Letting an AI near client files: the approval gate we built first

Fedix built human-in-the-loop approval gates into its Fedix and MYLEDGER accounting software for Australian accountants, so that AI-proposed changes to client records and Xero data wait for a person to approve or reject them before taking effect. The company also scopes its FediPilot desktop assistant to the signed-in user's existing permissions rather than a shared service account, ties desktop AI actions to enrolled devices, and logs sensitive actions centrally for audit. The Copilot can draft tax return preparation for an accountant to review, and MYLEDGER supports previewing, applying and undoing ledger rollbacks.

If you have ever watched an AI tool confidently rewrite something it should have asked about first, you already know the real question for a practice: not "can it do the work?" but "who signs off on the change?" We build Fedix and MYLEDGER for Australian accountants, and we made that question the first design decision rather than the last. This post walks through the guardrails we think any AI assistant near client records should have, and how ours behave. It is general information about our software, not advice about your obligations. The basic pattern is a gate. When the MYLEDGER Copilot or a connected assistant asks for a change, that change waits for a person to approve or reject it. The AI proposes; your team decides. feature:acb-copilot-write-gating The same idea applies to Xero. Every change Fedix would make in Xero is confirmed by a person in a confirmation dialog first. feature:crm-xero-write-confirm For an engineer, this is a boring design: a queue of pending changes and a human-held key. For an accountant, boring is the point. It means a draft stays a draft until someone with the right judgement looks at it. A common shortcut in tool integrations is a shared service account with broad access. We went the other way. The FediPilot desktop assistant uses Fedix as the person who signed in, with the same permissions. feature:crm-mcp-user-jwt So if a staff member cannot see a client's file, the assistant working on their behalf cannot either. There is no separate, wider permission set to audit. Desktop actions are also tied to a specific device: AI actions from the desktop are tied to an enrolled device, and sensitive actions need approval. feature:crm-mcp-fedix-desktop The most common worry we hear is about tax returns. In Fedix, an AI assistant works through the client's documents and produces a first draft of the return preparation for an accountant to review. feature:crm-tax-return-agent-prepare The word that matters there is "review". The draft is the start of your professional process, not a replacement for it. Approvals are only useful if you can show them later. In MYLEDGER, sensitive actions are recorded centrally, with categories and retention, as evidence of who did what. feature:acb-audit-log When a client, a reviewer or your own quality process asks "who approved this and when?", the answer should be a record, not a recollection. Gates reduce mistakes; they do not remove the need for a safety net. MYLEDGER lets you preview a plan to roll a client's ledger back to an earlier time, apply it, and undo it if needed. feature:acb-restore-to-time These are the questions we would ask a vendor, including ourselves. You can read more about how we approach security on the Fedix site: https://home.fedix.ai/security?utm source=devto&utm medium=article&utm campaign=cold-start-2026-10&utm content=pub-fb0bb1110450 https://home.fedix.ai/security?utm source=devto&utm medium=article&utm campaign=cold-start-2026-10&utm content=pub-fb0bb1110450 If you want to see the approach in practice, start at https://home.fedix.ai/?utm source=devto&utm medium=article&utm campaign=cold-start-2026-10&utm content=pub-fb0bb1110450 https://home.fedix.ai/?utm source=devto&utm medium=article&utm campaign=cold-start-2026-10&utm content=pub-fb0bb1110450 .