They asked for a script. The problem was judgment.
The situation
A global travel marketplace was onboarding lodging supply by hand: three hours of skilled work per property, running at 4,000 cases a month. New hires took four to six months to become productive because the process lived in people's heads. And the errors were not clerical: a single manual mistake had cost $400K, and the company kept breaching the $500K annual ceiling it had set for error losses. Behind the queue sat a backlog of internal projects, contract reviews and audits worth seven figures, that the team never reached.
The diagnosis
They arrived with the solution already scoped: a simple RPA bot and a basic regex to map source files to data fields, aiming to cut the three hours in half. The diagnosis said that mapping fields was a tiny slice of the work. Most of those three hours were judgment: reconciling records across multiple cloud platforms and homegrown tools that disagreed with each other, and validating the result against strict business rules. A regex cannot referee a disagreement between systems. RPA would have automated the typing and kept all of the thinking, which means halving the hours while preserving every dollar of the error risk, the $400K kind included.
The prescription
Automate decisions, not keystrokes. Prove the workflow manually first, then replace it stage by stage, with hard acceptance gates between semi-automated, production, and full production, defined in writing before anything was built.
The build
An integration layer connecting the cloud platforms and homegrown systems that had never agreed, over a hardened ETL pipeline. A trigger in Salesforce so a new case launches the pipeline on its own, with no human dispatcher. LLM-driven fuzzy matching to reconcile the same property across systems, concatenating fragments from multiple sources into single validated fields, parsing free text into more than a dozen structured fields, and validating every output against the strict rules that used to live in people's heads. Automated QA on top, then rollout across the top 15 supply partners and a long tail of 130 more, with uptime and error-response SLAs the client could audit.
What the team kept
A cycle that went from three hours to fifteen minutes, roughly eleven thousand hours a month returned to the business. The class of error that produced the $400K incident is now caught by validation before it ships, keeping losses inside the ceiling the team used to breach. Ramp time stopped mattering because the judgment no longer lives in anyone's head. And the seven-figure backlog finally has a team available to work it.