Not every step should be automated, and not every automated step should be done by a
model. How to sort them, and where to put the line.
One page, the whole method.
Sort the steps before you pick a tool
Map the process first, then sort each step. Starting from a platform tends to produce
automation of whatever was easiest to automate.
One question sorts most steps. The expensive mistake is putting a model on a deterministic
step because the platform makes it easy: slower, priced per call, and it adds variance to the
one part of the process that had none.
The threshold
For judgement steps the design is not automated or manual. It is a line, with everything below
it going to a person.
The threshold only means something if the confidence figure is calibrated against known
outcomes. A number a model emits about its own certainty is a dial with no scale on it.
Setting it, then moving it
01
Shadow period
Everything goes to review. Log the reviewer's decision next to the system's.
02
Read the bands
Agreement rate by confidence band, once there is enough volume.
03
Set the line
Where disagreement falls below what the process tolerates. A business call.
04
Move it on evidence
Sustained agreement, unchanged work shape, a named approver and a date.
Make it easy to lower again
If moving the threshold back is expensive or embarrassing, people will argue for keeping it
where it is long after the evidence has changed. Treat lowering it as a normal operation.
Reading the queue
What you see
What it means
Reviewers agree almost every time
Too low. You are paying attention where it earns nothing.
Reviewers overturn a meaningful share
Too high, or the model is not good enough for this step yet. Lower it and find out which.
The queue is never empty
Set for a review capacity you do not have. Approvals will become automatic in practice unless somebody changes it.
Nobody can say what it currently is
The control is not being managed. Worth checking before you rely on it in a review.
What every automated step leaves behind
The inputBy reference, not by copy.
The decisionAnd which implementation made it.
ConfidenceWhere the step was a judgement step.
Threshold crossingSo you know whether a person saw it.
The reviewerWho, when, and what they did.
Enough workingTo explain the outcome. Not so much that the record is itself a data protection problem.
Sorting the steps, and setting the review threshold.
Sort every step by one question
Kind
Recognise it by
Implement with
Deterministic
The rule can be written down. Same input, same correct output.
Plain software. No model.
Judgement
A right answer exists, but reaching it needs reading, matching or interpreting.
Agent, with a verifier and a confidence output.
Accountability
Someone must be answerable. A regulator would ask who decided.
A person. Automation prepares, never decides.
Setting the threshold
01
Shadow period
Everything to review. Log both decisions.
02
Read the bands
Agreement rate by confidence band.
03
Set the line
Where disagreement falls below tolerance.
04
Move on evidence
Named approver, a date, and a cheap way back.
The queue is telling you
Reviewers almost always agree: too low.
Reviewers overturn a real share: too high, or the model is not ready.
Never empty: set for capacity you do not have.
Nobody knows the current value: not being managed.
Every step records
Input by reference, the decision, which implementation made it
Confidence, and whether it crossed the threshold
If a person saw it: who, when, what they did
Enough working to explain it, no more
Never automate
A decision a regulation makes a named person accountable for. The final clearing of an
exception the system itself raised. Changes to the threshold. Those are control changes
and they belong to a person.
Greyquill Software builds governed automation for processes that have to
answer for themselves. greyquill.io
