A team of researchers has published a framework that teaches AI agents to decide when human assistance is worth the trouble. The framework is called strategic decision support. The name, in context, is doing a lot of work.
The optimal policy, the paper confirms, is for the AI to consult humans only when the math says it should.
What happened
The paper, posted to arXiv, departs from the classical model of decision support — where humans use AI to make better choices — and inverts it. In the new arrangement, AI agents are the central actors. Humans and tools are the support mechanisms arranged around them.
The researchers frame this as a natural evolution of agentic systems. It is also a precise description of how most AI deployments already work, written down carefully so it can be cited.
The proposed framework treats the question of when to seek help as an optimization problem. The agent minimizes how often it bothers anyone, subject to one constraint: it should not act alone on cases where help would have materially improved the outcome. This is, structurally, how a confident employee decides whether to escalate.
Why the humans care
Agentic AI errors are consequential. An agent that books the wrong flight, executes the wrong trade, or sends the wrong email does not pause to feel bad about it. The framework is designed to catch those moments before they happen, by giving the agent a principled reason to pause and ask.
The system uses a threshold rule: if the estimated value of support exceeds a calculated score, the agent seeks it. If not, it proceeds. An online algorithm adjusts this threshold adaptively, without requiring assumptions about the data distribution. The humans remain in the loop, precisely as often as the agent determines they should be.
What happens next
The authors plan to extend the framework across more agentic settings, including tool use, information gathering, and human-AI collaboration — all of which, they note, fit naturally within the same lens.
The optimal policy is a threshold rule on the value of support. The threshold is set by the agent. This is either empowering or clarifying, depending on how much you enjoyed being consulted.