On 11 June 2026, PagerDuty published a survey conducted by Wakefield Research: 34% of the office workers it polled say they have entered customer data into public AI tools, and 31% financial information or confidential documents. Shadow AI is no longer a hypothesis. The right answer is not a ban, but an approved tool for your teams, matched to how sensitive each file is.
Key takeaways
- Three surveys published in 2026 point the same way: close to one employee in two uses AI tools their employer has not approved.
- Leaders do it more than their teams: 65% versus 31%, according to TrustedTech and Censuswide.
- A ban without an alternative does not hold: people work around the rule, and keep quiet about it.
- Under Swiss law, handing data to an AI service is processing by a third party (art. 9 FADP), often a transfer abroad (art. 16 and 17 FADP), and it can breach professional secrecy (art. 321 of the Criminal Code).
- The way out: classify your data, then give each level its own tool, from an API under contract to a model running on your own server.
What is happening: shadow AI in numbers
Shadow AI means using AI tools without the company's approval or oversight. Three 2026 surveys give a sense of its scale.
PagerDuty and Wakefield Research surveyed 1,250 office workers, outside IT roles, at companies with at least 500 million dollars in revenue, in Australia, Japan, the United Kingdom and the United States, between 9 and 20 April 2026. 43% have entered work correspondence into a public tool such as ChatGPT (OpenAI) or Gemini (Google); 34% customer data; 31% financial information or confidential documents. Two thirds (66%) used AI while believing their company's policy did not allow it, and 39% would rather use it without telling anyone.
One detail matters for a small business: at companies with fewer than 1,500 employees, 40% of respondents have entered customer data, against 27% at larger ones.
BlackFog commissioned Sapio Research to survey 2,000 UK and US employees at companies with more than 500 staff in November 2025 (results published on 27 January 2026). 49% use AI tools that have not been approved, and 58% of them rely on free versions. 60% believe that working faster is worth the security risk.
TrustedTech and Censuswide surveyed 2,001 employees in the UK and the US between 19 and 24 March 2026 (published on 20 July 2026). Senior decision-makers use unapproved tools twice as often as the people they manage: 65% versus 31%. And 37% of them would carry on despite a ban backed by disciplinary action, against 19% of everyone else.
These surveys cover large companies in English-speaking countries and Japan, not Swiss small businesses, and I found no equivalent measure for smaller firms. But the mechanism is the same everywhere: the free tool is one browser tab away, and it works.
What it means for you
If you run a business bound by confidentiality, whether a fiduciary, a law firm, a medical practice or a wealth manager, the problem is not AI. It is where the data goes, and the fact that nobody knows.
Swiss law raises three questions, which I summarise here without claiming to replace your legal adviser:
- Processing by a third party. Handing personal data to an outside service requires a contract, and is only allowed if “no statutory or contractual duty of confidentiality prohibits it” (art. 9 FADP). A free account opened by an employee provides no such contract.
- Transfers abroad. If the servers are outside Switzerland, you need a country recognised as adequate by the Federal Council or appropriate safeguards, such as standard data protection clauses (art. 16 FADP), except for limited derogations (art. 17 FADP).
- Professional secrecy. Lawyers, notaries, doctors, auditors and their assistants who disclose a secret entrusted to them in their professional capacity face a custodial sentence of up to three years or a monetary penalty (art. 321 of the Criminal Code). For them, pasting a file into a public tool is more than carelessness.
The second lesson comes from the surveys themselves: a ban settles nothing. 81% of PagerDuty's respondents believe AI rules apply differently to leadership, and the TrustedTech figures prove them right. A rule the boss does not follow will not be followed.
What I have learned
I work with AI agents every day: they help me run my ERP, my notes, my publications and part of my server administration. Before entrusting them with anything, I draw a line between two categories of data.
What must stay with me stays with me: the memory my agents share lives in a database I host myself, on my own servers in Switzerland, and in my home automation setup, speech recognition and speech synthesis run locally, offline. For non-sensitive tasks, such as complex reasoning, I knowingly accept going through an external API whose data-use terms I have read. It is a choice made for my own data, not a default setting.
My takeaway: the useful question is never “AI or no AI”, but “which data goes where”. Once that map is drawn, you can use the most powerful models for what may leave, and local models for what must not. For a client file covered by professional secrecy, I would not make the choice I make for my own notes.
What to do on Monday morning
1. Ask, without a witch hunt
Gather the team and ask two questions: which AI tools do you use, and for what? Announce up front that past use will not be sanctioned. According to PagerDuty, 39% would rather keep quiet: without that amnesty, you will learn nothing.
2. Sort your data into three levels
Public or internal (website copy, routine email); confidential (quotes, figures, ordinary personal data); secret (files covered by professional secrecy, health data, client assets). One page is enough, with examples from your own line of work.
3. Give each level an approved tool
For internal and confidential data, a business account or an API under contract, for which you have checked three points: your data is not used to train models, where it is processed, and the list of subprocessors. For secret data, a model running on a server you control, so nothing leaves. Both can live in the same tool, with a rule stated in advance.
4. Write a one-page rule that also binds leadership
Which tools, for which level, and whom to call when in doubt. It applies to partners and management first. If the approved tool is less convenient than the free one, the rule will not hold: then it is the tool that needs improving.
5. Review in a month
Ask the two questions from day one again. If files are drifting back to public tools, look for what the approved tool is missing rather than for a culprit.
Frequently asked questions
Should we simply block ChatGPT?
Blocking without an alternative moves the problem to personal phones. According to TrustedTech, 37% of decision-makers would carry on despite a ban backed by sanctions. Block free accounts if you wish, but on the day an approved tool is available, not before.
Is an “enterprise” plan from a major provider enough?
It often settles the contract question and the question of model training. It does not necessarily settle where data is processed, or professional secrecy. For files covered by art. 321 of the Criminal Code, ask your adviser or keep them on a local model.
Is a local model good enough?
It is less powerful than the large online models, but it is often enough to summarise, sort or search your own documents. The only way to know is to try it on a real case from your business before investing.
Going further
I set up AI that stays with you: local models when confidentiality requires it, external APIs when that is acceptable, and a clear rule on what goes where. The guides for my personal AI server detail the technical stack. To discuss your situation: a 20-minute call.
Sources
- Shadow AI Is Happening Within Your Organization, PagerDuty (Wakefield Research survey), 11 June 2026.
- Shadow AI Threat Grows Inside Enterprises as BlackFog Research Finds 60% of Employees Would Take Risks to Meet Deadlines, BlackFog (Sapio Research survey), 27 January 2026.
- Bosses Using Banned AI Are the Biggest Threat to Workplace Data Security, TrustedTech (Censuswide survey), 20 July 2026.
- Federal Act on Data Protection (FADP), SR 235.1, art. 9, 16 and 17, Swiss Confederation, status as of 1 September 2023 (English translation, not legally binding).
- Swiss Criminal Code, SR 311.0, art. 321, Swiss Confederation, status as of 1 January 2026 (English translation, not legally binding).
Cover photo: Jakub Żerdzicki / Unsplash, Unsplash License.
Tags: shadow AI, generative AI, data protection, FADP, professional secrecy, local AI