AI Automation for Professional Services Firms in Ireland: 15 Jobs Worth Automating

In the Irish professional services firms I talk to, it's usually the same picture underneath.

Somebody's going through three folders looking for the notes from the last meeting, and the meeting starts in forty minutes. The monthly report is due again and it doesn't feel like a month. There's an engagement letter to put together that's the same shape as the last forty. The CRM is waiting to be updated. So is the client.

None of it is big. That's exactly why it survives. Ten minutes here, half an hour there, and nobody ever puts it on an agenda, because putting it on an agenda would mean admitting how much of the week it takes.

I'd start there. The useful question isn't where you could use AI. It's which work keeps coming back, what it's costing you, and how much real judgement it needs.

I built my own answer to that before I sold anyone else one. Mercurious runs on a small fleet of AI agents doing exactly this kind of work inside my own practice: preparing the briefs, keeping what the firm knows, writing down how a job is really done, reading the correspondence for what changed, and learning how the firm sounds. They've been at it for months.

They have names, the Argus, the Archivist, the Structuralist, the Alienist, and the Profiler, and I'm aware that's a bit much for software. The jobs behind the names are deliberately ordinary. Before seven each morning there's a brief waiting for me with the two or three things that actually need me in it, and I've stopped opening five tabs to work out what happened yesterday.

Four things I check before handing a job over

1. It happens often enough to matter. A beautiful automation for a job that comes round twice a year will never pay back the afternoon it took to build. Frequency is doing most of the work on this list, and it's the first thing I'd count.
2. There's a method, even if nobody has written it down. AI can't follow a process the firm can't explain. If three people do the job three different ways and all three are right, that's a conversation before it's an automation.
3. The material it needs is actually there. A fluent answer assembled from half a record is still half a record, and it reads beautifully, which is the dangerous part.
4. Somebody sees it before the client does. Professional work has consequences. Nothing goes out without a person's name behind it.

The list below is a starting point rather than a verdict on your firm.

In this article, automate it means automating the preparation. A person stays responsible for whatever leaves the firm.

Before the meeting

1. Preparing a client meeting brief

The job isn't deciding what to say to the client. It's making sure whoever walks into the room isn't reconstructing eight months from memory in the car park.

A brief worth having pulls together the last few sets of notes, the actions still open, what was promised the last time, the deadlines coming up, and the thing this client always circles back to. It shows where each point came from, and it says plainly when it couldn't find something.

The emphasis is the part that stays human. Which issue deserves the room's attention, which one needs handling gently, and what shouldn't be raised at all this week.

2. Summarising a long client history

Somebody joins a matter that's been running two years. Reading everything is thorough, and it's also three days they don't have.

A summary can lay out the chronology, the decisions that were made, what's still open, and the documents worth opening. Every material statement should point back to the thing it came from, and the gaps and the contradictions belong in there too.

It's a way into the record rather than a replacement for it. If you're about to advise on a point, open the source. What you've saved is the finding, not the reading.

3. Gathering information across systems

A remarkable amount of a professional's week goes on moving between email, the document store, the practice system, a task tool, and a spreadsheet somebody made in 2023, in order to assemble one answer.

This is a good job for AI when the request is precise: everything relating to this meeting, the latest approved version of that, all correspondence on this matter, or which of the twelve items we asked for still haven't arrived.

Access is the hard boundary. Each agent gets what its job needs and nothing else, and it has to be able to tell you the difference between "there isn't one" and "I'm not allowed to look". Blur those two and it'll mislead the person trusting it.

After the meeting

4. Turning meeting notes into actions

Notes come out as pages of conversation with the decisions buried in them. Getting to a usable list of who's doing what by when is tidy, repetitive work.

A first pass can pick out the proposed actions, the owners, and the dates, then link each one back to where in the notes it came from. Where there's no owner or the deadline is vague, it should say so rather than fill the gap with something confident.

Somebody still has to confirm what was actually agreed, because "we should probably look at that" and "Aoife's sending it Friday" look far more alike on a page than they sounded in the room.

5. Drafting routine follow-ups

A firm writes the same handful of follow-ups over and over: the note after the meeting, the confirmation of next steps, the request for a document, the summary of what was agreed.

Drafted from the meeting record and the firm's own approved examples, what you get is a message that already has the right names, dates, and actions in it, sitting in a queue waiting for a person. Mine sound like my practice because the Profiler learned the voice from my own sent work rather than from an idea of how consultants write.

