AI Agents for Accounting Firms, Client-Side vs Back Office

Last reviewed: September 2026

What is the difference between a back-office and a client-side AI agent?

The dividing line is whether the agent must talk to somebody outside the firm to finish its job. A back-office agent has everything it needs the moment it starts: the bank feed, the ledger, the prior year file. Its work is transformation. A client-side agent starts with an absence. Its work is acquisition, and acquisition depends on another human deciding to act.

That sounds like a technical detail. It is the whole thing. Transformation is deterministic and improves with a better model. Acquisition is a negotiation with a distracted business owner, and no model upgrade makes a client open their filing drawer. What moves it is timing, channel, persistence and tone.

Do Xero, MYOB and QuickBooks already do the client-side chase?

Not the one an accounting firm needs, and the distinction is easy to miss because the marketing language overlaps. All three major ledger vendors ship agents that chase somebody. In every case the somebody is the business owner's own overdue debtor, not the firm's client sitting on a shoebox of receipts.

MYOB announced its AI suite on 2 March 2026: AI BAS, AI Business Insights, Smart Reconciliation and Smart Invoice Reminders, all "at various stages in beta, progressively rolling out". In MYOB's own words, Smart Invoice Reminders "prepares suggested actions based on late payer behaviour" and offers "tone suggestions for communications". Read the MYOB announcement and the pattern is clear: three back-office agents, and one that helps a business collect its own money.

Intuit follows the same shape. Its 1 July 2025 announcement of a virtual team of AI agents describes a Payments Agent that predicts late payments and sends invoice reminders, and an Accounting Agent that automates categorisation. Intuit's QuickBooks Online help documentation, updated in August 2026, goes further: the Accounting AI raises questions to request more information and handles back and forth between the business owner and their accountant. That is the closest any ledger vendor gets, and it still lives inside the client's own subscription, so it only helps clients on that ledger who log in.

Xero is in the same position with Just Ask Xero, which drafts invoices, automates reconciliation and sends reminders for overdue invoices. Again pointed at the business's debtors. None of this criticises the vendors: they build for the ledger subscriber, and the subscriber is the small business. Chasing forty clients for a hundred missing artefacts across three ledgers is not their problem. Separately, the connection layer into Xero and QuickBooks decides what any agent can see in the first place.

Is a document collection portal the same as a client-side agent?

No, and firms that already own a portal should be clear about what they are still doing by hand. Document request portals are a mature category and Australian practices use them well. Platforms such as Karbon build client requests and automated reminders into the workflow, so the client sees what to action and the firm sees what it is waiting on. That is valuable. It is not what is missing.

What is missing is the reasoning step in front of the portal. A portal sends the list a human built. Somebody still opens the file, checks the unreconciled items, remembers this client has a second credit card, checks whether the trust distribution minute was signed, and assembles the request. Then somebody maintains that list as documents trickle in, so the client is not chased for something they sent last Tuesday. That assembly and maintenance is the actual labour, and a portal does not touch it.

A client-side agent inverts the order. It derives the request list from the ledger rather than from someone's memory, keeps deriving it as the period progresses, and retires each item the moment the document lands. The chase becomes a consequence of the reconciliation, not a separate administrative project beside it.

What does "waiting on client" actually cost an Australian practice?

Almost no practice knows, because the cost never appears on a timesheet. Processing time is measured. Dead time is not. A job that needs four hours of work but sits in "waiting on client" for five weeks has consumed four billable hours and five weeks of working capital, and the second number is the one that hurts.

There is a useful adjacent benchmark. Xero Small Business Insights, published 30 July 2026, found Australian small businesses waited an average of 22.9 days to be paid after issuing an invoice in the June quarter of 2026, and were paid 6.0 days late. That is Australian data on invoice payment, not document supply, so do not read it as a document-chasing statistic. Read it as a proxy for responsiveness: even with money owed and a due date attached, the average runs three weeks. A request for a missing receipt has neither.

Two costs compound while the job sits. Work in progress ages, and the re-familiarisation tax accrues. Every reopen means rebuilding context someone had fully loaded three weeks ago: which entities, which queries were resolved, what the client said in the call nobody wrote up. Where a job is picked up and put down five or six times a quarter, that is a meaningful share of total effort, and every minute of it is pure loss.

A worked example, labelled as such rather than presented as a client result. Replace every figure with your own. A practice runs 120 compliance jobs a quarter. Each takes three chase touches at ten minutes: 60 hours. Each is reopened four times at eight minutes of re-familiarisation: another 64 hours. Roughly 124 hours a quarter, about 500 hours a year, spent on the mechanics of waiting rather than the work, none of it on a fee note. It is also the number a back-office agent does not touch.

What does a client-side AI agent actually do, step by step?

Six steps. The first is what separates an agent from a reminder schedule.

1. Work out what is missing by reconciling against the ledger. Unreconciled bank lines, transactions with no attachment, accounts with no statement, entities with no signed engagement. A per-client list nobody had to write.
2. Turn each gap into a request a non-accountant can act on. "Receipt for the $1,840 payment to Officeworks on 14 August" is actionable. "Please provide outstanding source documents" is not, which is why so many requests get ignored.
3. Contact the client on the channel they actually use. Email for most, but the point is their channel, not the firm's preferred portal. A request that needs the client to remember a password has an extra failure point.
4. Escalate on a schedule the practitioner sets. A gentle first touch, a firmer second, then hand off to a named human rather than a third and fourth message from a machine. The ladder is a firm decision, not a vendor default.
5. Stop chasing the instant the document arrives. The single most important behaviour, and the one most reminder systems get wrong. Match the inbound document to the open request, close it, file it against the transaction, remove it from every future message.
6. Route anything that is not a document to a human immediately. A reply that says "I'm going through a divorce, can this wait" is not an exception to queue. It goes to a person the same day.

