Finding shadow AI without surveilling your people

Once shadow AI surfaces, the delicate question is not which tools are running in the company. It is the question of who is using them. In most companies that can be found out technically. In Swiss employment law the limits start exactly there, and they start earlier than most people expect: with the one-off spot check.

The tool question is an inventory task. The person question is an intrusion into your employees' privacy. That the usage is already happening before anyone approved it is what we described in shadow AI in companies. This article is about the step after that, about how far you may go when looking for it.

May I check which AI tools my employees use?

At the level of the tools yes, at the level of the individual person almost never. Article 26 of Ordinance 3 to the Swiss Labour Act prohibits monitoring and control systems intended to monitor employee behaviour at the workplace. A list of subscriptions, accounts and admin approvals is untouched by this. The evaluation by name of who opened which AI window and when is the part that tips over.

What the ordinance actually says

The text is short. Paragraph 1 of Article 26 ArGV 3 reads, in the official German: monitoring and control systems intended to monitor the behaviour of employees at the workplace may not be used. Paragraph 2 adds that systems required for other reasons must be designed and arranged so that the health and freedom of movement of employees are not impaired.

The detail sits in the SECO guidance on that article. It spells out what counts as behaviour monitoring, and the list covers precisely what you would look at for an AI search: IT means that monitor the computer or mobile phone activities of employees in the company or in their home office, with email, spyware, activity trackers, app and web logs, mouse and keyboard logs named in the brackets behind it.

The same text makes a second point: IT means that use artificial intelligence for automated analysis and evaluation of employee-based data also count as behaviour monitoring. Pointing an AI tool at your workforce's usage data is therefore not the elegant shortcut. It is the same problem with more compute.

The definition itself is wider than the term suggests. According to SECO, behaviour monitoring is any monitoring that allows continuous or non-continuous control, meaning briefly periodic or sample-based control, of certain activities in detailed form. The single look into the logs is therefore not the harmless variant, it is the spot check the text names explicitly.

What does not count as behaviour monitoring are arrangements such as electronic badges for access control and time recording, or technical systems for quality control. The purpose makes the difference. The same tool can land on either side of the line, depending on what you use it for.

One more detail, because contracts often try it: SECO states that these provisions cannot be set aside by a private agreement, not even with the consent of the employees or their organisations. The clause in the employment contract that permits monitoring carries less than it looks like.

The detour through the logs you already have

In our experience the proposal rarely comes from the management team, it comes from the IT provider, and it sounds helpful: a report from the proxy showing which AI sites were visited, per user. We know the argument behind it almost word for word. We are not introducing a new system, the data is already there, proxy, firewall and endpoint protection have been logging all along.

True, and it does not change the question. Admissibility is decided by the evaluation, not by where the data came from.

On the control of internet use at the workplace, the Federal Supreme Court held that the evaluation by person of data recorded by name had taken place unlawfully (BGE 143 II 443). In that individual case the employer was still allowed to use the result after weighing the interests. The evaluation itself remained unlawful. That is not a reassuring finding: the route was wrong, and whether anything from it holds up in the end is decided by someone other than you.

Behind this sits Article 328b of the Swiss Code of Obligations. An employer may process data about an employee only insofar as it concerns the employee's suitability for the employment relationship or is necessary for the performance of the employment contract. The Federal Supreme Court has made clear that an employer who accesses an employee's private messages violates that employee's privacy (judgment 4A_518/2020 of 25 August 2021).

Then there is the Data Protection Act. In its overview of technical means of monitoring at the workplace, the Federal Data Protection and Information Commissioner describes the sequence a company runs through beforehand: before it uses direct monitoring by technical means, it has to clarify whether the goal can be reached with less intrusive measures. For the evaluation of peripheral data, meaning the logged information about who did what, when and with whom, proportionality requires the form with the least intrusion: anonymised, pseudonymised or in its original form, depending on what serves the purpose.

That turns the legal question into a practical one. Your purpose is a list of tools. For that purpose you do not need names.

When a control system is admissible after all

What is prohibited is not monitoring systems as such, but those aimed exclusively or mainly at monitoring behaviour. The Federal Supreme Court worked this through on the case of a GPS system in company vehicles: with a justification such as safety or work organisation, a system is admissible if it is in reasonable proportion to the goal and the people affected were informed in advance (BGE 130 II 425).

