New-client checks that took five days now take 1.2
A regulated EU financial services firm of about 400 people chose, tested and launched one use case in a 90-day project.
At a glance
- Sector
- Regulated financial services, EU
- Size
- About 400 people, several EU member states
- What we did
- A 90-day project to choose, test and launch one use case
- Length
- 90 days
- Result
- Median time to onboard a client down 76%, final error rate 0.9%
The problem
The firm serves small businesses and retail clients in several EU member states. Checking a new client's documents (KYC) took about five working days, needed six to eight operations staff, and 12% of applicants dropped out before onboarding was complete.
Over eighteen months, several AI vendors had pitched a solution, and each stopped at the same point. None could say clearly where its product sat in the EU AI Act risk classes, the compliance team would not accept a pilot without documented human oversight, and the firm had no way of weighing return against regulatory risk at the same time.
What we did
Choosing the use case
We looked at twelve candidate uses of AI across operations and compliance, and scored each for feasibility, return and risk class under the EU AI Act. KYC document review came first: high volume, clearly bounded, and classed in our assessment as outside the high-risk categories.
Building the prototype
We connected the firm's existing document management system to an extraction step using generative AI, with no tie to any one vendor. Human review points (Article 14) were in place from the first day, the technical documentation (Article 11) was kept up to date throughout, and every step left an audit trail.
A shadow run, then launch
For four weeks the system ran in shadow mode while staff still reviewed every file, so that confidence thresholds could be set against the firm's own appetite for risk. In production, any extraction with confidence below 92% goes to a person.
How a new client's documents are reviewed
- Client documentsFrom the existing document systemAI extractionNo tie to any one vendorConfidence checkThresholds set in the shadow run
- Straight through92% or aboveTo a personBelow 92%
- Record keptAn audit trail for every step
The results
| Figure | What it measures |
|---|---|
| 76% | reduction in median time to onboard a client, from 5 days to 1.2 |
| 4 FTE | of operations capacity moved to higher-value review |
| 0.9% | final error rate, against a 3.2% baseline |
| €480k | annualised value, against an engagement cost of €90k |
The firm is now applying the same method to two further use cases on its own.
In the client's words
“We had been told for two years that AI would solve our KYC bottleneck. The pilots that came before always died at the compliance gate. This was the first that didn't.”
Delivered as Arqmetrica, the name we worked under before Ontia Labs.
Other work
Industrial manufacturing, Portugal
Two pilots stopped, one use case in production
Read the caseTransport and logistics, Iberia
AI training and shared rules across the group in eight months
Read the case
Start with a 30-minute call
Bring one task that takes your team too long, ideally one that comes round every week.
Book a 30-minute call