This is the pattern Agentive's AI Employee for accountants is built around: it identifies unreconciled items in Xero, generates the missing document list, emails clients requesting receipts and statements, and follows up automatically if there is no reply within the timeframe the practitioner sets. The same engine does the back-office half, because step one is only possible if the agent can already read the ledger. Buy the chase and the processing separately and you maintain the list by hand anyway. The broader picture sits in our guide to the AI Employee for accounting firms.

How do you stop a chasing agent from annoying your clients?

Accept first that this is the hard part, not a settings screen. A chasing agent that irritates a good client costs more than every hour it saves, and the loss is invisible in the reporting because it surfaces months later as a client who quietly moves. Nobody writes "left because of the emails" on an exit form.

Agentive's position, a design position rather than a measured result, is that the escalation ladder belongs to the practitioner rather than the vendor, and that the safe direction to be wrong in is too slow. One touch too gentle costs a few days. One touch too aggressive costs an engagement. Vendors default to aggressive because it flatters completion metrics in a trial. Four controls are worth insisting on.

• A hard cap on touches, and a send window. Two or three messages, then a human, and business hours on weekdays only. An agent that can send unlimited messages will eventually send an inappropriate number, and a reminder at 6:41am on a Sunday reads as harassment however politely it is worded.
• Suppression rules you control. Pause chasing during a bereavement, a sale process, an ATO audit, or any period the partner nominates. One click, not a support ticket.
• Approved tone per client tier. Your largest client and a first-year sole trader should not receive the same escalation language. MYOB shipping tone suggestions in its own reminder product shows the vendors know this too.
• Full visibility of what went out. Every message, to whom, when, and what triggered it, readable by any staff member before they ring that client. Nothing is worse than a partner calling a client who has just received a chaser the partner has not seen.

The cadence design is the same problem as accounts receivable follow-up, worked through in automating overdue invoice follow-ups. The mechanics transfer directly. The difference: a debtor knows they owe you money, whereas a client sitting on receipts does not think of themselves as blocking anything.

What are the compliance limits on an AI that contacts your clients?

An agent that sends messages in your firm's name is providing a service on your behalf, and Australian tax and BAS agents have explicit guidance on that. TPB(GS) 55/2026, issued by the Tax Practitioners Board on 22 July 2026, sets out how the Code of Professional Conduct applies to AI in tax agent services. Three parts bear directly on a client-side agent.

First, confidentiality. Code item 6 provides that unless you have a legal duty to do so, you must not disclose any information relating to a client's affairs to a third party without your client's permission. The guidance recommends telling the client about the proposed disclosure, including to whom and where it is made, where data is stored and whether AI tools may be used, and notes the permission may come via a signed letter of engagement. A client-side agent reads the ledger to build its list, so consent is live from the first run, not the first email. The broader question of what may safely go into a tool at all is covered in is it safe to put client data into AI.

Second, supervision. Section 35 of the Determination requires that each entity providing tax agent services on your behalf is appropriately supervised, having regard to the services provided and your system of quality management. An unattended chasing agent with no named owner does not meet that description. This is why Agentive keeps a per-action audit log across both halves of the job: for a client-side agent the log is not a nice-to-have, it is the evidence that satisfies the supervision obligation.

Third, competency. The guidance is explicit that tax practitioners remain accountable for the accuracy of information and advice they provide to clients. If your agent tells a client a document is missing when it is already on file, that is your firm's error. The disclosure and consent picture is covered in our companion post on what tax and BAS agents must tell clients about AI, and the accuracy side in why AI invents figures in financial reports.

Where should an Australian practice start?

Start by finding out whether your bottleneck is actually on the client side. If it is not, none of this applies to you. Four steps.

1. Age your jobs by status, not by date. Split each job's elapsed days into days in progress and days waiting on client. If the second number is more than half, your constraint is acquisition and a back-office agent will not move it.
2. Count the touches. For ten stalled jobs, count the chase messages sent and the times the file was reopened, then multiply across your job volume. That is the hours figure to hold a vendor against.
3. Write the escalation ladder before you buy anything. Number of touches, interval, tone at each tier, suppression rules, hand-off point. If you cannot write it down, you are not ready to delegate it to software.
4. Pilot on your most tolerant clients, not your worst. The instinct is to point the agent at the client who never sends anything. Do the opposite. Run it on clients who will tell you honestly that the second email was too much, then tune, then widen.

One piece of context. Karbon's State of AI in Accounting 2026 report, published 20 January 2026 from nearly 600 accounting professionals across six continents, found 98% of firms now use AI and average savings reached 60 minutes per day per employee, but fewer than half invest in training and only 21% have an AI policy or strategy. That is global rather than Australian data, so treat the percentages as directional. The gap is not: adoption has run ahead of governance, and pointing an unattended agent at your client list without a written policy is exactly that risk.

The half of the job nobody is selling

The back office is the easier half and the smaller half. Easier because the data is already there and the failure mode is a number your review was built to catch. Smaller because processing hours were always on the schedule, whereas the weeks a job spends waiting on a client were never on anything. Australian bookkeepers and accountants have watched vendors compete hard over the first half while the second stayed manual.

The right answer is not to choose. A client-side agent cannot work out what is missing unless it can read the ledger, and a back-office agent stops the moment it hits a gap only the client can fill. Running them as disconnected products is how practices end up maintaining the request list by hand while paying for automation. Agentive builds an AI Operation Engine that does both under one audit log, which is why the bookkeeping capability and the client chase are not sold separately.

So ask every vendor one question this quarter: when the document is not there, what does your agent do? If the answer is that it flags an exception for a human to chase, you are being sold the easier half. Still worth buying. Just do not confuse it with the thing actually holding your jobs up.