Keeping AI Private: What UK Businesses Actually Need to Know

Using AI safely starts with understanding one simple thing: where your information goes when you use it.

When you type something into a tool like ChatGPT or Claude, your words do not stay on your laptop. They are sent to infrastructure operated by the provider, processed there, and the answer is sent back.

It happens so quickly that it feels local.

It isn't.

That does not make hosted AI unsafe, and it does not mean businesses should stop using it. In many cases, hosted tools are exactly the right option.

The important part is knowing when that journey matters.

Every AI prompt goes somewhere

For everyday work, the distinction can feel unimportant.

If you are asking AI to improve the wording of a marketing email, research a public topic or help structure an internal presentation, there may be very little sensitivity in what you are sending.

The situation changes when the same tool is asked to read client correspondence, financial records, contracts, employee information or material covered by a professional duty of confidence.

The technology may be exactly the same.

The data is not.

This is why asking whether a business should use ChatGPT, Claude or a private AI system is often the wrong question.

The better question is:

What information are we giving it for this particular job?

Open-weight, self-hosted and offline are not the same thing

These terms are often grouped together as though they all mean "private AI".

They do not.

Open-weight means the model itself has been published so that somebody else can download and run it.

Self-hosted means the model is running on infrastructure that you control.

Offline means the system is completely disconnected from outside networks.

That last option is relatively unusual in normal business use.

The important distinction is between the model and where it runs.

An open model accessed through somebody else's cloud service is still being run on somebody else's infrastructure. Your data is still being sent elsewhere.

Downloading and running that model on infrastructure you control is different.

That is where self-hosting can become useful.

Hosted AI is often the sensible choice

Running your own AI sounds attractive because it gives you more control.

It also gives you more responsibility.

With a hosted service, the provider manages the infrastructure, patches the systems and maintains the models. Business and API offerings from major providers also come with different data arrangements from ordinary consumer accounts, including commitments around how business data is handled.

For common tasks such as drafting, research, planning and summarisation of non-sensitive information, that can make hosted AI the practical choice.

The fact that a provider does not use your business data to train future models, however, is not the same as saying the information never leaves your organisation.

The provider is still processing it.

For everyday information, that may be perfectly acceptable.

For information governed by client confidentiality, contracts, professional rules or specific regulatory requirements, it may not be.

Privacy should be decided task by task

A business does not necessarily need to choose between "cloud AI" and "private AI" for everything.

In reality, many businesses will end up using both.

A team might use hosted AI for:

• drafting internal documents

• researching public information

• improving marketing copy

• planning projects

• summarising non-confidential material

The same organisation might use a more controlled environment for:

• confidential client correspondence

• sensitive financial information

• regulated records

• information that contractually cannot leave its systems

• work involving processors that require prior approval

The right setup depends on the work.

That is a much more useful distinction than trying to declare one AI platform "safe" and another "unsafe".

Running AI yourself does not automatically make it secure

There is another trap here.

Keeping AI inside your own infrastructure gives you control over where information is processed. It does not remove the need for security.

You still need to consider:

• who can access the system

• how information is encrypted

• where backups are stored

• how long data is retained

• how the server is patched

• whether the system is exposed to the internet

• who reviews anything that eventually reaches a customer or client

This became particularly visible in January 2026, when researchers identified around 175,000 servers running Ollama that were openly reachable across 130 countries.

Ollama itself is designed to listen locally by default. The exposure came from how those systems had been configured.

The lesson is simple.

In-house does not automatically mean secure.

Sometimes a professionally managed cloud environment is safer than a poorly maintained server sitting inside the business.

When does self-hosting make sense?

The extra complexity is easier to justify when there is a clear reason for it.

For example, when:

• client data is contractually required to stay within controlled systems

• a regulator, network or professional body restricts which processors can be used

• the same sensitive AI workload runs repeatedly at enough scale to justify dedicated infrastructure

It makes less sense when a business wants to run its own model simply because it sounds more private.

If the work is occasional, nobody can maintain the infrastructure properly, or model quality is more important than where it runs, a reputable hosted provider may still be the better option.

The useful question to ask before using AI

Businesses do not need to turn every AI interaction into a compliance exercise.

They do need a basic understanding of what they are sending and where it is going.

Before sensitive information goes into an AI system, ask:

What data is involved, where will it be processed, and are we allowed to send it there?

That one habit removes a lot of the uncertainty around AI privacy.

The goal is not to keep everything away from hosted AI.

It is to know which work can comfortably go there, which work needs tighter controls, and why.

At TUSTRA, we help businesses understand how AI fits into the work they already do, including where data is going, which tools are appropriate and where a more private setup genuinely makes sense.