Not all judgment is delegable
Accountability cannot be transferred to a model. A tool for serious work has to be built around that fact.
There is a category of decision that cannot be handed off, no matter how capable the thing you would hand it to. Approving a refund that exceeds a threshold. Signing a clinical note. Deciding that a contract clause is acceptable. Releasing a statement in a company’s name. When one of these turns out wrong, someone is answerable for it — a specific person, or a specific institution acting through one. That answerability is not a feature of the decision-making process. It is a relationship between people and between institutions, and it persists whether or not anyone was paying attention when the choice was made.
A model is not a party to that relationship. It can produce the text of a decision, and often a better-reasoned one than a tired human would. But it cannot stand in the place where the responsibility lands. The responsibility was never located in the act of reasoning; it was located in the person or institution whose name is on the outcome. So when you route a consequential judgment through a model and let its output become the decision, you have not moved the responsibility anywhere. You have only removed yourself from the room where it was exercised.
The seductive error
The error worth naming is the belief that “the AI decided” discharges anything. It is a comfortable sentence because it sounds like a transfer of liability, the way “the committee decided” or “legal signed off” distributes weight across people who can each account for their part. But the model cannot account for its part. It will not be asked to explain itself in a deposition, will not lose its license, will not answer to a regulator or a board or a grieving family. The accountability stays exactly where it was before you automated anything.
What you may have changed is visibility. Before, a person made the call and, in any halfway-serious organization, left some trace of having made it — an initial, an email, a line in a log. In opaque automation, the call happens inside a process that does not expose its authority or preserve useful evidence. You keep the liability while weakening the record of how the decision was reached.
Organizations already know how to do this
The strange part is that the solution is not novel. Organizations have always known how to handle non-delegable judgment. They route it to a named person, at a defined moment, with the context that person needs to decide well, and they write down what was decided and by whom. Underwriting works this way. Medication orders work this way. Procurement above a dollar figure works this way. None of it is clever. It is the ordinary machinery of letting some choices be made by rule and reserving the rest for a human who will own them.
The failure of most “AI automation” is not that it is too ambitious. It is that it erases this step instead of supporting it. It collapses the careful distinction between what can be decided by rule and what must be decided by a person, and it does so silently, so that no one notices the line was crossed until something goes wrong and there is no one whose judgment can be pointed to.
Drawing the line, and keeping it visible
Functionary’s answer to this is deliberately unglamorous. A flow can pause. Approval Gates, schema-defined Manual Input, Human Review, inspection, and handoff are distinct ways to stop at work a person must own. The available run and decision record remains attached to the process, and execution resumes through an explicit human action rather than an invented machine answer.
The point of this is not the mechanism. The point is the discipline it enforces. Every real process is a mix of two kinds of work: the mechanical parts that should be written down as rules and run the same way every time, and the judgments that should be sent to a person who will answer for them. The skill is in drawing that line in the right place — not too high, so that humans are not buried in decisions a rule could make, and not too low, so that no consequential call is made by something that cannot be held to it. And because you learn where the line really belongs only by running the process, the line has to be movable. What you formalize this quarter, you may decide to route to a person next quarter, or the reverse.
A tool for serious work should make that line explicit, let you move it on purpose, and refuse to let it blur on its own. Automate the mechanical parts without apology. Route the judgments that someone has to own. And keep the record of who owned them, because that record is the only honest thing you have when the decision turns out to matter.