Nothing sends itself. The moment a follow-up contains new advice, a disputed point, or a commitment by the firm, it isn't routine any more and it shouldn't be treated as though it is.

6. Chasing missing information

Chasing is the clearest candidate on the whole list. It repeats, it runs to a timetable, and nobody has ever enjoyed a minute of it.

The work is knowing what was asked for, whether it's arrived, what the last reminder said, and what the next one should say. The first reminder is routine. The fourth one usually isn't, because by then something else is going on.

So the sequence gets approved once, by a person, and the exceptions come back to a person. Good automation knows exactly where the routine ends, and that boundary is a decision the firm makes rather than one the software works out on its own.

The documents

7. Assembling a proposal

A proposal is part assembly and part commercial judgement, and firms tend to get this wrong in both directions at once: treating the whole thing as bespoke, or reaching for a template and sending something that reads like everyone else's.

The assembly is the automatable half. The approved description of the firm, the relevant credentials, the standard delivery language, the right examples, and whatever the prospect told you. A first draft can be complete and still leave the choices marked as choices.

Scope, price, risk, and what you're promising stay with a partner. So does the voice.

8. Drafting engagement letters

Repetitive, consequential, and built from wording somebody already approved. That combination makes engagement letters suitable for controlled drafting and completely unsuitable for anything looser.

The right version of this populates an approved template from the agreed matter details, picks clauses by documented rules, and shows every single departure from the standard. It never invents a term because a field was empty.

Someone authorised reads the scope, the fees, the obligations, the exclusions, and every non-standard clause. If the firm can't see which approved clause set produced the draft, it isn't ready to be doing this job yet.

9. Drafting routine client correspondence

Plenty of client letters and emails have a recognised purpose and shape without being identical: the acknowledgement, the request for information, the progress note, the confirmation of a position already agreed.

Drafted from the matter record and the firm's own precedents, with the sourced facts kept separate from the proposed wording, and anything it can't support flagged rather than smoothed over.

A person approves every send. Advice, commitments, complaints, and anything with feeling in it don't become routine just because they arrived as email.

The reporting

10. The weekly or monthly client report

The report matters. Most of the labour that produces it doesn't need a decade of expertise.

The figures have to be collected. The notes have to be found. The same headings get filled in again. What changed since last time has to be worked out. Then somebody loses Friday afternoon to pulling it into shape.

Assembling the inputs, comparing the periods, drafting the recurring sections, and flagging what's missing or doesn't reconcile is a good AI job with a fixed rhythm, which means the time comes back every single week. The owner still writes the interpretation and approves what circulates. On my own practice this is the one I'd miss most if it went away, because the version waiting on a Thursday night is already checked and structured, and looking it over is a different job from writing it.

11. Producing recurring internal reports

Pipeline summaries, utilisation, aged work lists, and the management pack. Fixed patterns, produced monthly, usually eating the time of the one person best placed to interpret them. Collect the approved inputs, apply the agreed structure, draft the commentary on what moved, and put the anomalies in a queue. A named owner still checks the numbers before anyone decides anything on the back of them.

One warning, because it's the common failure here. If three partners mean three different things by "active", settle that first. Automating an argument just makes it arrive faster.

12. Preparing project and status updates

Same assembly every week: what's finished, what's open, what's at risk, what needs a decision, and what's next. Gathered from the systems people already update, compared against last week's, with the stale items and the missing owners surfaced rather than quietly carried forward.

The project lead still judges whether a risk is real and whether the status somebody typed in is honest. Take the labour out of reporting by all means. Nothing in that automation should be capable of turning a difficult project green.

The upkeep

13. Updating the CRM

Small jobs, done badly, across a whole firm. A call ends, everyone moves to the next thing, and the record waits until Friday, when nobody can remember who said what.

A structured update prepared from the approved note gets you who was involved, what changed, the next action, the date, and where it came from. The straightforward entries go through; conflicting details and anything touching important client data go to a review queue. What you're after is a record people actually trust, which is a much higher bar than a fuller one.

14. Capturing what the firm knows

Every firm knows more than its shared drive suggests. The good method is in one person's notebook. The reason for that clause is in an email from four years ago. The lesson somebody learned the hard way is remembered right up until they leave.

Approved notes, finished work, and conversations about how a process runs can be turned into draft knowledge records with titles, summaries, source links, and suggested connections to what's already there. The people who know the work review it before it becomes the firm's approved version. The Archivist does this in my practice, and the Structuralist writes down the methods, which is how a process outlives the person who invented it.

