AI Automation for Accountancy Firms in Ireland: What to Automate First

Here's a Monday morning in a practice, the shape of one anyway.

8:42. Open the tracker and work out which clients still haven't sent their records. 8:55. Write three more reminders, all much the same as the last three. 9:20. Pull together the background for an eleven o'clock call: last year's file, what was promised in March, the thing they keep asking about. 10:05. Update the recurring report that was updated the same way last month. 10:40. Turn Friday's meeting notes into actions before they go cold. 11:00. Someone puts their head round the door to ask where a particular client's payroll information lives.

Six jobs, most of a morning gone, and here's the only question worth asking about any of them.

How much of that actually required an accountant?

I don't mean could a machine do it. I mean did it need years of training, a professional qualification, and the judgement that comes with both. For most of that morning the answer is no, and yet it was an accountant's morning that it took.

The opportunity isn't where the conversation usually goes

Ask most firms where AI fits in an accountancy practice and the conversation goes straight to the technical work. Tax computations. Accounts preparation. Analysis. Advisory.

Those may matter eventually. They're also the parts stuffed full of professional judgement, carrying professional and regulatory consequences, which makes them the hardest and riskiest place to start and the slowest to show anything for the effort.

Meanwhile the work piled around them is repetitive, high-volume, low-judgement, and almost entirely unbilled.

Accounting expertise is scarce, and the practices I talk to say hiring hasn't got any easier. There is no shortage whatsoever of administrative repetition. Pointing the first at the second is a strange way to run a firm, and most practices do it because the admin arrived gradually and nobody ever decided to.

The opportunity here has very little to do with accounting, and a great deal to do with everything stacked around it.

What makes a practice different from other professional firms

I've written more generally about the jobs professional services firms should look at, and much of it applies here. But an accountancy practice has one feature that changes the sums completely, and it's worth naming before going further.

Everything happens to everyone at once.

A consultancy chases a dozen clients for a dozen different things on a dozen different timelines. A practice chases three hundred clients for the same document, in the same six weeks, on a calendar somebody else sets.

Revenue and the CRO write that calendar. The Form 11 pay-and-file deadline lands on 31 October, extended to 18 November for those filing and paying through ROS. VAT runs in two-month periods, filed and paid by the nineteenth of the month after each one ends. A company's annual return and financial statements are due at the CRO within 56 days of its annual return date, and miss that one and the company can lose its audit exemption, which is the sort of consequence a client remembers for years.

None of that is negotiable and none of it moves. So the work doesn't arrive in a manageable trickle. It arrives in waves, it arrives for hundreds of clients simultaneously, and the same three sentences get written several hundred times by qualified people who could be doing something else.

That's what makes the admin in a practice worth counting properly. No single one of those jobs is heavy. The weight comes from the same light job running three hundred times inside a fortnight.

What I'd look at first

Chasing client records

This is the clearest candidate in the building, and every practice knows it.

Bank statements. Payroll information. Purchase invoices. Expense records. Confirmations. Identity documents for onboarding. The year-end pack that was requested in July and is still, in October, partly outstanding.

The job underneath the chasing is really a tracking job: knowing what was asked for, from whom, what's arrived, what's still missing, when it was last mentioned, and what the next message should say. Practices run this on a spreadsheet, or on someone's memory, and both fall over at volume.

Handled properly, the sequence gets approved once, the reminders draft themselves against what's actually outstanding, responses get recorded as they land, and the exceptions come back to a person.

That boundary is the important part. The first reminder is routine. The fourth one usually isn't, because by the fourth reminder something else is going on: the client's in difficulty, or annoyed, or has sent it twice already and nobody logged it. A person takes it from there, and the system's job is to notice that the moment has arrived rather than send a fifth cheerful nudge into a situation that's changed.

Preparing for a client meeting

Whoever takes the eleven o'clock call needs the same things every time: what was discussed last, what's still open, what was promised, what's changed since, which deadlines are close, and the thing this client always circles back to.

Assembling that from the file, the correspondence, and last year's notes is a twenty-minute job done properly and a five-minute job done badly, and under pressure it's usually done badly or not at all.

An agent can have it ready before anyone asks, with each point showing where it came from and a plain statement of what it couldn't find. The accountant decides what matters, what needs handling gently, and what shouldn't be raised this week at all. The file gets assembled; the person walks in prepared.

Turning the meeting into actions

Notes come out as pages of conversation with the commitments buried in them. A first pass can pull out the actions, the owners, and the dates, link each back to where in the notes it came from, and flag the ones where nobody actually said who or when.

Somebody still confirms what was agreed. "We'll have a look at that before year end" and "you're sending the loan agreement on Thursday" read far more alike on a page than they sounded in the room.

Once it's confirmed, the actions and the practice-management update can go through together, which removes the other half of this job: the bit where the meeting happened, everyone moved on, and the record waited until Friday.

The recurring reporting

Monthly management reports for clients. The internal pack. Aged work-in-progress. Where everything stands going into a deadline. Pipeline.

Same headings every time, same sources, same comparison against last period, and then somebody senior spends an afternoon assembling it.

Gathering the approved inputs, comparing the periods, populating the recurring sections, drafting first-pass commentary on what moved, and flagging what's missing or doesn't reconcile is a good job to hand over, precisely because the rhythm is fixed and the time comes back every single month.

The report owner still interprets the numbers and approves what circulates. The report may need professional judgement. Most of the labour that produces it doesn't.

Routine client correspondence

The information request. The acknowledgement. The deadline reminder. The progress note. The confirmation of something already agreed.

Drafted from the client record, the firm's own approved wording, and the facts of this particular matter, with sourced facts kept separate from suggested phrasing, and anything unsupported flagged rather than smoothed over.

