Aipedia: IT consultants and developers

LLM integrations and client data leakage

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

Integrating a third-party large language model means client data leaves the client's systems, so a poorly designed integration can expose personal data, secrets or confidential material. Claims usually follow a breach, and the developer's design choices are scrutinised.

Part of: AI in IT consultants and developers

In short

When you connect a client system to a third-party LLM API, prompts may carry personal data, credentials or commercially sensitive information to the model provider. That raises UK GDPR questions about controller and processor roles, international transfers and security, as well as confidentiality duties under BCS clause 3.4 and contract terms such as PPN 017 limits on training with buyer data. If data leaks, the client may face regulatory and customer claims, then pursue the integrator. Technology PI usually responds to negligent design, subject to the wording, with cyber cover often relevant too.

What LLM integration work involves

Typical projects add summarisation to a case management system, generate draft replies in a CRM, classify incoming documents, or let staff query internal data in plain English. The integrator chooses the model provider and API settings, decides what data goes into each prompt, and builds logging, retrieval and error handling. Those choices determine how much client data leaves the client's environment, where it is processed and how long it is retained.

The duties that apply

The ICO's AI guidance hub covers AI and data protection, explaining decisions made with AI and an AI and data protection risk toolkit. In general terms, UK GDPR requires the parties to identify who is controller and who is processor, put processor terms in place, apply appropriate security, and handle any transfer of personal data outside the UK lawfully. The integrator may be a processor for the client and may also be arranging a sub-processor relationship with the model provider.

For BCS members, clause 3.4 prohibits disclosing confidential information and clause 1.1 requires due regard for privacy and security. Public sector contracts using PPN 017 may bar the use of buyer data to train AI without written approval, so provider settings on training and retention need checking.

What can go wrong

Illustrative scenario: a developer integrates an LLM into a recruitment firm's applicant tracking system to summarise CVs. To save time, the integration sends whole records, including health disclosures and right-to-work documents, in every prompt, and debug logs containing prompts and an API key are written to a shared storage location. A misconfiguration exposes the logs. The client must assess and possibly report the breach, handle candidate complaints and review its processor arrangements, then claims its losses from the developer.

Quieter failures include using a provider tier that retains or trains on data contrary to the client contract, or sending data to a region the client had not approved.

How PI and cyber usually respond

A client claim that your integration design was negligent, or that you breached confidentiality, is commonly within technology PI, subject to the wording and any data or cyber exclusions. Your own breach response costs, and some third-party data liability, are usually the business of cyber insurance. Regulatory fines are a separate question and may not be insurable. Read PI versus cyber for IT consultants and Article 82 compensation and processors for more on where claims fall.

Practical controls

Map what data each prompt contains and minimise it; strip identifiers where you can.
Never place credentials or secrets in prompts or logs; manage keys in a vault.
Check the provider's retention, training and location terms against the client contract.
Agree controller and processor roles and sub-processor approval in writing.
Support the client's data protection impact assessment before go-live.
Log carefully and restrict who can read logs. See building AI chatbots and training models on client data.

Questions to settle with the model provider

Before recommending or wiring in a provider, the integrator should be able to answer a few basic questions for the client, and keep the answers on file:

Does the provider retain prompts and outputs, and for how long?
Is customer data used to train or improve models, and can that be switched off contractually?
Where is data processed and stored, and which sub-processors are involved?
What security commitments and incident notification terms does the provider offer?
Can the client delete data and audit compliance?

These map onto the supply chain expectations in principle 7 of the AI Cyber Security Code of Practice. If the answers change after launch, for example because the provider updates its terms, the client should be told. Agreeing who monitors provider changes avoids a later argument over whose job it was.

How Apex helps

Apex arranges technology PI and cyber cover for IT consultants. A broker reviews how your policies treat data leaks through AI integrations you build. See the IT consultants AI hub for related guides.

Sources

Frequently asked

Is sending client data to an LLM API a data breach?

Not automatically. It can be lawful if roles, processor terms, security and any international transfer are properly handled and the client has approved it. It becomes a problem when data is sent without a lawful basis, beyond what was agreed, or is then exposed.

Who is the processor in an LLM integration?

It depends on the arrangement. The client is usually the controller. The integrator may act as its processor, and the model provider may be a sub-processor or contract directly with the client. Agree and document the roles before launch.

Does PI cover a data leak from an integration I built?

A client claim alleging negligent design or breach of confidence is commonly a PI matter, subject to the wording and any cyber or data exclusions. Breach response costs are more often cyber. Check both policies and how they interact.

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.