---
title: "The last human in the loop is a signature"
url: "https://demertis.com/blog/the-last-human-in-the-loop-is-a-signature"
description: "Most automation programmes stop at the same place, and it is not a technical boundary. Somebody has to be accountable, and accountability needs a person."
---

# The last human in the loop is a signature

September 4, 2026·2 min read·Demertis

Follow almost any automation programme to its end and you find the same shape. The system gathers, decides, prepares — and then stops, holding something for a person to approve. Teams usually describe this as a maturity problem: the model is not accurate enough yet, the edge cases are not covered, next year we will remove the step. Years pass and the step is still there, because it was never really about accuracy.

## What the signature is actually doing

It is attaching a decision to somebody who can be held responsible for it. That is a legal function, not a technical one, and no amount of model improvement satisfies it.

This becomes obvious the moment you ask what would change if the system were perfect. If it made no mistakes at all, who would the counterparty sue over a contract it agreed to? Who would the regulator sanction? Whose duty of care was breached? The answers do not improve with accuracy, because they were never about accuracy.

## Three things the step provides

**A party.** Contracts, orders and payments need somebody with the capacity to be bound. A process cannot be bound to anything.

**A defendant.** Somebody must be answerable when the outcome is wrong. Liability that lands nowhere is liability the legal system will relocate onto whoever is nearest.

**A moment of legibility.** The approval is when a human last understood the state of the system. Remove it without replacing it and you lose the only regular point where anybody looks.

The first two cannot be automated today in any jurisdiction we are aware of. The third can be, and that is where the engineering work actually is.

## What replaces it, if anything

Not a better model. What replaces the approval step is a structure that makes the loop governable without a person standing inside it: a permanent record of what was observed, concluded and done; bounds on what the loop may touch, enforced outside it; an interrupt the system cannot influence; and a named legal owner who carries the consequences and has the information to act on them.

Notice that the last item still contains a human. The point is that they are outside the loop rather than inside it — accountable for the operation as a whole, rather than clicking approve on individual decisions they have no realistic capacity to evaluate.

That is a much better place for a person to be. The approval step, in practice, is often a person confirming things at a rate that makes genuine review impossible, which produces the appearance of oversight and none of the substance.

## Why this is the honest framing of the problem

Saying the goal is a business with no humans in it is a claim about the law, and the law does not currently allow it. Saying the goal is a loop that can run unattended, under an owner who can be held to account and can actually see what it did, is a claim about engineering — and that one can be worked on.

The rest of this programme follows from that distinction. Nearly every hard problem in autonomous operation turns out to be a question of who is answerable, what was recorded, and how it stops.

-   automation
-   accountability
-   approval
-   control loops

## Keep reading

-   [Self-healing is not a feature, it is a different architecture](https://demertis.com/blog/self-healing-is-not-a-feature)
-   [Can a machine own anything?](https://demertis.com/blog/can-a-machine-own-anything)