Every useful idea hits the same clause
Your team could save hours summarising case files, drafting reports from client data, or searching years of project records. Every time someone suggests AI for it, the same objection comes up: our client agreements say their data must not be shared with third parties, must stay in the UK, or must not be used with AI services without consent.
So the AI work stops at the ideas stage. Meanwhile competitors seem to be using AI, and some of your own staff may be quietly pasting client material into public tools anyway, which is exactly what the contracts are trying to prevent.
Why the answer is usually "it depends on the wording"
Contract clauses about data vary a great deal. Some forbid any sub-processor without consent. Some require data to stay within a region. Some prohibit use of data to train AI models, which is a narrower thing. Some were written before anyone was thinking about AI and are ambiguous. The first mistake is treating all of them as a flat ban.
The second mistake is assuming AI means sending data to a US consumer service. There are several ways to run capable models, and they differ in where data goes, who can see it, and what terms apply. Matching the architecture to the contract is a design task, not a yes or no.
What the stalemate costs
| If nothing changes | The effect |
|---|---|
| AI ideas shelved | Repetitive work on client files stays manual |
| Staff use AI anyway | The breach the contract prevents happens informally |
| Competitors move ahead | Clients start asking why you are slower |
| No documented position | You cannot answer a client who asks how you use AI |
The informal use is the real exposure. A ban with no approved alternative tends to produce exactly the behaviour it is meant to stop.
How we design AI that keeps client data where it belongs
- We read the relevant clauses with you, and your legal adviser where needed, and write down in plain terms what each type of client data is and is not allowed to do.
- We map the data flow for each proposed AI use: what leaves which system, where it is processed, whether anything is stored, and who could access it.
- We choose a hosting option to fit. That might be an open-weight model such as Llama or Mistral running in your own Azure or AWS account or on your own server, or a managed service such as Azure OpenAI or Amazon Bedrock in a chosen region under business terms, where the contract allows it.
- We keep retrieval, indexing and logs inside your environment, so the document store and the history of questions and answers never sit with a third party.
- We add redaction or pseudonymisation where it helps, so names and identifiers can be replaced before text is processed and restored afterwards.
- We produce a short written description of the setup that you can share with clients who ask, based on what was actually built.
Self-hosted models are generally less capable than the largest hosted ones, so we test the actual task on your data before committing. For many internal jobs, such as summarising, classifying and searching, a well-chosen smaller model is enough.
What you end up with
AI working on client material in a way you can explain in one paragraph: where the data goes, where it does not, and what model is used. Staff have an approved route, so the informal workarounds lose their reason to exist.
When a client sends an AI questionnaire, you can answer it from a real description of your setup rather than a hopeful one.
Where a contract is stricter than the rest, that client's data can be handled differently, for example processed only by the self-hosted model, while other work uses a more capable hosted one. The routing is set by the data, not left to each member of staff to remember.
Is this your situation?
- Client agreements restrict sharing data with third parties or outside a region.
- AI projects stop as soon as client data is mentioned.
- Clients have started asking how you use AI on their information.
- You suspect staff are using public AI tools on client material regardless.
- You need AI on private data without sending it outside your environment.