.png)
Exceptions Dominate Execution
A sales order processing operation handles forty inbound documents per day — unstructured emails, attachments, mixed formats, variable field positions. Before automation, every document required manual handling: reading, interpreting, and keying data into the system. Automated ingestion now extracts line items, quantities, delivery addresses, and pricing. Thirty-eight orders process without error. Two do not.
The team reports a 5% error rate and cannot understand why labour hasn't been removed.
The answer is not in the 5%. It is in what the 5% does to the other 95%.
At a 5% error rate, the system cannot identify which orders are clean. The two errors per day are not labelled. They are distributed invisibly across forty outputs, indistinguishable from the thirty-eight correct ones until a human checks. The result is that every order must be reviewed — not to process it, but to confirm that automated processing was safe. Automation has removed the data entry labour from thirty-eight documents per day and replaced it with verification labour across all forty. In most operations, verification is not materially faster than the original processing it replaced, because the reviewer must understand the document well enough to confirm the system's interpretation of it.
The 95% success rate has not resulted in a 95% labour reduction. It has shifted the labour from processing to checking, and applied it to every order in the stack.
The trust threshold for an automated system is not lower than for a human. It is higher. When a person makes an error processing a sales order, it is absorbed as a normal operational event — caught, corrected, moved on from. When a system makes the same error, it is treated as a system failure, because the assumption is that machines either work or they don't. A 5% human error rate is tolerated and managed. A 5% machine error rate produces a decision to check everything, because the errors are unpredictable, the system cannot explain them, and acting on a wrong order without knowing it is wrong carries consequences that most operations will not accept.
This means the labour reduction that automation promises is not unlocked gradually as accuracy improves. It is unlocked at the point where the output can be trusted without verification — and that point is close to zero errors, not somewhere on a sliding scale. Reducing the error rate from 5% to 2% does not remove 60% of the checking labour. It still produces full review, because any invisible errors in a live order processing environment cannot be tolerated.
The problem is that accuracy and trustworthiness are not the same thing. A system can get 95% of orders right and still provide no basis for knowing which 95%. Without that signal, the only safe response is to check everything — and a more accurate system that still cannot identify its own uncertain outputs does not change that calculus. It just makes the errors rarer and harder to anticipate.
Before automation, forty documents per day required manual processing. That labour was visible and proportional to volume. Automation absorbed the processing work from the thirty-eight routine cases. What it introduced, at a 5% error rate, was verification labour across all forty.
Scale the operation to four hundred documents per day and the same dynamic holds. The verification workload grows by a factor of ten, applied to the full population, driven not by how many errors occur but by the fact that the system cannot tell you which outputs are safe.
Checking every order in a system that gets 95% of them right is not a quality control problem. It is a trust problem.
What actually unlocks labour removal is simpler: can the system tell you which orders are safe to act on, and can you trust that signal? If it can't, you check everything.
You will keep checking everything regardless of whether the error rate is 5% or 1%, because the trade-off is binary: accept undetected execution errors, or retain full verification labour. The way out is not more accuracy. It is a question of whether the system can tell you which orders are safe — and until it can, verification stays universal.