Automating compliance review workflows with AI in the loop
Cutting document review cycles by 70% using an LLM-assisted intake and triage system — with human-in-the-loop controls satisfying second-line risk.
The situation
- A 24-person compliance team was spending 60%+ of their week on first-pass document review and routing.
- Backlog had reached 11 days, holding up origination revenue and frustrating front-line partners.
- Prior automation attempts had been blocked by risk and legal due to model explainability and audit trail concerns.
How we engaged
Map
Time-and-motion study across 40 reviewer hours to identify which decisions were repeatable, which required judgement, and where errors originated.
Redesign
Re-cut the process into 'machine-decidable' triage, 'AI-assisted' synthesis, and 'human-only' adjudication — each with its own controls.
Build
Deployed an LLM-assisted intake that extracts, classifies, and pre-fills review artefacts, with full source citation and reviewer override.
Embed
Co-designed audit, sampling, and model monitoring with the second line so the system shipped with risk on side rather than blocking it.
What we built and ran
- An intake pipeline that extracts structured facts and routes by risk tier
- Reviewer cockpit showing AI-suggested findings with source highlights and one-click accept/edit/reject
- An immutable decision log capturing reviewer rationale and any deviation from AI suggestions
- Continuous evaluation harness scoring AI agreement vs. final human decision, by document type
What we'd carry forward
- In regulated workflows, the controls plan is the product. Build it before the model.
- Citations and override paths buy reviewer trust faster than model accuracy claims.
- Treat AI agreement rate as a first-class metric — drift shows up there before it shows up in outcomes.
Want a tailored example for your context?
Tell us about your industry and the outcomes you're after. We'll send a relevant scenario and a recommended path forward.