warning shot". For developers, that means a data protection impact assessment is part of the build, not an afterthought. BCS members must also have due regard for privacy and security (1.1).

What can go wrong

Illustrative scenario: a consultancy builds a support chatbot for an online retailer. The bot confidently tells customers that a return window is longer than it really is. Customers rely on it, and the retailer decides to honour the answers. A tester later shows that a crafted message makes the bot reveal other customers' order details from its retrieval index. The retailer claims its costs from the consultancy, alleging the knowledge base was poorly scoped and guardrails were not tested.

Hallucinations, prompt injection and over-broad data access are the three recurring failure modes. Each can also create a hallucination liability argument between client and developer about who should have caught the problem.

How PI usually responds

Technology PI is usually built to respond where a client alleges your professional services were negligent, so a poorly designed chatbot is the kind of claim it is meant for, subject to the policy terms. Loss caused by a data leak may involve both PI and cyber: third-party claims for negligent design commonly fall to PI, while incident response and notification costs are usually cyber territory. Some wordings may include AI-specific clauses, and contractual indemnities you give may go beyond what PI covers.

There is no compulsory PI for developers; contracts set the limit. Our page on AI chatbot insurance explains the cover options in more detail.

Allocating liability in the build contract

Define who owns content, the knowledge base and the decision on what the bot can do.
Agree testing scope, including adversarial and prompt-injection testing, and who signs off.
Make clear that the client is responsible for monitoring live answers unless you are paid to do it.
Cap liability at a level your PI limit supports and avoid uncapped indemnities for AI output.
Agree data processing terms and where prompts and logs are stored. See AI clauses in IT contracts.

Practical controls

Run a data protection impact assessment with the client before launch.
Restrict retrieval to the minimum data needed and enforce per-user access.
Document system prompts, model versions and test results.
Show users that they are talking to AI and how to reach a person.
Monitor answers after launch and agree a kill switch. For connected tools, read deploying AI agents with system access and LLM integrations and data leaks.

How Apex helps

Apex arranges technology PI and cyber cover for firms that build AI for clients. A broker reviews how your policies treat chatbot failures and the indemnities in your contracts. See also the IT consultants AI hub.

Sources

Frequently asked

Who is liable when a chatbot gives a wrong answer?

Usually the business that deploys the chatbot is answerable to its customers first. It may then claim against the developer if the build was negligent or breached the contract. How far that succeeds depends on the scope you agreed, testing and any liability caps.

What did the ICO say about chatbots in the Snap case?

The ICO issued a preliminary enforcement notice in October 2023 for failing to assess the risks of a generative AI chatbot properly. It closed the case in May 2024 after a compliant risk assessment and called it a "warning shot" to assess data protection risk before deployment.

Does PI or cyber cover a chatbot data leak?

Often both are relevant. A client claim alleging negligent design is commonly a PI matter, while breach response and notification costs usually fall under cyber. The answer depends on the wordings, so check both policies together.

Should a chatbot build include prompt-injection testing?

It is good practice for any bot that reads user input or retrieves data. Agree the testing scope in the contract, record the results, and make clear who is responsible for retesting after the client changes content, prompts or the underlying model.

Check your PI covers how you use AI

Tell us about your firm and the work you do. A broker will review how your current PI treats AI-assisted work, or arrange new cover. No fee for a quote. Or call 0117 325 0027.

Start a PI quote Request a callback

Apex Insurance Brokers Limited is authorised and regulated by the Financial Conduct Authority. Registered in England and Wales, company number 07014570. This page is general information about how AI is used in the profession and how insurance may respond. It is not legal, regulatory or insurance advice on your circumstances, and it does not guarantee that cover will be available or on what terms.