INI-006
·Em revisãoAutonomous agent for bank reconciliation
Henrique Vargas · TreasuryAgentic workflowCriticality CriticalAutonomy: Fully autonomous
60Overall
Automatic evaluation
Business Value70
AI Quality58
Security42
Privacy80
Opinion
High-value initiative but autonomy level is inadequate for criticality. We do not recommend fully autonomous mode in a financial system on day one. Part of the problem can be solved by deterministic rules — AI should only handle exceptions.
Aieval qualifies and recommends — the final decision is always human.
Gaps
2- No rollback plan in case of wrong entry
- Agent permissions in ERP not scoped
Risks
3- Fully autonomous agent in critical system — high risk surface
- Misclassification can cause irreversible financial entry
- Prompt injection via statement descriptions
Recommendations
4- Reduce autonomy to 'executes with review' for the first 90 days
- Scope ERP permissions (reconciliation account only)
- Implement volume/value circuit breaker
- Re-evaluate whether a deterministic workflow fits — AI may be unnecessary here
Discovery
Answers collected during intake
- Problem
- Manual reconciliation of 4 daily bank accounts.
- Objective
- Agent that runs reconciliation end-to-end and records in ERP.
- Impacted users
- Treasury (4 people)
- Vendor / model
- LangGraph + GPT-4o · Multi-step agent with tool access
- Data used
- Bank statements, ERP entries
- Integrations
- Banks via Open Finance, SAP, Sharepoint
- KPIs
- % automatic reconciliation · Discrepancies detected
- Estimated cost
- $220,000
- Risk flags
- Uses PII: NoCustomer-facing: NoProduction access: YesExternal sharing: No