This isn't an invitation to hoover up everything you own. Knowledge with no owner and no way of retiring it is just a bigger pile.

15. Preparing a morning priorities brief

What's due, what changed overnight, what's stuck, and what needs a decision from you. Just as usefully, it leaves out the hundred things that don't need you today.

The assembly is a good AI job. Priority isn't. A system can show you deadlines, dependencies, and what's still unanswered. It can't know, unless you tell it, that a quiet client is worth more of your attention this week than a loud internal deadline.

The best brief is short, sourced, and easy to argue with. It helps you decide where the day goes rather than deciding it for you.

Five jobs I wouldn't hand over

There's a difference between using AI in the preparation and making it answerable for the job. These are the ones I'd keep on your side of the line.

1. The advice itself

An agent can gather the record, test a draft against an approved checklist, or show you the issue from an angle you hadn't taken. The advice belongs to the qualified person who understands this client, this context, and what happens if it's wrong. "The system produced it" isn't an approval and it won't hold up as one.

2. The first reply when a client is unhappy

A complaint, a dispute, a bereavement, bad news, or a conversation about a bill. Those need more than the correct words in the correct order. Prepare with AI by all means. The reply belongs to whoever holds the relationship, and it should read like it came from them.

3. Anything written about a person

Reviews, references, notes about performance, anything that ends up on someone's file. Firms are full of people, and the moment the subject of the sentence is a colleague rather than a task, it stops being administrative work. This is the same principle the Census runs on: count the work, never the people.

4. Work that fails quietly

A limitation date. A filing deadline. A calculation nobody checks because it's always been right. If getting it wrong wouldn't be noticed for six months, the review that catches it is worth more than the ten minutes automating it saves. The jobs on my list above all fail loudly.

5. Deciding what to hand over in the first place

A tool shouldn't get to choose its own work. The people doing the job know where it catches and which exception breaks the rule. Management knows what the firm is trying to protect. Both need to be in the room before anything gets picked, which is most of why the Census is a day with your staff around a table rather than an afternoon with me and a laptop.

The loudest job is rarely the heaviest

Fifteen jobs, and they're not worth the same in your firm.

A ten-minute task that thirty people do every day will beat a monthly report nobody enjoys. A job that looks simple can hold one exception that changes the risk entirely. And a process everyone wants rid of is sometimes just badly designed, in which case fixing it costs a fraction of automating it and works better afterwards.

So count before you choose. How often does it happen? How long does it really take, including the hunting and the correcting around the edges? Whose time goes on it? What doesn't get done while it's being done? How predictable is the method? Where does judgement come into it? What happens if the first draft is wrong, and who'd notice?

That count reorders the list nearly every time. The thing people complain about loudest is hardly ever the heaviest thing they carry.

And a thing the counting only shows you later: the first automation is a project. By the fifth, you have a small operating environment, and it needs an owner — somebody who knows what's running, what it costs, and what should change next. Which kind of help owns that is its own question.

Want real numbers on one of the fifteen? The Repetitive Work Calculator is free, takes about two minutes, and runs the same five-way sort.

The jobs worth looking at first are usually the ones that:

• happen often;
• take more time than anyone in the room thinks;
• follow a method that holds up when you write it down;
• can be checked quickly by whoever knows the work;
• and need very little judgement, right up until an exception turns up.

When one of those five isn't true, the answer is usually assistance, a process fix, or leaving it with a person, rather than automation.

Don't know which of these is costing your firm the most?

That's what the Mercurious Census, a one-day count of the recurring work across a firm, is designed to find out.

One facilitated day inside your firm, no laptops in the room, and a printed roll down the length of the boardroom table. Your people write the recurring work on it in their own handwriting, we count how often it happens and weigh what it costs, and then we run every job through the five-way sort: automate it, assist it, fix the process, keep it human, or leave it alone for now. Often enough the answer to a good chunk of it is the fourth or the fifth one.

Nobody signs anything on the day. The day ends with a direction, the written AI Roadmap follows within 48 hours, and the roll stays with you whatever you decide.

How the Census works →

When it's working, the firm doesn't feel automated. It feels like the report is already in the drafts folder on Thursday night, checked and ready to look over. The brief for tomorrow's client meeting is sitting there with the last three sets of notes in it and nobody stayed late assembling it. The chasing went out without anyone having to steel themselves for it.

And the person who used to spend a third of the week on all that is back on the work your clients are actually paying for, which is the part they came to your firm for in the first place.

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