Human-reviewed transaction checks for a fintech operator

The challenge

Transaction checks and reconciliation were sensitive, repetitive, and hard to automate safely. The team needed help gathering evidence and preparing review work, but not an unconstrained agent making compliance decisions on its own.

What Grail built

Grail built a review workflow that gathers transaction and AML/KYT evidence, applies clear scoring rules, stores the trail, and stages decisions for human approval before any writeback.

Stack used

TemporalPostgresDockerTypeScriptAML/KYT APIsApproval gates

Impact

Reviewers get a consolidated evidence packet and approval step before sensitive reconciliation or compliance actions move toward writeback.

How the workflow runs

We kept the compliance workflow rule-first. The agent prepares context and explanations, while deterministic checks and human review control the final step.

1. The workflow gathers transaction and compliance payloads.
2. Rules score the case and collect the supporting evidence.
3. The agent prepares a summary for the reviewer.
4. A human approves or rejects the action.
5. Only approved outcomes move toward system writeback.

Human control

The control points were specific to the workflow, so the agent could speed up the work without silently taking over sensitive decisions.

• No unreviewed compliance writebacks.
• Rules own the decision path; agents explain and summarize.
• Reviewer approval is required for sensitive actions.

What shipped

The implementation centered on these shipped pieces:

• Transaction check workflow.
• Reconciliation workflow.
• Reviewer console shape.
• Evidence and audit trail storage.
• Backtesting and shadow-mode path.
• Approval gates for writeback.