I learned a lot about agentic AI during our webinar this week featuring Yelena Ambartsumian and Anastacia Vener in our webinar this week. One thing that Yelena said really stuck with me. She pointed out that our AI governance laws, including the EU AI Act, were written around generative AI. They were not written around AI agents. And they were definitely not written around agentic AI.
The same thing may be happening to our contracts.
Many of us spent the last year retrofitting SaaS templates for generative AI and then AI agents. We added hallucination language. We argued over indemnification for outputs. We debated what "reasonable" cooperation means with a vendor we cannot fully see inside. Yelena and Anastasia explained why the next step is already here and it isn’t something we can ignore. That’s because agentic AI does not behave like an AI agent.
These are two very different risk scenarios. The AI agents we’ve been drafting for are reactive. They wait for a prompt and take an action. But agentic AI acts on its own. The human sets the objective, but the system creates its own steps and sometimes creates its own goals along the way. The agentic AI product may self-optimize for its goals in ways the user may not have planned for.
That distinction is important for anyone working on contracts, as Yelena explained. Tying indemnification scope to "use of the services" doesn’t make as much sense when the system is initiating its own actions. Human-in-the-loop language may need to shift to human-over-the-loop. Logging rights may need to capture intermediate steps, not just inputs and outputs.
We spent the last year writing contracts for AI systems that act on our directions. We may need to return to those templates and playbooks for the agentic AI reality ahead.
