Blog & Insights

You found an unapproved AI tool holding company data. What do you do first?

Find out what is in it before you do anything to it. Then stop more going in, secure the account, read the vendor's terms, talk to counsel, and decide its future.

Find out what is in it before you do anything to it. Then, in this order: stop more sensitive data going in, put the account under the company’s control, read what the vendor does with the data, take the facts to counsel, and decide with the person who uses the tool whether it stays. Do not start by shutting it off.

The order matters because each step depends on the one before it. You cannot judge the exposure until you know what went in, and you cannot decide the tool’s future until you know what it is doing for the people who use it.

1. Find out what you are dealing with

This comes first because every later decision rests on it, and the people using the tool hold most of the facts.

Talk to them before you touch anything. Make it clear that you are there to make the tool safe to use or to find a safe replacement. People who expect to be punished stop telling you things, and then you are working from guesses.

Write down the answers to these questions.

  • What is it used for? Which piece of work, how often, and what happens to the output.
  • What data has gone in? Customer records, contracts, employee information, financials, source code, credentials. Ask for examples.
  • Whose account is it? A personal account on a personal email address, or an account on a company address. A free plan or a paid one.
  • Who else uses it? One person, a team, or people outside the company.
  • What is it connected to? Uploaded files, a link to email or file storage, an integration with another system, an API key.
  • Does anything depend on it? A report that goes to a customer, a step in a process, a deadline.

2. Stop more sensitive data going in

Do this next because it is the one thing that is still getting worse. Whatever is already in the tool is in it. What you can change today is what goes in tomorrow.

Ask the users to stop putting sensitive data in while you sort it out, and name the kinds of data you mean. Let the rest of their work with the tool continue if it can. This step costs nothing, can be undone, and destroys no evidence.

3. Put the account under the company’s control

An account that belongs to one person’s email address leaves when that person does, and nobody else can see what is in it or close it.

  1. Make a company email address the owner of the account.
  2. Put a strong password for it in the company’s password manager and turn on multi-factor authentication.
  3. Review what the tool has been given access to. Remove connections to email, file storage or other systems that the work does not need.
  4. Replace any API key or password that was pasted into the tool. Treat it as known to someone else.

Do not delete anything yet. A record of what went in is what you and counsel will need to judge the exposure, and deletion is hard to reverse.

4. Read what the vendor does with the data

What happened to the data depends on the vendor and on the plan that was used, so read the terms for that plan. You need these answers before you can say how serious this is.

  • How long the vendor keeps what users put in, and what users get back out.
  • Whether the vendor may use that content to improve its models, and whether a setting turns that off.
  • Who else the vendor shares data with.
  • Where the data is stored.
  • How to export the data, and how to have it deleted.
  • Whether a business plan offers different terms from the plan that was used.

Save a copy of the terms as they stand today. Terms change, and the version in force when the data went in is the one that matters.

5. Take the facts to counsel

Whether anyone has to be told is a legal question, and it is not IT’s to answer alone. IT’s job is to bring counsel accurate facts quickly.

  • Contracts. Agreements with customers and partners often say where their data may be stored and who may process it. Putting their data into a tool they did not approve may be something you are required to tell them about.
  • Regulated data. Personal information, health information and payment card data come with legal duties, and in some circumstances those duties include notifying the people affected or a regulator. Which rules apply depends on the data, the people it describes and where they live.
  • Insurance. A cyber insurance policy may require notice within a set time.

Give counsel what you know: what data, how much, whose it is, when it went in, and what the vendor’s terms say. Confirm with counsel before you tell anyone outside the company, and before you conclude that nobody needs to be told.

6. Decide what happens to the tool

Decide this with the person in the business who owns the work, because they know what the tool does for them and what they would lose without it.

  • Approve it, with changes. Move to a business plan on company accounts, turn on single sign-on, set what data may go in, and add it to your inventory.
  • Replace it. If an approved tool does the same job, move the work there and help the team make the change.
  • Retire it. Export what the business needs, ask the vendor to delete the rest, close the account, and keep a record that you did.

Whichever you choose, tell the users what was decided and why. If they hear nothing, the next tool will not be reported.

7. Fix the reason it happened

Someone had a real need and no quick way to meet it. Ask what they tried first. If asking IT took longer than signing up, the same thing will happen with the next tool.

Look for others as well. One unapproved tool usually means there are more. What is shadow IT, and how is shadow AI different? covers how to find them. Then write down what is approved and what may go into it: the four-sentence AI policy is a short way to do that.

Common questions

Should we shut the tool off as soon as we find it? Not before you know what depends on it. Ask the people using it to stop putting sensitive data in, secure the account, and learn what the tool is doing for them. Then decide.

Should we delete the data straight away? No. Deleting it can remove the record of what was exposed, and counsel may need that record. Find out what the vendor keeps and how deletion works, then delete as part of a decision.

Is this a security incident? It may be. That depends on what data went in, what the vendor does with it, and what your contracts and the law require. Collect the facts and take them to counsel. Do not decide it alone.

What if the person who set it up is a senior leader? The steps are the same. Senior people are often the first to adopt a new tool because they have the authority to buy one. Treat it as a need the business has and work out how to meet it safely.

How do we stop it happening again? You will not stop people wanting better tools. Approve a small number of tools on company accounts, say plainly what may go into them, and make asking for a new one quick.

If the tool was connected to a system the business depends on, a Security Review of that system is the fixed-fee first step.

Related service: Security & Compliance

More like this, in your inbox. Subscribe for our latest thinking, resources, and events.

Get started

Want results like these?