AI Agents in Accounting: Tasks and Limits 2026

AI agents in accounting can use connected tools to carry out tasks, such as requesting missing information or preparing a bookkeeping entry. Their permissions determine whether they can propose a change, seek approval or make it. Set that boundary before giving them client work.

What Makes an AI Agent Different?

An AI agent can choose a next step and use a tool to carry it out. A chat answer stays in the conversation until someone acts on it. An agent with email access can send a message, while one with permission to edit records can change them.

The basic sequence starts with a task and the information available. The agent chooses an allowed tool, receives its result and decides whether to continue or stop. A failed tool call can change the next step. A successful call can create a real change outside the conversation.

A fixed AI workflow follows steps chosen in advance. An agent has more freedom to choose its steps. Both can take actions, so the label alone tells you little about what the software can change.

Keep action status visible to staff:

Status	What has happened	What staff need to see
Proposed	The tool has suggested a change	The proposed change and supporting information
Approved	An authorised reviewer has agreed to the change	Who approved which version and the permitted action
Completed	The connected tool has carried out the change	The actual result and its record in the destination

An approval is permission to act. Completion needs confirmation from the destination. An agent saying that it sent an email does not establish that the email service accepted it.

Accounting Jobs and Action Boundaries

Give an accounting agent permission for a defined job, with review before actions that change client records or communicate a professional judgement.

A client follow-up task can have a narrow instruction: identify a missing document and ask the client to upload it. Specify the recipient and the permitted request. If the client replies with a tax question, route it to a staff member instead of extending the task into advice.

An information-review task can compare answers with attached documents and identify a mismatch. The reviewer then decides whether the client needs to clarify an answer. Requesting clarification and replacing an answer are different actions with different permissions.

A Bookkeeping Posting Example

In this fictional example, a company uploads an invoice for equipment. A bookkeeping AI agent suggests an expense account based on the supplier description. The invoice describes an asset, so a bookkeeper must check the proposed account and GST treatment before posting.

With read-only access, the agent can leave its suggestion for review. Permission to create a draft entry allows it to prepare the entry without posting it. Write access that includes posting can change the client ledger immediately.

For this task, require approval from the practice’s authorised bookkeeping reviewer before posting. That approval must cover the selected client and invoice, the account, amount and GST treatment. A changed proposal needs fresh approval. After posting, confirm the saved entry against the approved version.

An account suggestion is not a completed entry. Treating it as posted would leave the transaction missing from the client’s records. Posting the unchecked suggestion would put the wrong account into those records.

Permissions Before the First Action

Choose the narrowest access that completes the job. Keep client selection fixed and separate reading from permission to edit, send or post. Set a stop point for uncertain information, an unexpected reply or a failed action.

The Tax Practitioners Board’s AI guidance for tax practitioners requires professional judgement when relying on AI for tax agent services. It recommends reviewing generated content throughout the work and documenting those steps. Tool permissions do not transfer that responsibility to the agent.

Before disclosing client information to a third-party AI provider, obtain the client’s permission unless you have a legal duty to disclose it. A staff member approving a software action is a separate decision from the client agreeing to disclosure.

Use only the client information the task needs. The OAIC’s commercial AI privacy guidance recommends keeping personal information out of publicly available generative AI tools.

Records and Corrections

An action record must let your reviewer reconstruct what the agent attempted and what actually changed. A chat transcript alone can miss a rejected email, a duplicate posting or an edit made to the wrong client.

For each consequential action, retain:

• The client and task identifier, with the source information used.
• The proposed action and the version the reviewer approved.
• The reviewer’s decision, identity and reason for any change.
• The tool used and the time of the attempt.
• The destination’s result, including a saved-entry reference or failure message.
• Any correction and the staff member who approved it.

These are practical record fields for this workflow. The TPB guidance also calls for documented review and ways to understand and contest AI decisions.

If a bookkeeper finds the equipment invoice posted to the wrong account, stop further posting on that task first. Compare the saved entry with the invoice and the approval record. Establish whether the agent followed an incorrect approval or changed the instruction after approval.

Use the authorised correction procedure in the bookkeeping software and preserve the correction record. Review any later work that relied on the entry. Resume only after the permissions or task instruction causing the error have been corrected and checked.

If an action times out, inspect the destination before trying it again. A timeout can leave staff uncertain whether an entry was saved. Checking first prevents a second attempt from creating a duplicate.

Tax Genie’s Automatic Client Fixes

Tax Genie takes the checking and chasing out of client onboarding through a defined sequence. It reviews submitted forms and sends fixable issues straight back to the client.

1. Tax Genie runs rule-based checks for ABN, ACN and TFN format, register matches, missing documents and cross-field consistency.
2. Every TFN is removed before the AI review runs.
3. AI reviews the answers for common client mistakes, such as GST registration with no reporting frequency.
4. Tax Genie emails the client a plain list of what to fix and reopens only the affected sections on the same private link.
5. The client corrects those sections, and the practice receives the corrected form.

Staff can turn individual checks on or off. They can also ask a client to change specific sections. When staff correct a client’s answers, Tax Genie records who changed what and when.

In a fictional form example, a company says it is registered for GST but leaves its reporting frequency blank. The fix asks the client to complete that answer. The practice gets the corrected information without writing the follow-up email.

This is a bounded review-to-fix task. Its action is asking for corrections to the form. Permission to chase that answer does not give an agent authority to decide the company’s tax position or post a bookkeeping entry.

Use the AI for accountants topic guide to connect this reviewed form handover with other practice tasks. When choosing an agent for a separate job, define the permitted action and its reviewer before connecting it to client records.