For developers building AI agents that touch production data, the hardest part isn't making the model smart enough to workβit's making it safe enough to trust. Australian accounting software providers Fedix and MYLEDGER have tackled this by making human approval the primary design constraint for their AI assistants. In a recent post on DEV.to, the team detailed how their MYLEDGER Copilot and FediPilot desktop assistant operate under a strict 'propose, then approve' workflow, ensuring that no change lands in a client's ledger or tax return without a human explicitly signing off on it.
The Boring Architecture of Trust
The core mechanism is what the team calls a 'boring design': a queue of pending changes held behind a human key. When the MYLEDGER Copilot or a connected assistant requests a modification, the change waits for a person to approve or reject it. This same logic applies to integrations with Xero, where every potential write action from Fedix must be confirmed via a dialog before execution. For engineers, this is a standard pattern of asynchronous job queues with manual triggers; for accountants, it eliminates the risk of an AI confidently rewriting financial data based on a hallucination.
Context-Aware Permissions
A frequent pitfall in AI tool integration is the use of shared service accounts with broad, super-user permissions. Fedix avoids this by tying the FediPilot desktop assistant to the specific user who signed in. The assistant inherits the user's permissions exactly, meaning if a staff member cannot view a specific client file, the AI cannot either. Additionally, desktop actions are bound to enrolled devices, and sensitive operations require further approval. This approach simplifies auditing because the AI acts as the user, not as a separate entity with wider access rights that need their own permission matrix.
Drafts, Not Final Products
The team is explicit that AI outputs in Fedix are treated as first drafts, particularly for complex tasks like tax return preparation. The AI works through client documents to generate a draft, which then enters the accountant's professional review process. This distinction is crucial for liability and quality control: the AI accelerates the drafting phase but does not replace the professional judgement required to finalize a return. The system is designed to start the workflow, not end it, keeping the human in the loop for the final decision.
Auditability and Rollback
Approval gates are only useful if the history of those approvals is transparent and immutable. MYLEDGER records sensitive actions centrally, categorizing them and applying retention policies to create a clear audit trail. This ensures that when a reviewer or client asks who approved a change and when, the answer is a system record rather than a human recollection. Furthermore, the system includes a safety net for bulk changes, allowing users to preview a plan to roll a client's ledger back to an earlier time, apply it, and undo it if necessary.
Key Takeaways
- Human-in-the-loop approval gates prevent AI from making irreversible changes to client data without explicit consent.
- AI assistants should inherit user permissions rather than using shared service accounts with broad access to simplify auditing.
- AI-generated outputs in professional software should be framed as drafts that require human review, not final deliverables.
- Robust audit logs and time-travel rollback features are essential safety nets for any AI system modifying production records.
The Bottom Line
The 'boring' design of human approval gates is actually the most critical feature for AI in regulated industries. Trust isn't built on model intelligence, but on the assurance that a human always holds the final key.