AI for Accounting Firms: Automate the Work Around the Ledger

CA ANZ's February 2026 article cites global research in which 92% of chartered accountants wanted AI training and 30% had received it. Those figures describe a training gap, not the percentage of Australian firms using AI or the amount of work that can be automated. Use that distinction when buying tools. A team can learn to write better prompts and still spend its afternoons chasing missing attachments. Conversely, a reliable reminder and status update may remove repeated work without using a language model. Ask staff to record where a job waits, what they copy between systems and what they repeatedly explain to clients. Review a recent completed job alongside a delayed one. The difference often identifies a workflow with a measurable starting point. Source: CA ANZ on AI capability and training.

Create a checklist for the particular job, rather than sending every client the same document request. Each item needs a status such as requested, received, needs clarification or accepted. A reminder should read those current statuses before it is sent. In an illustrative bookkeeping workflow, a client uploads an invoice to the agreed folder. The connection associates it with the correct job. A reviewer accepts it or requests clarification. The next reminder asks only for the items still outstanding. A client replying by email should have a defined route back into that checklist. AI may propose a document type or summarise a reply. Keep an exception for a file that cannot be confidently linked to a client or period. Do not turn an uncertain match into an accepted document merely to make the dashboard look complete. The owner of this workflow needs to be able to pause reminders, correct a match and see what was sent. Read the more detailed accounting client onboarding guide for intake and approval steps.

First review what the existing accounting product already does. A custom integration should address a remaining gap rather than reproduce standard features. Possible gaps include preparing a review queue, attaching supporting records or routing an unresolved item to the person who knows the answer. For a proposed connection, distinguish reading transactions, preparing a suggestion and changing the ledger. They have different authority requirements. Give the reviewer the underlying record and the reason for the suggestion. Record the decision so the team can explain the outcome later. Xero's published Accounting API specification includes invoice, bank-transaction and invoice-attachment operations. That establishes possible connection points; it does not establish that an integration can perform every reconciliation action through the API. Verify the required operation and access scope before promising the workflow. Test with an unfamiliar supplier, an incomplete document and an ambiguous reference. The useful outcome may be a well-prepared exception, rather than an automatic posting. Source: Xero's official Accounting API specification.

Build the reporting pack from an identified source and period. Keep a record of when the data was retrieved and what has been approved. A template can assemble the agreed figures without asking a model to invent or recalculate them. If AI drafts commentary, supply the relevant approved figures and instructions about what it may infer. Ask it to flag missing context. A difference in revenue does not, by itself, establish its cause. The professional reviewing the pack should be able to trace a statement back to the figures and supporting information. Use an approval step before a client receives the report. If source data changes after the draft is prepared, mark that draft as needing review. Keep the previous approved version available so staff can explain what changed. A sensible first scope can be one report type for one group of clients. Adding every report and every exception at once makes it harder to establish whether the connection or the review process needs adjustment.

A follow-up workflow should know why the message is due. That might be a missing document, an unanswered clarification or a report awaiting approval. It should stop when the task is complete and pause when the responsible staff member intervenes. Use approved templates for routine requests. AI can propose a clearer draft for a complicated reply, with a person reviewing it. Give the team a history of requests so a client is not chased by several people for the same item. Decide how the workflow behaves when a client responds through a different channel. If the portal and email are disconnected, a received document can still look missing. Agree on one authoritative status and who can change it. Measure the delay between a request and a usable response. Counting messages sent rewards activity, while the business needs the job to progress. Discuss your accounting workflow with a recent example of where that progress stops.

Write a short baseline before the build. Record the staff minutes spent processing a job, the time it waits on the client, how often an exception is reopened and whether the job arrives at review complete. Measure the same items during the pilot. For an illustrative calculation, saving 20 minutes on each of 30 comparable jobs releases 10 hours of processing time. Subtract time spent reviewing outputs, correcting failures and maintaining the workflow. Released time is not automatically extra revenue; the firm still needs a plan for how it uses that capacity. Agree on what counts as failure. A request sent to the wrong client, a repeated reminder after completion and a silent connection error should be visible in the pilot review. Include a way for the team to return to the manual process while the issue is resolved. Make the decision to expand after reviewing both completed and failed jobs. A good demonstration on one tidy file is not evidence that the process handles the firm's actual workload.

Use business-owned accounts for the systems and integrations. Document who can reconnect each service, change a reminder rule, review an exception and approve an external message. Review provider arrangements, access and retention requirements with the appropriate people before sending client information to a new service. The operating instructions should show a routine job and a failed one. Include a connection inventory, the role list, alert destinations and the procedure for a departing staff member. Test recovery with non-client data where possible. Training should cover those operating tasks, rather than ending with a demonstration of prompts. The firm needs to know what the workflow does when it cannot complete its job. Our automation handover guide sets out the documents and ownership decisions. The best next project is the one the team can run after the builder leaves.

What should an accounting firm automate first? Start with a repeated, bounded workflow that has an owner and a measurable result. Document collection, exception preparation and client follow-up are candidates. Check the accounting product's existing capabilities before adding a custom connection. Does accounting automation need AI? No. Rules and integrations can handle document status, reminders and record updates. AI is useful where interpreting unstructured material or preparing a reviewed draft adds value. Can an integration automatically reconcile everything in Xero? Do not assume that. Verify the specific API operation, permission and review requirement. Reading bank transactions or preparing a suggestion is different from completing every reconciliation action. How do we measure whether the project worked? Compare comparable jobs before and after the pilot. Include processing time, waiting time, reopened exceptions and maintenance effort. Track the result the workflow was bought to produce.