Nobody Could Say What Percentage We Checked.

The question that stopped the room was not about AI. It was: what percentage of your carrier invoices do you actually audit?

Nobody knew. Not the finance lead, not the operations manager, not the five people whose job it was. They knew how many hours they spent. They knew roughly what they recovered. They had never expressed the work as a fraction of the work available.

The answer turned out to be about five percent. Five people, roughly three weeks a month, reviewing five percent of the invoices from a business shipping more than half a million packages a year. They recovered around $50,000 annually and considered that a functioning process. In fairness, it was. Five percent was what five people could physically get through.

But the sample rate had quietly stopped being a decision. It started as a fact about staffing and became a fact about the business, and at some point everyone stopped seeing it.

The baseline nobody writes down

Every automation proposal I read includes a target. Very few include a baseline anyone can defend.

That is the actual reason so many AI projects produce enthusiasm and no provable value. Not because the technology underdelivered, but because nobody established, in advance and in writing, what the current state was. Without that, improvement and activity look identical on a slide. Effort gets shifted rather than removed and reported as a gain, because there was never a number to argue with.

And the baseline that goes missing most often is not speed. Everyone measures cycle time — it is visible and easy to instrument. The one that disappears is coverage: what fraction of the eligible population gets examined at all.

Five numbers worth having before you automate anything

Not a framework, just the short list that has repeatedly turned out to matter:

• Coverage. What share of eligible items is actually examined today. If nobody can answer this, stop and find out — it is usually the largest number in the room.
• Cycle time. How long from item arriving to item resolved, measured end to end rather than in the fast middle section.
• Cost per transaction. Fully loaded, including the review nobody counts as work.
• Error and rework rate. How often something has to be done twice, and what triggers it.
• The deadline. Almost every reconciliation process has a window after which acting is worthless — a dispute period, a filing date, a claims window. A finding that arrives after it closes has no value regardless of how correct it is. Teams that measure only accuracy and never timeliness routinely automate their way into producing correct findings too late to use.

What happened when we measured first

With the baseline established, the case for full coverage made itself, and the results were arguable against a real prior number rather than a feeling: more than $400,000 recovered, an 800% return in the first year, audit workload down 80%, dispute resolution from three weeks to two hours, and a 95% reduction in recurring errors as findings started flowing back to the carriers and fixing the source.

Those numbers mean something specifically because the $50,000 and the five percent were written down first. Had we skipped that, every one of them would have been a claim rather than a measurement.

The one question

Before approving any automation, ask what percentage of the work is checked today.

If nobody can answer, you have not found a gap in your reporting. You have found the business case.

Find out what percentage you actually check. We help teams put a defensible number on coverage, cycle time, cost per transaction, error rate and the deadline that governs them — before anything gets automated against it. Book a baseline review.