Quick answerBuild a short, explicit, board approved list of roles and decision types that stay human no matter how good the underlying model gets, reviewed on a fixed schedule rather than left to drift. Good candidates almost always include final termination decisions, decisions that waive a legal right, anything requiring a licensed professional's signature by law, and any irreversible action above a defined severity threshold. The list should exist independent of any single incident and should not shrink automatically just because the technology improved.
Why this needs to be a standing list, not a case by case call
Most companies handle edge cases through an exception approval workflow that lets a human override an automated rule when the situation warrants it. That is the right tool for handling unusual situations inside decisions that are otherwise automated. It is the wrong tool for deciding, once and for all, that an entire category of decision should never be automated in the first place, no matter how the exception process performs. Those are different governance questions and deserve different documents.
Why a single domain refusal pattern is not enough either
Plenty of companies have solid refusal boundary design for a narrow domain like medical or legal advice, where the agent is trained to decline and redirect. A standing never automate list is broader and more structural: it is not about the agent recognizing a request it is not equipped to answer, it is about a small number of decision types the company has committed, as policy, to never route through automation at all, regardless of how confidently the agent could technically answer them.
What actually belongs on the list
Good candidates share a common shape: they are hard to reverse, they carry disproportionate legal or human consequence relative to their frequency, or the law specifically requires a licensed or authorized human. Final employment termination decisions, anything that waives a customer's legal rights, medical or legal determinations requiring a licensed professional's sign off, and any single action above a defined financial or safety severity threshold are common entries. The list should stay short. A list with fifty items stops functioning as a clear boundary and starts functioning as a document nobody actually reads before shipping a new automated flow.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
Keeping human in the loop from becoming a moving target
There is a real temptation to quietly shrink this list as the underlying model gets more capable, on the theory that a stronger model earns more autonomy. Resist making that call implicitly. If an item comes off the list, that should be its own explicit, reviewed decision with its own sign off, similar in weight to how the list was approved in the first place, not something that happens because nobody re-litigated an old boundary while shipping an unrelated feature. This overlaps with the same discipline used for human in the loop approval design more generally, applied here at the policy level rather than the individual decision level.
Who owns keeping the list current
Assign a named owner and a fixed review cadence, at minimum annually, independent of any specific incident. A list that only gets revisited after something goes wrong is really an incident response document wearing a governance label. The goal of a standing list is to have already decided the hard cases before they happen, not to be writing the policy for the first time in the middle of a crisis.
FAQ
Does this list need board approval?
For the highest severity items, yes. A list that any single team can unilaterally edit is not really a standing commitment, it is a preference that happens to be written down.
How is this different from a safety refusal built into the model's prompt?
A prompt level refusal is a technical control that can fail, be bypassed, or degrade with a model update. The standing list is an organizational commitment about which categories of decision are routed to a human by design, regardless of what any individual prompt does or does not catch.