← Insights

Method / 7 min read

What to automate first, and how to know you are wrong

The first automation decision is rarely a technology decision. It is a decision about which mistakes you can afford to make.

The loudest process is rarely the right one

Ask any operations team which process to automate and you will be pointed at the one that generates the most complaints. Complaint volume measures friction, not opportunity. The process everyone hates is often hated precisely because it is full of exceptions — which is the property that makes it hardest to automate well.

A better first question: which process is high in volume, low in judgement, and cheap to get wrong? That intersection is small, unglamorous, and where the first win almost always lives.

Score on three axes, and weight the third

Volume tells you the size of the prize. Error cost tells you what a bad output does downstream. Reversibility tells you whether a bad output can be quietly corrected or whether it reaches a customer, a regulator, or a ledger.

Reversibility deserves the heaviest weight. A reversible process with a moderate error rate is a good candidate; an irreversible process with an excellent error rate is not, until a review gate makes it reversible again. This is the single distinction that separates automation programmes that compound from those that get switched off after one incident.

Design the gate before the model

Human-in-the-loop is often treated as a fallback. It is better treated as the design: define the confidence threshold below which a human must approve, define who that human is, and define how long they have. Then measure how often the gate fires.

A gate that never fires means the threshold is wrong or the reviewer has stopped reading. A gate that always fires means you have automated nothing and added a queue. The healthy range is visible within two weeks of a pilot.

The tests that tell you the choice was wrong

Run the pilot against a manual baseline you measured beforehand — cycle time, error rate, cost per item. If you did not measure the baseline, the pilot cannot succeed or fail; it can only be argued about.

Three signals mean stop and re-scope: exceptions exceed roughly a fifth of volume, the review queue grows faster than it clears, or the team quietly builds a spreadsheet to work around the new system. The last one is the most reliable indicator in this list and the least often reported upward.

If this is the kind of thinking you want applied to your own operation, a first conversation costs nothing.

Start a conversation