The bottleneck in your accounting firm isn't the work. It's chasing the records.

Accounting firms do a lot of knowledge work that AI can take the drag out of: reading documents, drafting, checking what’s missing, tidying notes, digging up past advice, summarising a client’s history. The opportunity is real. So is the risk, and pretending otherwise is how a firm ends up with a data problem, so client confidentiality has to shape the system from day one rather than get bolted on after the first scare.

But start by being honest about where the week actually goes, because it isn’t mostly the accounting. It’s the chasing. The document that never arrived. The client who sent eleven of the twelve things you asked for. The workpaper prep that has to happen before anyone does the work that needs a brain. That’s the bottleneck in most firms, and it’s exactly the low-judgement, high-friction work AI is suited to, which is why it’s the right place to look first, not the clever analysis a partner should be doing anyway.

Where the hours go in a mid-sized firm

Picture a senior accountant in a Toowoomba firm in the thick of tax season, with clients spread from town out across the Downs: graziers, ag suppliers, a transport operator or two, the usual mix. She has thirty jobs in progress. On Tuesday she opens the one she meant to finish last week and finds it’s still waiting on a livestock summary and a loan statement. So she writes the third follow-up email to that client, switches to a different job, spends forty minutes rebuilding context she had cold a fortnight ago, and gets twenty minutes of actual accounting done before the front desk puts a call through.

Multiply that across the firm. Every stalled job costs the chase itself, maybe ten minutes, plus the context-switch either side of it, which is the expensive part. Firms that track this honestly tend to find seniors losing a day a week or more to chasing and re-orienting, none of it billable at anything like the rate their judgement is worth. The work that needs a brain gets squeezed into whatever’s left, which is why review happens at 7pm and why capacity feels short even in years when the client list hasn’t grown.

The confidentiality question comes before the use case

The fastest way to create a problem is to let staff drop client material into whatever public AI tool they’ve got open, usually with perfectly innocent intent, to tidy a letter or reformat a schedule. The firm still needs clear rules on what can go where and which tools are approved, before it becomes a habit nobody can walk back, because client data that’s left the building can’t be recalled.

So the safer first step is a controlled workflow around one bounded task, document intake, missing-information checks, workpaper prep, client email drafts a person reviews, internal policy search, rather than a general “use AI” free-for-all. Match the architecture to the data in front of you: plenty of firms are fine using approved cloud tools for low-risk tasks, while sensitive document sets warrant private processing, role-based access and logging you can actually point to. What decides that is the data, not the sales deck.

Writing the rules down is the unglamorous half of this, and it can’t be skipped. An approved-tools list staff can find without asking, a plain statement of what client material can and can’t be pasted where, and a named person to go to when the answer isn’t obvious. Two more questions worth settling early: what happens to prompts and outputs when someone leaves the firm, and what you’d tell a client, or the professional body, if they asked exactly how their file was processed. If the honest answer today is a shrug, settle it before the next busy season makes everyone improvise.

AI prepares, a person still signs

In a firm the rule is simple and it doesn’t move: AI does the prep and stops short of finishing the job. Let it read a document, extract the fields, summarise a client history, draft a query, flag something that doesn’t add up. A person reads the output before any of it touches advice, a lodgement, a bill or a client. The efficiency is real and the accountability doesn’t shift, same as it never has.

Put that review point in the workflow explicitly, rather than leaving it to whoever happens to be careful that day. The one thing worse than doing the work yourself is filing something that only looked done because a model was confident, and models are confidently wrong often enough that “it came from the AI” can never be the reason something went out. Write the review step down, and teach it.

Here’s why that step stays even after the tools improve. An extraction model reads a scanned trust deed and pulls the distribution date confidently and wrong, because the scan was skewed and a 3 read as an 8. Unchecked, that lands in a workpaper and everything downstream inherits it. Checked, it costs a reviewer four seconds against the source page the answer links to. That’s the trade the whole setup rests on: the model does an hour of reading in a minute, and the person spends seconds, not hours, confirming it against the documents.

Cut the chasing, which is the actual bottleneck

Missing records hold up accounting work more than anything else, so this is where the time comes back. Automation can scan the intake folder, work out what hasn’t arrived, draft the follow-up and update the task list. AI can read the client’s reply and sort the attachments, while the workflow still decides what happens next. Less chasing, and fewer files stuck waiting on a step nobody had defined.

Internal knowledge is the other strong early win, because most of a firm’s useful knowledge sits in templates, checklists, old workpapers, emails and internal notes, and a private assistant can find the answer to the procedure question someone gets asked twice a week. AI grounded in your own documents suits this because the answer cites the source it came from, which keeps it trustworthy and keeps staff using it.

Own the process, don’t just adopt the features

AI features are appearing everywhere inside accounting, document and practice-management tools, and some are handy. But a feature being available is not the same as it being safe for your work, and the firm still has to decide how each one fits its review standards, its promises to clients and its data rules. That’s a decision to make deliberately, not to inherit because a vendor shipped a button.

For the higher-value workflows, a small custom layer is often cleaner than stitching a dozen disconnected tool features together, because it puts the firm’s rules, logs, review points and source links in one place you can see and stand behind.

Measure review time, or you’ll never know if it worked

The measure that matters isn’t how many staff used the AI. It’s whether the jobs moved. Before the first workflow goes in, write down the current numbers: average days from records-in to job-out, how many follow-ups the average client needs, how many hours a senior spends on prep versus review. Check the same numbers a quarter later. If turnaround shrinks and review time grows, the workflow’s paying for itself. If staff are enthusiastic but jobs are stuck in exactly the same places, you’ve bought a toy, and it’s far better to learn that after one bounded workflow than after a firm-wide rollout.

That’s also the honest pitch to a sceptical partner who’s watched software promises come and go for twenty years. Not a transformation story. One process, these numbers, reviewed in ninety days.

So pick one recurring process staff already hate, document intake, missing information, workpaper prep, client history summaries, knowledge search, and build a controlled workflow around it. For a Toowoomba firm the point was never to look like a tech company. It’s to spend less of the week on admin and more of it on the judgement clients are paying for. If you know which bottleneck eats the most hours, tell us what it is and we’ll scope a bounded first workflow.