Where does your data go when your business uses AI?

The question to ask first

Every automation that uses AI sends something to a model: an email, a document, a spreadsheet. The question isn’t “is AI safe?”. It’s “where does this particular data go, and is that acceptable for this particular job?”.

The answer depends on the job. A model reading product descriptions is a different risk from one reading customers’ medical details. The UK regulator, the ICO, makes the same point: it expects businesses to show how they’ve handled the risks in their own context, even when the underlying technology is the same.

Option 1: commercial AI services

This means sending data to a provider such as OpenAI or Anthropic through their business products or API.

The main providers say they don’t train their models on business data by default:

OpenAI says it does not train its models on data from its API platform or business products by default, offers a Data Processing Addendum, and lets eligible API customers store data at rest in regions including the UK and Europe.
Anthropic says that by default it will not use inputs or outputs from its commercial products, including its API, to train its models.

These are the providers’ own statements, and terms change. Check the current terms for the specific product you use, and note the exceptions: for example, sending feedback can change how data is treated.

Good for: jobs where the data isn’t especially sensitive, or where the provider’s terms and a signed data processing agreement meet your obligations.

Option 2: models on servers you control

Open models can run on your own server or a server you rent in a UK data centre. The data never goes to an AI company.

Good for: sensitive data, or businesses that simply want to keep control.

Trade-offs: you, or whoever runs it for you, are responsible for the server, its security and its updates. Open models can be less capable than the best commercial ones, so whether one is good enough depends on the job. That’s something to test on real examples before committing.

Option 3: no AI at all

This one gets overlooked. If a job can be done with plain rules, such as reading a supplier’s fixed-format order email, it needs no AI model and sends nothing anywhere new.

It’s often cheaper and easier to check, too. Use AI where it earns its place, not by default.

How I decide

Whichever fits the job. In an audit, each recommended automation comes with the approach it needs and why, so you know before anything is built.

The ICO’s eight questions

In April 2023 the ICO published eight questions that developers and users of generative AI need to ask. They’re a useful checklist for any business using AI with personal data:

What is your lawful basis for processing personal data?
Are you a controller, joint controller or a processor?
Have you prepared a Data Protection Impact Assessment (DPIA)?
How will you ensure transparency?
How will you mitigate security risks?
How will you limit unnecessary processing?
How will you comply with individual rights requests?
Will you use generative AI to make solely automated decisions?

The last one is worth dwelling on. If a system makes decisions about people with legal or similarly significant effects, with no human involved, people get extra rights under Article 22 of UK GDPR. For most admin automation, like entering orders or building reports, that doesn’t apply. For anything that decides who gets a job, a loan or a service, it might.

Questions to ask anyone building AI into your business

Which AI service, if any, will my data be sent to?
Where is it stored, and for how long?
Is it used to train anyone’s models?
Is there a data processing agreement I can see?
Does this job need AI at all, or would rules do?
What personal data does it touch, and can it touch less?

If you can’t get straight answers, don’t send the data.

A note on what this isn’t

This is a practical overview, not legal advice. For anything involving sensitive personal data, check the ICO’s own guidance and take proper advice.