# Stop Making Your Customers Click Approve - Emil Chetty, Wise

## Executive summary

The presentation argues that current AI assistants force users into an 'approval queue' by requiring manual sign-off for every step. The proposed solution, demonstrated using Wise's payment infrastructure, is to shift the authorization model from step-by-step approval to a single, immutable, locked agreement. This allows the agent to plan and execute the job autonomously within defined bounds, while the underlying trusted software stack enforces the terms, mitigating risks like prompt injection.

## Key takeaways

- The Problem: The Approval Queue: AI assistants often force users to manually approve every action (e.g., scraping, reading documents), turning the user into the 'API' and undermining the promise of reduced complexity. This creates a constant need for human intervention.
- The Solution: Intent to Locked Agreement: Instead of approving steps, the customer agrees to the overall 'job' or intent. This agreement is verified by trusted software and becomes an immutable, machine-readable promise that governs the agent's operation.
- Enforcement and Resilience: The system's core strength is that the downstream services (e.g., payment services) enforce the agreed-upon bounds, regardless of the agent's internal state or if the agent is tricked by a prompt injection. The software stack, not the model, provides the security boundary.

## Technical details

- Agent Workflow Design: The process involves the agent exploring with limited read-only access to gather context, which is then surfaced to the user as a draft job. The user approves the job, making it immutable and establishing the operational bounds.
- System Architecture: The execution phase utilizes a trusted provisioning platform that checks the intersection of the confirmed job, Wise's permissions, and the customer's current rights. Execution is split into specialized workers (e.g., a payment worker and a reconciliation worker) to maintain separation of concerns.
- Security and Compliance: The system relies on a robust, long-standing software stack (since 2011) to enforce the agreed-upon terms. This ensures that even if the LLM is compromised (e.g., via prompt injection), the downstream services prevent unauthorized actions.

## Practical implications

- When designing agent workflows, focus on defining the scope of the job (the 'intent') for customer approval, rather than listing every possible step.
- Build robust, downstream services that act as the ultimate authority, enforcing the agreed-upon state and preventing unauthorized actions, even if the AI layer fails or is compromised.
- Implement clear separation between the agent's planning/suggestion layer and the execution/enforcement layer (e.g., using dedicated payment and reconciliation workers).

## Topics

AI Agents, Financial Technology (FinTech), Workflow Automation, Immutable Records, Security Architecture, Prompt Injection Mitigation, Emil Chetty, AIE CODE summit

Source: https://www.youtube.com/watch?v=V2UqEUA8NjM