SECO turns this into three conditions that have to be met together: a clearly overriding other interest, for example the safety of the staff or of the operation; proportionality between the employer's interest and the employees' interest in not being monitored; and the participation of the employees in the planning, installation and operating times of the system as well as in the storage period of the data.

Paperwork belongs to it. Where permitted performance or safety monitoring cannot be cleanly separated from prohibited behaviour monitoring, the company has to examine its overriding interest and the proportionality and justify both in writing. That justification has to be submitted to the competent cantonal labour inspectorate on request.

That is the honest price of the technical answer. It is not prohibited, it costs a named overriding interest, a weighing of interests, the participation of the workforce, a defined storage period and a document that holds up in front of an authority. For the question of whether someone in marketing drafts proposal texts with a free account, that effort is out of proportion. This applies on operational grounds alone, before anyone talks about law.

Finding shadow AI without introducing a control system

The route that works in our mandates needs none of that justification, because it evaluates no personal data.

First, the documents you already have. The subscription list, the admin consoles of Microsoft 365, Google Workspace and Zoom, credit card and expense statements, industry applications with new assistants. This list names tools, not people, and it covers the largest part. What such an AI inventory looks like, and which five places are enough for a first pass, we wrote up separately.

Then ask the team, without the threat of sanctions. In our experience that round delivers more than any log, because it supplies the reason along with the answer: which process is too slow and which company tool is missing. That information sits in no log file.

Third, the rules on one page, including the question of what private use is allowed on company devices. The Commissioner places this in the context of Article 321d of the Code of Obligations: against that background, the employer sets out in a usage policy how employees may use the internet and email for work and privately, and defines what counts as abusive use. Without that, you later lack the yardstick to judge anything against. How short such a page may be is in do we need an AI policy?

What comes before all of this is transparency. The Commissioner recommends that companies draw up a set of rules for monitoring, describing how it works, what kind it is and when it happens, and states that processing has to take place in a way that is recognisable to the person concerned. Saying beforehand what gets logged costs nothing and closes most of the later discussions.

A word on consent, because it often shows up in projects as the solution: the Commissioner treats consent in the employment relationship with caution, because employees feel under pressure and restricted in their free decision. A checkbox in onboarding does not replace the weighing of interests.

When there is a concrete suspicion

The case exists. There are indications that someone is putting customer data into a tool the company knows nothing about. The same sequence applies, only concretely: a named purpose, the mildest measure that fulfils it, one person who carries the decision, and a note on why there was no other way. Access to the data has to be regulated and limited to the people authorised to evaluate it, as the Commissioner puts it. Honestly, in such individual cases the line stays blurred, and that is not a comfortable answer. It is the reason why the decision has to sit with a named person and not with whoever happens to hold read rights on the logs.

What does not fall within your remit is criminal prosecution. SECO writes it plainly: punishing criminal offences inside a company is in principle not a matter for the company but exclusively for the police. Those who ignore this end up with monitoring data that is mostly unusable in criminal proceedings.

Purpose, responsibility and a set of rules are the part of AI governance and secure AI adoption that comes before any technical question. And once the list exists and the tools are meant to deliver something in daily operations, this is where it continues: AI in operations.

Two lists on the table

In the end two possible lists sit in front of you. One names people and timestamps and is the harder route in employment law terms. The other names tools, accounts, data categories and owners.

The second one looks like the cautious substitute. It is the only one a decision follows from: approve, consolidate, stop. The decision is made about a tool, never about a person, and for that decision you do not need a single name.

If you want to know what this looks like in your company before anyone opens the logs, a free first conversation is the right setting. No obligation, and usually shorter than the discussion that arises without this assessment. Book a slot here.

Frequently asked questions

May I evaluate the proxy logs to find AI usage?

Not on a per-person basis before purpose, proportionality and prior information are settled. The Federal Supreme Court judged the evaluation by name of recorded internet data to be unlawful (BGE 143 II 443). For a list of tools an anonymised or aggregated view is enough, and usually the subscription list and admin consoles already answer it.

Do I have to inform employees in advance when something is logged?

Yes. Processing has to be recognisable to the person concerned, and the Federal Supreme Court requires prior information of those affected for admissible monitoring systems. The Commissioner recommends a set of rules describing how the monitoring works, what kind it is and when it happens. Announcing it costs nothing.

What belongs in a usage policy for AI tools?

Which data categories may go into approved tools, which tools are approved, who decides about new ones, and what private use is allowed on company devices. The Commissioner places private use of internet and email in the context of Article 321d of the Code of Obligations. One page is enough to start.