Designing AI for accountants: approval gates, masking and an audit trail

If you run a practice, an AI assistant that can touch client records raises four plain questions. Who approves a change? What leaves your systems? Whose permissions does the assistant use? And afterwards, can you show what happened?

These are engineering questions as much as policy ones, and they are worth asking of any tool, ours included. Here is how we think about each, and what Fedix actually does.

1. A person approves the change

The most useful rule we have seen in AI tooling is simple: the model proposes, a person decides. In MYLEDGER, changes requested by the Copilot or a connected assistant wait for a person to approve or reject them.

The same idea applies to Xero. Every change Fedix would make in Xero is confirmed by a person in a confirmation dialog first.

For an accountant, the useful test is whether you can look at a proposed action and decide whether it goes ahead. Approval is a point where your judgement applies; it does not replace checking the underlying figures yourself.

2. Control what goes to the model (MYLEDGER Copilot)

Client files are full of personal identifiers. Personal identifiers can be masked in each Copilot session before content is sent to AI model providers.

If you are deciding whether a tool fits your practice, ask where masking is applied, how you switch it on, and what it does not cover. Our security page is the place to start: https://home.fedix.ai/security?utm_source=devto&utm_medium=article&utm_campaign=cold-start-2026-10&utm_content=pub-192a41f25b1d

3. The assistant acts as the signed-in user

A common design mistake is giving an assistant its own broad service account. We took a different route. The FediPilot desktop assistant uses FEDIX as the person who signed in, with the same permissions.

So the assistant sees what that person is allowed to see, not more. Desktop AI actions are also tied to an enrolled device, and sensitive actions need approval.

4. Keep a record you can point to

When something goes wrong, or a reviewer asks, you want evidence rather than recollection. In MYLEDGER, sensitive actions are recorded centrally, with categories and retention, as evidence of who did what.

Where AI helps with return preparation

On the return preparation side, an AI assistant works through the client's documents and produces a first draft of the return preparation for an accountant to review. The word that matters there is review: the accountant stays responsible for the return.

Questions to ask any vendor

• Can I see and reject each proposed change before it is applied?
• What personal information reaches the model provider, and can I mask it?
• Does the assistant inherit my permissions or hold its own?
• Is there a central record of sensitive actions?

This article is general information about product design, not tax or legal advice. If you want to see how Fedix approaches these questions, start at https://home.fedix.ai/?utm_source=devto&utm_medium=article&utm_campaign=cold-start-2026-10&utm_content=pub-192a41f25b1d and read the security page.