In short: As long as AI only suggests text, its risks are manageable and AI risk management stays simple. As soon as agents act, placing orders, planning, answering customers or changing data, new kinds of risk emerge that most risk reports still don’t cover. Good AI risk management names five of them explicitly: autonomy, cost, vendor dependence, knowledge leakage and regulation. A board should have them reported with owners, metrics and thresholds.
Why does AI belong in the risk report?
Because AI agents intervene in processes that matter for revenue, customers and liability. In Germany, management’s duty to detect developments that threaten the company’s existence early is written into law, for stock corporations in the Stock Corporation Act and for managing directors of legal entities in general in the Corporate Stabilization and Restructuring Act (StaRUG). Whether AI risks threaten a company’s existence in a given case depends on the company. That they should be captured in risk management as soon as agents act operationally is, in my view, beyond dispute.
What risks do AI agents bring?
| Risk | What it’s about | Possible metric |
|---|---|---|
| Autonomy | Agents act in areas where nobody has explicitly decided that they may | Share of agents with a documented autonomy level and approval rule |
| Cost | Compute costs grow with usage and case difficulty, and errors such as endless loops can multiply them | Cost per completed case, variance from plan, number of times cost caps were exceeded |
| Vendor dependence | Core processes depend on a single model or platform provider whose prices, terms and availability the company doesn’t control | Share of applications with an interchangeable model, fallback plan in place yes/no |
| Knowledge leakage | Corrections and exceptions that employees teach an outside system improve that vendor’s product, not your own | Contractual rules on learned knowledge per vendor |
| Regulation | Obligations under the EU AI Act, data protection law and co-determination are not met | Completeness of the AI inventory, open review items |
What does autonomy risk mean in practice?
Agents often get their rights step by step, first for suggestions, then for routine cases and eventually for more. If that expansion isn’t decided explicitly, autonomy emerges out of habit. The risk lies less in any single error than in nobody knowing anymore what a system is allowed to do, and in being unable to trace who approved it when something goes wrong. A board should therefore ask for an overview: which agents act where, at what autonomy level, who decided that and when was it last reviewed?
Why is vendor dependence a board-level issue?
Many companies build their AI applications on the models of a single large provider. That’s convenient, but it turns prices, terms of use and availability into external factors the company doesn’t control. There’s also a subtler form of dependence. Gartner explicitly advises vendors to capture and retain customer-specific knowledge (Gartner, July 1, 2026). From the customer’s point of view, that can mean that what your own employees teach an outside system stays with the vendor. If you don’t address this in the contract, you may be funding improvements to a product your competitors use too.
Which regulatory risks matter?
The EU AI Act applies in stages. Prohibitions on certain practices and the duty to promote AI literacy have applied since February 2025, transparency obligations since August 2026, and obligations for high-risk systems under Annex III take effect in December 2027. Fines for prohibited practices can reach €35 million or 7 percent of global annual revenue. On top of that come data protection and, in Germany, co-determination rights whenever systems can track employee performance or behavior.
What questions should the board ask?
- Is there a complete inventory of the AI systems that act autonomously in the company?
- Is the autonomy level of each of these systems documented and approved by an accountable person?
- Are there cost caps per case and a handoff to humans when they are exceeded?
- Can we switch model providers, and how long would that take?
- Who contractually owns the knowledge our employees teach outside systems?
- How do these points feed into the regular risk report, with owners and thresholds?
Frequently asked questions
What risks do AI agents pose for companies?
Key risks include unclearly granted autonomy, hard-to-plan compute costs, dependence on individual model providers, leakage of company knowledge to vendors and regulatory risks under the EU AI Act, data protection law and co-determination.
Does AI have to be covered in risk management?
As soon as AI systems intervene in business-critical processes, their risks should be captured, assessed and reported like other operational risks. The specifics depend on legal form, size and area of use.
How can you reduce dependence on AI vendors?
Through architectures in which models are interchangeable, fallback plans for critical applications and contracts that settle who owns the knowledge learned in operation.
This article does not replace legal advice on the duties of management and boards.
Further reading: EU AI Act for advisory boards · AI expertise in the boardroom · Five questions for management
Sources: Gartner press release, July 1, 2026 · Regulation (EU) 2024/1689 (AI Act), EUR-Lex
