Can your AI supplier read your client files?

Probably, yes.
Most AI suppliers’ trust pages say “only authorised personnel” can access your data, and stop there.
For a firm holding client files, the real questions are who authorises that access, for how long, and whether anyone would ever tell you it happened.

Three kinds of supplier access

When a supplier says its staff can see your data, it usually means one of three things.
Each needs its own answer.

• Support access. Someone helping you with a problem opens your account to see what you see.
• Engineering access. Developers and operators reach the database and logs to keep the service running.
• Abuse review. Some AI providers keep what you send for a period and let staff read anything their systems flag.
  Google’s paid Gemini API, for example, keeps prompts for 55 days for this.

An AI tool usually sits on top of one or more AI providers.
You can face all three at the supplier, and the third again at each provider underneath.

What goes wrong when nobody checks

In 2023 the US Federal Trade Commission settled with Ring, the Amazon doorbell company, for $5.8m.
It alleged that for years every employee and many contractors could view customers’ videos, and one employee watched thousands of them.
Ring disputed the allegations.

The FTC’s complaint made a second point that matters more to a buyer.
Ring had no monitoring of who viewed what, so it could not say how far the problem went.
Without a log, a supplier cannot answer your question even if it wants to.

In 2025 Coinbase disclosed that criminals had bribed support agents at an outsourced contractor to copy customer records, including ID documents, for 69,461 people.
The agents had the access their jobs required.
The weakness was how much of it they had, all the time.

What the law already asks

UK GDPR requires your processing agreement to bind the supplier’s staff to confidentiality, and to give you the right to audit.
Regulated firms have more.
The SRA expects solicitors to understand a provider’s safeguards before client information goes into AI, and the FCA holds a firm responsible for anything it outsources.
Neither will accept “the supplier said it was fine”.

Four questions to ask

• Is staff access to my records off by default? The answer you want is yes, with a named way to switch it on.
• Who approves it, and does it need a reason? Microsoft’s Customer Lockbox makes an engineer wait for the customer’s approval.
  That is a high bar, and a useful one to compare against.
• Does it expire? Access granted for one problem should end when the problem does, not stay open.
• Is each access logged where I can find it? A log only the supplier can read protects the supplier.

Then ask the same four about the AI providers underneath, and get the list of every sub-processor, including the analytics and support tools.
In November 2025 OpenAI told API users that an attack on its analytics provider had exposed their names and email addresses.

How we answer them

When our staff open a customer’s live account to help, they see how it is set up: which tools, tables and settings it has.
They do not see the records inside.
The person opening the account cannot choose to see more at that point.

To read records, one of us has to give a written reason.
The system writes that reason to the customer’s own audit trail before access starts, and refuses if it cannot.
The access ends with the session, within an hour.
You can read that trail yourself, in your account’s Settings under Privacy.

It has two limits, and we prefer to name them before you find them.

• It hides records, not labels. If your team types a client’s name into a task title, that title shows up in the account’s structure.
• The database sits underneath it. Whoever runs the platform’s database can reach data below the software, and that access does not yet write to your audit trail.
  We use it to run and repair the service.
  Any supplier who says this access cannot exist at all is describing their software, not their database.

Our trust page, data processing agreement and sub-processor list are public.
If you want an outside view of the AI tools your firm already uses, our AI audit covers them.

This post explains questions to ask.
It is not legal advice, so for a decision about your own firm, check the ICO’s guidance and your regulator’s.