A fintech client came to us with a problem: their compliance team was spending 20+ hours a week manually reviewing transaction reports for regulatory flags. They knew AI could help, but three months of internal discussions hadn’t produced a clear plan.
We scoped the entire project in one week. Here’s how.
Day 1: Understand the actual work
We didn’t start with a requirements document. We started by watching.
We sat with two compliance analysts and observed their process end-to-end. Not what they said they did — what they actually did. The difference matters.
What we found:
- They reviewed ~500 transactions per day
- 85% were clearly fine and took 10 seconds each
- 10% needed a quick lookup in an internal database
- 5% required genuine judgment and sometimes escalation
The first 85% was the opportunity. If we could auto-clear those, the team could focus entirely on the cases that actually needed human attention.
Day 2: Map the data
We spent a full day understanding the data: where it came from, what format it was in, what fields mattered, and what the rules were for flagging.
The compliance team had an informal checklist in their heads — patterns they’d learned to watch for over years. We turned that tacit knowledge into explicit rules:
- 12 hard rules (regulatory requirements, automatic flags)
- 8 soft signals (patterns that warranted a closer look)
- 3 escalation triggers (things that always needed a senior reviewer)
This became the spec for what the AI system needed to learn.
Day 3: Evaluate approaches
With the problem mapped and the data understood, we evaluated three approaches:
- Rule-based system — encode the 12 hard rules as code, flag everything else for review
- ML classification — train a model on historical decisions to predict clear/review/escalate
- Hybrid — hard rules for the obvious cases, ML for the gray area
We chose the hybrid. Pure rules would miss the soft signals. Pure ML wouldn’t guarantee regulatory compliance on the hard rules. The hybrid gave us both.
Day 4: Prototype
We built a working prototype in a single day. Not production-ready — but functional enough to run against real historical data and see results.
The prototype:
- Processed a batch of 500 historical transactions
- Applied the 12 hard rules
- Used a simple classifier for the soft signals
- Produced a report showing: auto-cleared, flagged for review, escalated
We ran it against three months of historical data where we knew the correct outcomes.
Day 5: Validate and plan
Results from the prototype:
- Accuracy on auto-clear decisions: 97.3%
- False negatives (missed flags): 0.4% — within acceptable range with human review on flagged items
- Estimated time savings: 16 hours per week
We presented the findings to the compliance team and leadership with a clear proposal:
- Phase 1 (2 weeks): Harden the prototype, add audit logging, integrate with their transaction feed
- Phase 2 (2 weeks): Run in “shadow mode” alongside human reviewers to validate accuracy
- Phase 3 (1 week): Go live with human-in-the-loop for flagged items
Total timeline: 5 weeks from scoping to production. Total cost: a fraction of what three months of committee meetings had produced in “strategy.”
Why this worked
A few things made the one-week scope possible:
We watched the work. Not a presentation about the work — the actual work. This eliminated assumptions and revealed the real process.
We focused on the 85%, not the 5%. The temptation is to solve the hardest cases first. But the ROI is in the volume — automating the easy decisions frees up time for the hard ones.
We built to validate, not to ship. The day-4 prototype was rough. But it answered the only question that mattered: can this work?
We had decision-makers in the room. No waiting two weeks for feedback loops. The compliance lead and a VP were available every day for quick decisions.
The takeaway
Long scoping cycles usually mean one of two things: the team doesn’t understand the problem well enough, or they’re afraid to commit to an approach. Both are solved by getting closer to the actual work and building something small to test your assumptions.
If you’ve got a project that’s been stuck in the “planning” phase, let’s talk. We might be able to unstick it faster than you think.