Document automation works when you stop trying to cover everything

Volume is never evenly distributed

Ask a team which documents eat their time and you get a list of everything they handle. Measure it and you find a small number of types carrying most of the volume, and a long tail of formats that appear a handful of times a year.

That distribution is the single most useful thing an assessment produces here, and almost nobody has measured it before we arrive. The perception of the problem is shaped by how annoying a document is, not by how often it appears.

Scope to the concentration and a project becomes tractable. Scope to the perception and it becomes a platform.

The long tail is where accuracy targets go to die

High-volume documents are usually structured. Same layout, same fields, predictable variation, which is exactly the condition under which extraction is reliable.

The tail is the opposite. Unusual formats, unstructured correspondence, one-off documents from a party who sends one a year. Extraction on those is unreliable, and the errors are hard to spot because there is no pattern to check against.

Setting one accuracy target across both is how projects promise something they cannot deliver. Accuracy targets belong per document type.

Confidence thresholds beat blanket automation

A well-built extraction system knows when it is unsure. Using that signal to route uncertain documents to a person is the difference between a system that quietly corrupts data and one people trust.

The threshold is a business decision. Set it high and more work reaches humans but errors are rare. Set it low and throughput rises along with risk. It should be tunable and revisited once you have real numbers.

Systems that force a confident answer for every input are the ones that eventually produce a costly mistake nobody catches.

Deciding not to automate is a deliverable

The hardest part is writing down what stays manual. It feels like an admission of failure, and teams often relitigate it repeatedly because the decision was never recorded with its reasoning.

Record it explicitly: this document type stays manual, because the volume is low and the variation is high and an error would be expensive. Then it can be revisited when volume changes, rather than argued about every quarter.

That section is regularly the most referenced part of an assessment, because it settles arguments that had been running for a year.