The situation
A compliance-heavy financial product for small businesses had been built by accountants, for accountants. Every case moved through fourteen steps and took roughly two weeks, processed by hand by a 26-person operations team — reading documents, categorising expenses, running eligibility questionnaires and drafting the final filing. It worked, but it couldn't scale: headcount rose in lockstep with volume. Conversion sat at 7–8% and NPS was in single digits.
Why it was hard
Automating a workflow is straightforward when nobody gets hurt if it's wrong. This wasn't that. Every output was a regulatory filing — a false positive meant a customer claiming something they weren't entitled to, with a professional's name attached. The compliance team's objection wasn't obstruction; it was correct. Any credible answer had to hold up against a regulator, not just a demo.
What we built
The redesign came out of a five-day sprint, with compliance in the room from day one. Three friction points carried most of the cost, and each got a different treatment:
- Eligibility — a manual questionnaire became public web data scraped to pre-fill answers, which the customer approves or edits.
- Eligibility review — roughly a week of human review per case became a RAG system over historical filings and regulatory guidance: instant.
- Documents — manual upload and categorisation became OCR plus a classification model.
- Final report — written by hand, now generated as a draft and reviewed by compliance before delivery.
- Tier-1 support — a human queue, now an assistant handling routine queries at ~40% deflection.
Fourteen steps became seven; time to value went from two weeks to two days.
How we made it safe to ship
We set the bar against the human baseline, not perfection. Working with external counsel, we established a false-positive threshold under 2%, benchmarked against the published error rates of top-four accounting firms. Automated eligibility decisions landed at 88% alignment with human reviewers; anything below a confidence threshold escalated to a person automatically, and only high-confidence cases were ever surfaced to customers.
We also took liability off individuals — resolving the compliance team's personal exposure structurally, with counsel, and positioning the professionals as partners running quality control rather than the thing being replaced.
What we deliberately left alone
Compliance review of every generated filing stayed human. Edge cases escalated to a person. Relationship management stayed with the professionals who owned it. One feature we built — automated drafting of customers' outbound emails — was rejected by customers outright, so we removed it. Knowing which parts of a workflow people actually want automated is part of the job.
Results
- Workflow steps: 14 → 7
- Time to value: 2 weeks → 2 days
- Operations headcount: 26 → 10, at twice the case volume
- Conversion: 7–8% → 12–17%
- Retention: 30% → 60%
- NPS: single digits → 65
The operations reduction came from two decisions, and it's worth being precise: automation carried most of it, and a deliberate narrowing of the customer segment carried the rest.