Case study · Module 01, screening

−70% officer hands-on time, measured in live production.

A leading Swiss bank. 8 compliance officers, 4.5 weeks, real cases. Their numbers, not ours.

−70%
officer hands-on time
average, their own journals
300+
live screenings
including after formal close
63,019
articles read in full
across 37 languages
−97%
adverse-media noise
removed on a typical case

The facts of the pilot.

Client
a leading Swiss bank · not named on this page
Users
8 compliance officers, in production
Duration
4.5 weeks, 2026 · formal close 26 June 2026
Environment
live production · the bank's own cases
Scope
screening · sanctions, PEP, adverse media
Measurement
the officers' own KPI journals, reconciled against production data

The problem was volume.

Adverse media does not arrive as a tidy list. One case can surface hundreds of articles.

≈238
articles found on a typical case
before anyone reads one
1,993
peak, on a single case
in the pilot

A manual review realistically reads 10 to 20 of them. What stays unread is still risk, and it stays on the file.

What the agents did.

Read every article in full. 63,019 of them, across 37 languages.
Gather
Ranked what mattered, dropped what did not, kept the source on every field.
Structure
No discretionary authority. Approved deterministic rules can complete small cases.
Bounded authority

Exceptions and judgment calls route to a named officer. Every automated or human path is recorded.

A source that fails is reported as incomplete coverage, never as a clean result.

The full position for a model-risk function: AI governance.

An officer reports a gap. The fix goes live.

Every reported issue was resolved while the officers were still using the product.

Reported to live in productionPilot workflow
Reported
An officer reports a gap.

Raised from inside the bank, against live production work.

Minutes to hours
Root-caused, resolved and live.

The correction returned to the officers in the same working window.

Officer report · professional resolution · production fixminutes to hours
210
improvements implemented
during the pilot itself
Minutes to hours
officer-reported gap
to fixed in production
4.5 weeks
continuous work
with the bank's officers

Every conclusion divergence was root-caused, resolved and returned to production during the pilot.

“Whenever we needed a fix, it was shipped extremely fast.” · Compliance officer, pilot journal

Measured in the officers' own journals.

Same dossiers, manual versus Compl.AI. Logged by the officers, reconciled line by line against production data.

875 → 278 min
the same dossiers
manual, then Compl.AI
≈43 h
officer hours returned
over the pilot
≈238 → ≈6
articles per case
found, then surfaced

The 278 minutes include platform run time. The officer's clock kept running while Compl.AI worked, so active officer time alone was lower. We publish no second estimate.

No accuracy claim here. A number from our own test set is not evidence to your model-risk function.

The method is published instead: how −70% was measured.

Small, quick cases can run straight through.

Bank-approved deterministic rules can fully automate low-risk checks.

Anything outside those rules goes to a named officer with the full evidence attached.

Run the same measurement on your cases.

Same dossiers, manual versus Compl.AI, logged by your officers. We run the protocol with you.