The situation. Onboarding a single family meant
reading trust deeds, articles, share registers, foreign registry
extracts and source-of-wealth evidence by hand, then screening whoever
that reading turned up. Standard investment files took two committee
cycles to clear. The incumbent screening process had no threshold
calibration, no entity resolution, and matched names by eye. Vendor
tools were evaluated first: none resolved trust structures (settlors,
trustees, beneficiaries, controllers and underlying corporate
ownership) into a single screened entity per client structure, which
was the firm's core requirement. So it was built in-house.
What was built. A document pipeline that returns
structured fields with a confidence score, each routed to a compliance
reviewer, because KYC requires a person to sign off every field.
Retrieval happens at page level before extraction rather than pushing
whole documents at the model, which is where most of the accuracy came
from. The same index answers questions over the deed corpus in natural
language and cites the source clause. On top of it, each family's
holding structure is reconstructed as a graph — settlor, trustee,
protector, beneficiaries, underlying companies, Companies House PSC
data — and resolved into one screened entity set. The PSC register is
treated as the unverified self-reported source it is: corroboration
rules, documented failure modes, and no direct-owner entry accepted as
a beneficial owner on its own.
Calibration, not guesswork. The engine screens
against UK and OFAC sanctions data, PEP lists (OpenSanctions) and the
Companies House PSC API. Match thresholds were set by above-the-line /
below-the-line testing against the firm's own logged history: roughly 3,000
client and counterparty checks spanning two years, replayed and scored
against the incumbent process. The recall condition was fixed up front
— every standing PEP and sanctions match on the client book preserved,
and no alert previously escalated by compliance allowed to auto-clear.
Independently validated before go-live.
Where automation stops. Agreed in the same forum as
the thresholds and documented for internal compliance sign-off and
external audit: suspicious activity reporting kept out of scope
entirely; beneficiary distributions, tax and succession excluded;
low-confidence extractions and rejections go to mandatory human review
and are never auto-declined. Record-keeping meets MLR 2017 and HMRC
Trust Registration Service duties.
The result. Alerts requiring manual review down by
about a third. Median file-open-to-decision time roughly halved since
go-live, with standard investment files clearing in one committee cycle
instead of two. The pipeline was triaged into the compliance workflow
already in use rather than run alongside it, with runbooks written and
the internal team trained; it now drives the annual review of the
existing book as well as new files.
