A JSE-listed broker was reviewing 1,200 onboarding packets a week by hand. We built a document agent that did 78% of it for them — auditably.
The client is a JSE-listed broker offering FX and equity products to South African retail and institutional clients. Like every FSCA-regulated firm, they have to run full KYC on every new account — FICA, source-of-funds, PEP screening, sanctions, the lot.
Their onboarding flow was working. The compliance team behind it was not. Analysts were averaging 17 minutes per packet, 35 packets a day, and the queue kept growing. New hires took 6 weeks to ramp. Two senior analysts had resigned in the previous quarter, citing — verbatim — "the work is killing us."
Two prior vendors had attempted automation. Both failed. The first sold a no-code OCR product that broke on anything that wasn't a perfect bank statement scan. The second pitched a custom build but quoted 7 months and R3.2M — the client passed.
"We don't want a tool that solves 90% of the easy stuff and leaves the 10% that matters. We want something that knows what it doesn't know." — Compliance Lead, kickoff call
Four agents in a Temporal workflow. Each agent owns one thing and outputs structured data with citations. If any agent's confidence falls below threshold, the packet goes to human review — with the agent's notes attached.
Textract + custom layout model. Handles scans, photos, multi-page PDFs.
What is this? ID, proof of address, bank statement, payslip, deed of sale.
Structured fields. Cross-check against application. Run PEP / sanctions API.
Pass / queue for human / hard-reject. Every decision linked to its evidence.
The hardest part wasn't the agent. It was the eval suite. We built it from 600 historical packets the compliance team had already reviewed, weighted toward the edge cases that had previously slipped through. By week 3, every code change ran the suite. By week 5, the system was matching senior-analyst accuracy on every category except source-of-funds — which we intentionally left to humans, because the cost of a false negative there is regulator-level.
"They built in five weeks what our last vendor quoted six months for. And it actually works in production — which after three failed AI projects, was the part we'd stopped believing in." — Head of Operations
If there's a workflow eating >20 hours a week, we can probably build it down by a third within 6 weeks.