Private banking onboarding: from twelve days to three
Document AI reads the pack, verifies what it can, and assembles the file — while every screening judgment and every acceptance decision stays with a human.
The problem
Onboarding a private banking relationship took a median of twelve days and, in complex structures, considerably longer. The delay was almost entirely document handling: certified identity documents, source-of-wealth evidence, trust deeds, corporate structure charts, tax residency declarations — collected by email, checked by hand, chased when incomplete, and re-checked when a version changed.
The cost was not only operational. Onboarding friction is where wealth relationships are lost, and the people most likely to abandon the process were exactly the clients with the most complex — and most valuable — affairs.
The hypothesis
Two distinct activities were tangled together: evidence assembly, which is mechanical, and risk judgment, which is regulated and consequential. Automating the first without touching the second would compress the timeline dramatically while leaving the control environment intact.
The build
- Intake and classification — inbound documents identified by type and jurisdiction, with extraction models pulling names, dates, identifiers, addresses, and ownership percentages, returning confidence per field.
- Cross-document consistency checks — the same entity's details reconciled across every document in the pack, with mismatches surfaced explicitly rather than resolved silently.
- Structure interpretation, drafted — for trusts and corporate holdings, the system drafts the ownership chart and proposes who the beneficial owners appear to be, citing the clause in the deed. An analyst confirms or corrects it.
- Gap detection with drafted requests — missing or expired items are identified up front and a single consolidated request is drafted, instead of the three separate chases that used to happen.
- A hard boundary at risk decisions — screening hits, source-of-wealth adequacy, politically-exposed-person determinations, and final acceptance are all human. The system assembles evidence and does not score risk.
Design choice that mattered: we deliberately declined to build a risk-scoring model. Financial crime controls are examined, and "an algorithm assessed the source of wealth" is not a position a bank wants to defend. Assembly automated, judgment untouched — that division is why compliance sponsored the project rather than tolerating it.
Rollout
We ran parallel processing on live cases for six weeks, comparing extracted fields against analyst-verified values field by field, and set thresholds per field type — an identifier requires near-certainty, a stated occupation does not. Complex trust structures were excluded from automation entirely until accuracy on the simpler tiers had been demonstrated for a full quarter.
Results
Median onboarding fell from twelve days to three, most documents cleared verification without a human touch, and analysts spent their time on structure interpretation and risk judgment rather than data entry. Client abandonment during onboarding dropped, which was the outcome the business actually cared about.
What we'd tell you
- Separate assembly from judgment, then automate only the first. This is the whole design in regulated onboarding.
- Confidence per field, thresholds per field type. A single document-level score is too blunt.
- Surface inconsistencies rather than resolving them. Silent reconciliation destroys the audit trail.
- Exclude the complex tier until you have earned it, and say so plainly to your regulator.
More from the field.
Consumer lending, decided in hours
Early-warning credit risk
The 6 a.m. advisor brief
Bring us a hypothesis. Leave with a system.
Tell us what's eating your team's time. We'll give you an honest read on whether AI is the right tool — and if it is, a scoped v1 with a timeline and cost.