Shadow AI in the legal department: one week of work

In one week the legal department can map shadow AI, stop the worst, offer an approved alternative and write the rule. Shadow AI means staff using AI tools on their own accounts at work, with no contract, no settings and nobody knowing what has gone in. Here is the order of work, and what a week cannot do.

Updated by Nils-Erik Jansson

Shadow AI ends up with the lawyer for two reasons. The first is that it is a contract and data protection question. A private account means that the employee has accepted the terms, the company has not, and the company has no data processing agreement with the vendor. Commitments that apply under commercial agreements, such as content not being used for training, do not apply to private accounts [1]. The guidance from the Swedish Bar Association (Sveriges advokatsamfund) expressly distinguishes between tools where the user accepts the terms as a private individual and tools the firm has contracted for in its business [2].

The second reason is that the legal department uses the tools itself. Whoever writes the rule for the company therefore starts with their own private accounts.

The week in a table, with the details under the headings below.

Day What Who
Day one Find out the position: which tools, how many users, for what The lawyer, with the department heads and IT's logs
Day two Stop the two or three risky uses until an approved tool is available The lawyer, in the teams where those uses are
Days three and four Offer an approved alternative: company contract, data processing agreement, settings and login through the company's identity system The lawyer, with the company as contracting party
Day five Write the rule on one page and announce it on the day the approved tool opens The legal department
The week after Two-hour training, a new look at the logs after a month The legal department

Day one: find out the position

Ask, without threatening. Three questions to the department heads are enough: which tools are used, by roughly how many people, and for what. Add IT's logs of traffic to the most common services if they exist. The aim is a picture of the position, and of which uses are harmless, such as rewording emails, summarising long texts and translating.

What you are looking for are the two or three uses that are not harmless: HR matters, customer data in volume, ongoing disputes, and anything involving clients if you are a law firm.

Day two: stop only the worst

Go to the teams where the risky uses are and ask them to stop just those until an approved tool is available. Say why: no contract, no control over retention, and for law firms no basis for disclosing client information. Say at the same time that an approved tool is coming, and when. A ban with no date for the alternative is hard to maintain.

If something has already gone in that should not have: assess whether it is a personal data breach (see the FAQ), ask the employee to delete the conversation, and note that deletion at the vendor takes time and does not reach everything.

Days three and four: offer an approved alternative

The alternative is what makes shadow AI go away. A company agreement with a vendor, with a data processing agreement, retention settings and login through the company's identity system, is what makes staff switch. Which vendor depends on what the company already has. If you have Microsoft 365, Copilot is the quickest to put in place. If the legal department wants a tool that can be configured with its own rules and contracts, Claude Team or Enterprise is an option, see Claude for Legal. The choice between them is a question of technology selection. For the step out of the shadow, it is enough that there is a contract and that logging in is easy.

The five data protection questions should be answered before the first matter goes into the new tool, see the checklist. Until then the rule from day two applies.

Day five: write the rule

One page. Which tools are approved and under which contract, what may be entered, which matters are kept out even in the approved tool, and whom to contact. The structure is in the guide on AI policy for the legal department. Announce the rule on the day the approved tool opens. Announced earlier, the rule is only a ban, with no way to do the right thing.

The week after: follow up

Two hours of training for those who use the tool, documented as a measure under Article 4 of the AI Act, see the guide on the AI Act for in-house counsel. Look at the logs again after a month. If traffic to private services has not gone down, the alternative is too cumbersome, and it is the alternative that needs fixing.

What a week cannot do

Get a full picture of everything already entered. Get the vendor to delete it. Replace tools the business has already built routines around. That is work for the coming months, and it only works if the approved alternative is there first.

HOVR Legal produces the five answers for the tool you choose in the data protection package, and the guideline and the training are included. If the department also needs an operating model for how AI is used in daily work, see AI operating model for the legal department.

Frequently asked questions

Is it a personal data breach when an employee has pasted customer data into a private AI account?
It can be. The data has gone to a recipient with no data processing agreement and to a third country with no transfer basis. Assess whether it is a personal data breach. If it is, it must be notified to IMY (the Swedish Authority for Privacy Protection) within 72 hours, unless it is unlikely to result in a risk to the data subjects. Document the assessment, even when the conclusion is that no notification is needed.
Can we just block ChatGPT and Claude in the firewall?
You can, but use then moves to mobile phones and to tools you do not know about. Blocking works when there is an approved alternative that is just as easy to reach. Without an alternative, blocking creates more shadow use.
Is it enough to tell everyone not to enter personal data?
No. The rule cannot be followed in practice, because so much text about work contains personal data: names of counterparties, employees, contact persons. The rule has to say which tool may be used, under which contract, and which matters are kept out even there.

Statements about Anthropic's terms and products were read on Anthropic's pages on 9 November 2026 and are checked at the start of every engagement.

Sources

  1. [1]Commercial Terms of Service, Anthropic
  2. [2]Vägledning om användning av generativ AI i advokatverksamhet, Sveriges advokatsamfund (Swedish Bar Association), 12 June 2025 (in Swedish), section 1 on open tools with private user terms
  3. [3]GDPR-vägledningen, Sveriges advokatsamfund (Swedish Bar Association), 10 September 2026 (in Swedish)

Back to AI for legal

More under AI for legal