Do you have to navigate data residency?

The job: 21 client PDF reports, five pages each, and page four of every one had to come out before they went to clients. The contract: this data is processed in the UK and nowhere else.

The reflex is to drop the folder into a chatbot. The moment you press enter, the prompt and the files have left the UK. That is true for ChatGPT, for Claude and for Gemini, and it is true before any policy, warning or guardrail gets a look in.

Fiona MacLeod runs a GP practice in the Highlands. She wants to know where the patient data goes before anyone mentions saving an hour. This one is for her.

Where your prompt goes.

Anthropic's privacy centre says Claude data is stored in the US. There is no UK option on a consumer or Team plan. OpenAI will store Enterprise data in the UK, but storage is not processing: the GPUs that read your prompt run in-region only for Europe, the US and the UAE. Not the UK.

So if your contract says UK-only, every chat route breaks it. Not after a review. On enter.

Build a tool instead.

The model is good at working out how to do a job. It does not need the data to do that. So ask it for a tool, not an answer.

I opened Claude Code, the terminal version of Claude, and asked for a script to remove a named page from a PDF. It wrote one in Python, tested it on a fake five-page file it made itself, and handed it back. I ran the script on my laptop across the folder. All 21 reports came out at four pages. The model wrote the tool. The tool read the data, where the data lives.

You do not need to be technical to do this. I am a business analyst, not an engineer. Claude Code, OpenAI's Codex and Google's Gemini CLI all work the same way: you type what you want in plain English and the tool writes the program, runs it, fixes it when it breaks and explains what it did. You will need to install a couple of things first. The tool tells you what, and walks you through it. If you can follow a recipe, you can follow this.

That is as far as the recipe goes here. Describe the job and the shape of the files, never the contents. Have it build and test on a fake. Run the finished tool yourself, where the data already lives. If you want the exact prompts, the guards that stop it overwriting anything, and the two places the pattern still leaks, drop me a DM on LinkedIn.

Send the model the job. Keep the data. Case Study 008

What this means for Fiona.

Fiona, you have a patient recall list to tidy up: duplicates, old addresses, a column in the wrong format. The quickest-looking route is to attach the spreadsheet to a chatbot and ask it to sort the list out. Here is what happens the moment you press enter. The list is uploaded to a server in the US. The model reads every row to do the job. A copy sits in your chat history on that server until it is deleted. Your patients' names, dates of birth and health details have left the practice, left Scotland and left the UK, in one click.

That is health data, the most protected kind under UK GDPR. Sending it abroad is a restricted transfer and needs a lawful route you can show. If it goes wrong it is a breach, reportable to the ICO within 72 hours, and your health board will want to know why a patient list was on an American server at all. You already knew all that, which is why you said no to the AI note-takers.

So do not send the list. Send the job. Open the tool, describe the list in plain English (which columns it has, what needs fixing) and give it one made-up patient so it can test. It writes a small program. You run the program on the practice computer. The real patients never leave the building, because they were never sent. There is no transfer to justify, nothing in a chat history abroad, and nothing to explain to anyone. The model only ever saw the column names and a patient who does not exist.

Two limits. The program can tidy, sort, de-duplicate and reformat. It cannot summarise or draft from what the rows say; that needs the model to read them, and you are back where you started. And a program the tool wrote is still software, so your health board's rules on what runs on practice machines apply.