A managing partner in a Midlands firm asked me a fair question over coffee: could a small AI swarm replace one of the seniors who had just left? I told him that was the wrong way to frame it. The right question is which parts of which roles a swarm can absorb, and which parts you should keep human on principle. We sat down and did the audit. What follows is the shape of that work — twelve people, six full-time roles in scope, and three places we said no.

The firm and the brief

The firm is typical of Irish professional services: twelve staff, partner-led, mix of compliance and advisory, clients across Tipperary and the south-east, a small Dublin book. Two partners, four seniors, four juniors, two on the office side. Margin under pressure. Hiring hard. The brief was simple — work out where AI productivity in Ireland actually lands for a firm of this size, without burning the practice down to find out.

I have a bias I should name up front. I spent twenty-odd years inside Tesco, Dunnes Stores and Oracle before IMPT. Big-company process. I know what real automation looks like and I know what dressed-up nonsense looks like. SME AI rollout is mostly the second thing at the moment. So we were careful.

How we ran the audit

We took each of the six in-scope roles and broke a normal week into tasks. Not job descriptions — tasks. Then for each task we asked four questions:

• What is the input, and is it already in a form a model can read?
• What is the output, and who signs it off?
• What happens if it's wrong — client harm, regulatory harm, or just rework?
• How often does the task repeat, and does it repeat the same way?

That last one matters more than people think. Role automation works on shape, not on volume. A task you do fifty times a week with fifty different shapes is harder to automate than one you do five times a week the same way. We sorted every task into swarm, assisted, or human-only. The swarm bucket is where a small group of agents — a reader, a drafter, a checker, a filer — handle the task end to end with a human approving the result. Assisted is where a person still drives, but the model removes the boring half. Human-only is what it sounds like.

Role one: the bookkeeping desk

This was the obvious one and it still surprised us. Bank feed reconciliation, supplier invoice coding, expense capture, month-end packs — the bulk of a junior bookkeeper's week is shape-stable and low-stakes per item. We put a swarm on ingestion and coding, with a human review step at month end. The gain was not that the work got done faster. It was that the senior who used to spend Friday afternoons untangling a junior's coding errors got Friday afternoons back.

The honest read on AI ROI here: the saving is not the junior's salary. The junior is still there, doing better work. The saving is the senior's attention, which is the actual scarce resource in any Irish professional services firm.

Role two: compliance drafting

Annual returns, CRO filings, statutory letters, engagement renewals. High volume, high template-density, low creative content. A drafting agent with access to the firm's prior filings and the relevant statutory shape produces a first cut that a senior corrects in minutes rather than composes in an hour. The checker agent runs the draft against a list of the firm's own historical mistakes — which turned out to be the most useful single thing we built, because every firm has a private catalogue of recurring slip-ups and nobody writes it down.

We did not let the swarm file anything. Submission stays human. That line matters and I'll come back to it.

Role three: client correspondence and triage

The inbox is where small firms drown. We put a triage agent on the shared mailbox that classifies, summarises, drafts replies for routine items, and flags anything with a deadline, a complaint marker, or a Revenue reference. It does not send. A person sends. But the partner who used to start every morning by reading two hundred emails now starts by reading twenty summaries and approving fifteen drafts.

This is the area where Irish professional services AI gets oversold the most, by the way. The vendors will tell you the agent can run the inbox. It cannot. It can prepare the inbox for a human to run in a fifth of the time. That is already a serious win. Don't ask for more than that yet.

Role four: research and advisory prep

When a client rings about a transaction, a restructuring, a Revenue query, a senior used to spend two or three hours pulling background — prior file, sector notes, legislation, recent guidance. A research swarm now produces a briefing pack before the senior sits down. The senior reads it, throws out what's wrong, adds what's missing, and meets the client. The meeting is better because the senior went in with the file already in their head rather than building it during the call.

One caveat we wrote into the operating manual in red: the agent can pull material, but the legal and tax position in the advice itself is the senior's, full stop. The model is a librarian, not a counsel. If you cannot tell the difference between those two things in your own firm, do not start with this role.

Role five: practice management and scheduling

Booking, rescheduling, deadline tracking, WIP chasing, timesheet nudging. Pure shape-stable admin. A swarm handles it with a calendar and a practice management tool, and the office manager moves up the value chain into client onboarding and supplier negotiation — work that was always meant to be hers and never got done because the diary ate her week.

Role six: marketing and BD

Newsletter, LinkedIn presence, case study writing, pitch decks for tenders. A drafting swarm with the firm's tone of voice baked in produces first drafts. A partner edits. The cadence went from monthly-when-we-remember to weekly-without-thinking. Whether that wins new work is a longer story than this audit can tell, but it costs almost nothing to maintain now, which was the point.

The three places we said no

This is the part most write-ups skip. An honest SME AI rollout has refusals in it.

One: we did not put a swarm on tax advice itself. Drafting around it, yes. Research around it, yes. The position taken on a client's behalf, no. Not because the model can't read the legislation — it reads the legislation fine. Because the responsibility for getting it wrong sits with a named human under Irish professional rules, and that human needs to have actually thought about it. Outsourcing the thinking to the agent and signing the output is the worst of both worlds. You get the model's confidence without the human's caution.

Two: we did not automate the difficult client conversation. The fee increase, the bad-news call, the complaint, the disengagement letter. The agent can prepare a script. A partner makes the call. Firms that try to send a model into emotionally weighted conversations lose clients they should have kept, and the loss is invisible for about six months until renewal season makes it visible all at once.

Three: we did not connect the swarm to anything that moves money or files with the State. Read access, yes. Write access to the bank, to Revenue Online, to CRO — no. The cost of a wrongly-filed return or a wrongly-paid supplier dwarfs any productivity gain you'll book in a year. Keep the human in the loop on the irreversible actions. This is not a temporary position while the tech matures. This is the design.

What twelve people running six roles actually looks like

After the rollout, the headcount did not change. The shape of the week did. The juniors stopped doing the boring half of their job and started doing more of the senior work, supervised. The seniors stopped doing the boring half of their job and started seeing more clients. The partners got an evening back, which they mostly used to take on work the firm had previously turned away. Capacity went up without anyone working longer hours. That is the version of AI productivity in Ireland I think is real. It is not the version where the firm gets smaller. It is the version where the firm takes on the work it had been quietly declining.

The gains were uneven across the six roles. Bookkeeping and practice management gave the largest hour savings. Advisory prep gave the largest quality lift. Marketing gave the largest cadence change. Compliance drafting and correspondence sat in the middle on both. If a firm asked me where to start with one role, I would say practice management, because it pays back fastest and it builds the muscle for the harder rollouts later.

What to do this week

If you run a firm of this size, do not buy anything yet. Take a week and do the task-level audit on your own people. Sort every task into swarm, assisted, or human-only, with the four questions above. Be honest about which tasks are actually shape-stable and which only feel that way. Then pick one role — not six — and run a four-week pilot on the swarm tasks only, with a senior reviewing every output. That is the work we are doing now at IMPT alongside the platform build, because the same audit logic sits behind the AI-native booking agent we are bringing into IMPT.io. The firms that get this right in 2026 will not be the ones who moved fastest. They will be the ones who refused most carefully.