← Tech blog

The judgment line moves: evidence, ceilings and the delegation ladder

· Prabhu Eshwarla

This is an update to the judgment line, applied to a whole process, which I published in September.

What the September version said

Six steps. List your processes. Shortlist them on evidence, measuring touch time against elapsed time. Break each process into steps. Choose what executes each step: rules, programs, AI or people, always the most deterministic one that can do the job. Draw the judgment line, deciding for each step whether the machine finishes it, the machine prepares and a person finishes, or a person does it with the machine's briefing. Write it down.

Why I am updating it

Writing the delegation ladder, about which work to hand to AI first and how trust is built, exposed a gap. The ladder describes work moving from "a person decides" to "AI acts within limits" as evidence builds. The September piece drew the judgment line as if it were drawn once. Both cannot be right, and the September piece was the one that was incomplete: it had no element of time.

What stays the same

What changes

The rest of this piece explains each change.

The line is drawn on day one's evidence

A judgment line is drawn on the evidence you have on the day you draw it. On day one you have very little. Six months later you have a record. When a machine first takes on a step, nobody knows how often it is right. The honest answer is usually that the machine prepares and a person finishes. The person reads the extracted fields, checks the flagged mismatch, approves the draft. That is not a failure of ambition. It is the only defensible place to start, because the step has no record yet.

What changes is the record. If a person has confirmed the machine's output four hundred times and corrected it twice, the case for that person reviewing every instance has weakened. A sample, checked on a schedule, would catch the same errors at a fraction of the cost.

The delegation ladder is the line, moving

The delegation ladder sets out seven kinds of work, from the least autonomous to the most. Each rung corresponds to a place on the judgment line.

RungWhere the judgment line sits
1 to 4: finding answers, reading and extracting, checking and comparing, draftingThe machine prepares, and a person finishes
5: following upThe machine finishes, within rules agreed in advance
6: running repeatable processesRules and programs finish on their own, because the exact steps are code
7: working with partnersThe same, with the line running between two organisations

Climbing a rung means moving the line for a step from "a person finishes" to "the machine finishes, and people audit a sample". The original piece gave each step a destination. The ladder says destinations are not permanent.

Two questions, and only one of them moves

The judgment line asks two kinds of question, and they behave differently over time.

Can the machine do it? Has this been done many times in essentially the same way, and can someone check the output faster than they could produce it? This is a question about track record. Evidence changes the answer.

May it act alone? When this goes wrong, does someone outside the company expect a named person to answer for it? This is a question about accountability. Evidence does not change the answer.

So every step has a ceiling. A supplier expects a named person to answer for what was paid, and no record of correct matches changes that. What evidence can change is where the ceiling applies. A finance team that has reviewed a year of exceptions may decide that payments below a threshold, matched exactly to a purchase order, no longer need a signature. The principle holds for everything above it.

This is the point I would most want a reader to keep from this update:

Evidence moves the line for what a machine can do. It never moves the line for what a person must answer for.

A process climbs step by step

A team is not on one rung. Each step of each process is.

Take the invoice approval process from the original piece. Routing is a rule and finished on its own from day one. Reading the invoice is an AI step that starts with a person confirming the fields, and can move to sampled review once the corrections dry up. Matching against the purchase order is a program and finishes on its own. Approving payment stays with a person, except perhaps below a threshold the finance team sets on the strength of the record. Posting to the ledger finishes on its own.

The steps sit on different rungs, and one of them has a ceiling. Asking "which rung are we on" of the whole process hides exactly the information you need.

What it takes to move the line

Three things, for each step, before the line moves:

The line can move back

The ladder suggests one direction. The line moves both ways.

A model is replaced. A new supplier sends invoices in a layout nobody has seen. A regulation changes what a step must check. Each of these resets the record for the steps it touches, and the honest response is to move those steps back down: a person finishes again until the record rebuilds. The sampled audit is how you find out. If the sample starts showing corrections, the step has told you it is time.

A seventh step

The September version ended at writing the line down. It needs a seventh step. Revisit the line on a schedule, step by step, against the record. Move up the steps the evidence supports, up to their ceiling. Move down the steps where something changed. The judgment line is not drawn once. It is maintained.

Which of your steps has a person reviewing every instance, when the record says a sample would do?

← More from the Forge tech blog