Amol Kapoor and Clifford Ressel of Nori Agentic explain a human-reviewed data-cleaning workflow in this 20-minute AI Engineer talk.
Give the database an investigation queue
- A security master is the shared record of an asset manager’s securities, feeding risk, accounting, and compliance work.
- Private investments arrive through contracts, amendments, PDFs, and email, often without common identifiers.
- Compare yesterday’s records with today’s, then investigate each changed cell before downstream analysis depends on it.
Assemble the evidence before asking for approval
- Nori describes a Fortune 200 asset manager where each changed cell previously required about an hour of manual investigation.
- An agent searches Box documents, financial-data APIs, and a limited email collection to explain the change with citations.
- It recommends keeping or discarding the change and sends the responsible person a Slack review request.
- Kapoor says this reduces the human portion to roughly five minutes; the talk gives no measured error rate or sample size.
Reuse the runtime, customize the context
- The described deployment uses an on-premises open-source model, despite the broader “cloud agent” framing.
- Instructions identify document locations, trusted sources, and criteria for accepting a change.
- Skills, configuration files, credentials, and triggers adapt a general agent to the task.
- The same underlying infrastructure can support audits, code review, and other back-office investigations.
Keep a person accountable
- Kapoor recommends aiming for substantial automation while retaining a short human review, rather than chasing unattended completion.
- His “95% automation” is a design goal, separate from the reported one-hour-to-five-minute human-effort example.
- Domain experts supply the rules and judgment; agents absorb the repetitive work of reconciling formats and finding evidence.
“AI adoption is primarily a collaboration problem.” — Amol Kapoor, 17:27