The Boss in the AI Shadow Economy

It is 10 p.m., the office empty. A board member copies an internal strategy draft into a personal ChatGPT to clean up the language. An hour later the result lands in the inbox of the supervisory board chair. Nobody sees it. Nobody documents it. From August 2, 2026, this is a compliance breach with consequences that do not end at the company.

Since December 6, 2025, Germany's NIS2 implementation act has been in force, without a transition period. Around 30,000 organizations are in scope, well above the roughly 5,000 under the previous IT security law. Paragraph 38 BSIG puts leadership personally on the hook. They have to approve risk management measures, oversee implementation, and attend cybersecurity training. If they fail in that duty, they are liable with their private assets. A waiver of that liability is excluded by law. In serious cases, executives can be barred from holding leadership positions.

On August 2, 2026, the EU AI Act becomes fully applicable. Article 99 sets three sanction tiers. Up to 35 million euros or 7 percent of global annual turnover for prohibited AI practices. Up to 15 million or 3 percent for breaches of obligations for high-risk systems. In practice, the AI Act pretty much forces a complete inventory of all AI systems in the company, because neither conformity assessment for high-risk systems nor the literacy obligation in Article 4 can be met without one. If you don't have it, you have a problem.

Here it gets interesting. A representative study by the ifo Institute together with ZEW, IAB, BIBB and BAuA, based on around 9,800 interviews with employees subject to social insurance contributions in Germany, shows that 64 percent of them have already used AI at work. Only about one in five uses it regularly. For two thirds of those who use it, the use happens without any formal introduction by the company. Only for one third was it officially rolled out by the employer. A 2025 Bitkom survey of 604 companies adds the company side. In 8 percent of firms, the use of private AI tools like ChatGPT is already widespread, in another 17 percent there are individual cases. At the same time, only 26 percent of companies provide their staff with official access to generative AI.

That leaves an uncomfortable picture. The person signing in August that no undocumented AI system runs inside their company is statistically the same person who just opened one. It is also explainable. Executives write more, decide faster, have less oversight, and more deadline pressure than the back office. Whoever needs one more slide at 9 p.m. for a board meeting grabs the fastest tool available. Anyone holding that against them has never sat in a board meeting.

The usual reaction to numbers like this is a ban or a memo. Both have been failing in predictable sequence for two years. Shadow AI does not need code, admin rights, or a VPN. It needs a browser and a deadline. Bans push usage onto personal devices and personal accounts, where compliance teams cannot reach. The third way, beyond ban and allow, is to replace. Offer an environment that is faster than the private tab, visible to IT, documented for audit. Sounds simple. In practice it is rare, because internal platforms are too often built with sign-in forms, approval rounds, and ticket queues, rather than starting with the one question that matters. Does this open in two clicks, and does it give a better answer than what is freely available outside.

In my own work the pattern repeats. Leadership is convinced that at home, everything is under control. Then you sit down with a marketing or sales team, and it does not take half an hour before tools land on the table that no one on the management floor has heard of. That is not bad faith, it is organized self-help. There is no German mid-sized company without Shadow AI. There are only those that do not know it yet, and those waiting for someone else to sort it out. And there is no employee using ChatGPT out of malice. They use it because the official tool is slower than their private one. Behind every private AI tab is an open request to the organization itself.

In August, EU regulation does what no IT strategy managed. It makes looking away expensive. Personally expensive. It also has a side effect the Brussels drafters never intended. It forces executives to engage with the actual working reality of their own staff.

That shifts the question. From August on, it is not how to ban Shadow AI in my company. It is two questions. First, do I actually know what my people are using. Second, is what I officially offer as good as what they open in a private window at night. Whoever answers the first one honestly has the inventory. Whoever answers the second honestly will sleep better in August.