Cutting a week-long quality check to two days

Situation

One of the Quality Check processes was conducted to catch inconsistencies, and it was almost entirely manual. Analysts would work through a batch of documents, read the extracted data, compare it against the guidelines, flag what didn’t match, and correct it. Two or three senior analysts would lose the better part of a week to it each time.

This needed experienced people. What it didn’t need was for those people to spend most of the week on the manual part before reaching anything that required a decision.

Constraints

AI could not run the process. Client data governance ruled out anything agentic, anything operating on data unsupervised, and anything that moved it outside the systems already approved for it. Whatever I built had to work inside the tools the team already had.

There was also no budget and no extra time. This was built alongside normal delivery, with no new licences and no additional people.

What I did

1. Mapped what was actually happening. Steps, methods, tools, time spent at each stage. The bottleneck wasn’t the judgement calls - it was the manual check, and that was most of the week.
2. Designed the logic. I knew this process well enough to write down the rules the checks were really applying: what counts as an inconsistency, what to compare against what, what to flag. I described that logic to Claude with the data anonymised, since the real data couldn’t go near an AI tool.
3. Built it inside Excel and checked it. Claude wrote the macros and scripts from those rules. I assigned them to buttons so an analyst runs the check by clicking, with nothing operating unsupervised and no data leaving where it belonged. Then I ran it in parallel with the manual check on the same documents. It was faster, and it caught details a person reading for hours would miss.
4. Left the decisions where they belonged. The automation flags; the analyst decides whether something should be corrected. That was deliberate - it’s the part of the check that needs judgement, and it’s the part the constraints wouldn’t have allowed me to automate anyway.
5. Handed it over properly. I prepared instructions, a presentation, and a training session for the team.

Result

A full check went from a week to two days. The bigger change was where the analysts’ time went: from reading data to checking results - the part of the work that needs their judgement rather than their patience.

Accuracy went up as well. When I ran the automation alongside the manual check on the same documents, it caught inconsistencies a person reading for hours would miss.