A person approves every send. Advice, commitments, complaints, and anything with feeling in it are not routine correspondence, whatever inbox they arrived in.

Document intake

This is the one most particular to accountancy, and at volume it's enormous.

Records arrive by email, by post, through a portal, and in a carrier bag. Somebody works out what each document is, which client and which year it belongs to, whether it's the thing that was actually asked for, and what's still outstanding.

A good version of this identifies the document type, matches it to the client and period, checks it against what was requested, extracts the routine fields, and puts anything it isn't sure about into a queue rather than filing it confidently in the wrong place. Every extraction points back at the source page.

The rule I'd hold hardest here: an uncertain match must look uncertain. A system that files a document in the wrong year with total confidence has created a problem that surfaces months later, in front of a client, and costs more than the whole automation saved.

Finding what the firm already knows

Every practice has a Mary. How was this treated for that client three years ago? Where's the approved wording for this letter? Which partner dealt with the last case like this, and what did they conclude?

That knowledge lives in people's heads and in old email, and it walks out the door at five o'clock every day, permanently at some point.

Made findable, with the source attached to every answer, it stops depending on who happens to be in. The requirement is that the firm's approved position is identifiable as approved, has an owner, and has a review date, because a confidently retrieved answer from 2021 is worse than no answer at all when the treatment changed in 2023.

Where I'd assist rather than hand over

Some work is repetitive in its assembly and genuinely judgement-heavy at the point of decision. The preparation goes to AI and the decision stays where it belongs.

Commentary on management accounts. Proposal and engagement scoping. Preparation for technical research, where the material gets gathered and read but the conclusion is drawn by a person. Difficult client correspondence, drafted to save the blank page and then rewritten by whoever knows the relationship. Assembling a file for partner review.

In every one of those the model prepares and the accountant decides. The assembly was never the expensive part of any of them.

What I'd keep human

The accounting judgement itself. An agent can gather the file, check a draft against an approved list, and show the treatment used last time. The conclusion belongs to the qualified person who understands the client and carries the consequence.

Client advice. Same reason, and it's what the client is paying for.

Any difficult conversation. Bad news, a dispute, a client in financial difficulty, a complaint, a conversation about fees. Prepare with AI. The conversation belongs to whoever holds the relationship.

Anything written about a person. Reviews, references, notes that end up on a file. The moment the subject is a colleague rather than a task, it stops being administration.

The final sign-off. Someone identifiable is answerable for what the practice puts its name to. The standard isn't whether AI can produce an answer. It's whether the firm is willing to be answerable for it.

And a category worth naming separately: work that fails quietly. A filing date. A limitation period. A calculation nobody re-checks because it's always been right. The CRO annual return that's four days late and costs a client its audit exemption. Those failures don't announce themselves, and a job whose failure nobody would notice for six months is a bad place to remove a human pair of eyes, however repetitive it looks.

Five hours a week

Say a manager or a partner loses five hours a week to chasing, preparing, assembling, and writing up. Not a wild number; most people who do the sum find it's higher.

Across a working year of about forty-six weeks, once annual leave and the public holidays come out:

5 hours × 46 weeks = 230 hours a year.

I'd stop the sums there. I don't multiply them by anybody's salary, and I wouldn't in your boardroom either, because the moment you do the conversation stops being about work and becomes about people, and you lose the honesty you need from everyone else in the room.

I'd rather ask what those hours could have been instead:

• Client capacity you didn't have to hire for.
• Turnaround that stops being the thing clients complain about.
• Advisory work that keeps getting postponed because compliance eats the week.
• Proper partner review, on time, rather than squeezed.
• Fewer evenings in October and November.

The commercial argument has nothing to do with AI being cheaper than an accountant. It's a question about how a scarce and hard-to-replace sort of capacity is currently being spent, and whether the firm would choose that split if it were choosing it fresh today.

Don't automate the first annoying thing you notice

Everything above is a starting list, and it isn't a plan. The loudest complaint in a practice is rarely the heaviest piece of work in it.

Frequent small jobs beat dramatic occasional ones, almost always, and nobody believes that until they've counted. Some tedious processes are tedious because they're badly designed, in which case automating them just makes a bad process run faster, and half a day rearranging it would have been the cheaper fix. And some work should stay exactly where it is.

That's why I'd map the whole practice before choosing anything. I've written up how to find the work AI should actually do: find it, weigh it, ask how much judgement is in it, then run every job through the five-way sort. Or start smaller again and put one task through the free Repetitive Work Calculator.

What an assessment should leave you with

If you're paying somebody to assess this, whether it's me or anyone else, here's what should be in your hands at the end of it.

An inventory of the recurring work, written down. For each job: how often it happens, roughly how long it takes, who does it, how much judgement it needs, whether the source material actually exists, whether the output is easy to check, and what happens if it goes wrong.

Then a decision on each one, from five: automate it, assist it, fix the process, keep it human, or leave it alone for now.

Then a ranked shortlist of what to do first, with the capacity you'd expect to get back, and the reasoning shown so you can argue with it.

Count the work before you buy the technology. That holds whatever the answer turns out to be, and for a fair number of practices the answer is to wait a while first, and the day tells you that too.

When it does go the other way, the practice doesn't feel automated. The records chase went out on Monday without anyone having to write it. The file for the eleven o'clock is sitting there assembled. The monthly pack is drafted and waiting to be read rather than built. And the person who used to spend a day a week on all that is doing the work they actually qualified for.

None of it requires the partners to become the firm's AI experts, and they shouldn't have to. They stay accountable for the professional work; somebody else keeps up with the tools, the workflows, and what deserves attention next. The end state worth wanting is simple: when an AI question comes up, everyone in the practice knows who's looking after it.

Don't know where the time is actually going?

More on what I do and the firms I do it for: AI consultant for professional services firms in Ireland.