AI for accountants in Ireland: what's actually worth adopting

A practice manager at a six-partner firm told me recently that her team had "basically already done AI" — someone had used ChatGPT to draft a client letter, thought it was clever, and that was that. Multiply that story across most small and mid-sized Irish practices and you have an accurate picture of where the profession actually stands: individual curiosity, no structure, and a genuine uncertainty about what's safe to do with client data in a tool built by an American software company.

That uncertainty is reasonable. Accountants handle exactly the kind of information — financial records, PPS numbers, banking details, commercially sensitive figures — that makes "just paste it into ChatGPT" a genuinely bad instinct without a considered policy first. This piece is about what's actually worth adopting, in what order, and what still needs a qualified person doing the judgement.

Start with what you're already paying for

Before any new tool enters the conversation, it's worth being honest about what's sitting unused inside software you already licence.

Xero, Sage, and most modern practice-management platforms have shipped AI features into their existing subscriptions over the past two years — automated bank reconciliation matching, receipt-line-item extraction, anomaly flagging on unusual transactions. Most firms are paying for these and not using them, either because nobody told the team they existed or because turning them on felt like a project rather than a toggle.

This is genuinely the highest-value first step for most practices, and it's the one almost nobody takes, because "you're already paying for this, you're just not using it" isn't a sentence that sells anything. It's also the honest starting point of the Four Levels framework we use in every Assessment:

• Level 1 — use what's already licensed. The AI features inside Xero, Sage, or your existing practice software, switched on properly.
• Level 2 — one considered standalone tool for the specific work native features can't touch — drafting client correspondence, summarising meeting notes, first-pass analysis of a set of accounts.
• Level 3 — connect what you have so outputs land where the work actually happens, rather than being copied and pasted between a chat window and your real systems.
• Level 4 — custom implementations — genuinely valuable for specific cases, mostly more than a small practice needs.

Most practices should spend most of their attention at Level 1 and 2. Almost nobody starts there, because Level 4 gets the marketing budget.

What's genuinely safe to hand to AI right now

First drafts of routine client correspondence. Engagement letters, standard queries, meeting follow-ups. The judgement and the final sign-off stay with a qualified person; the blank-page problem is what AI removes.

Summarising long documents before a human reads them properly. A 40-page lease or a year of bank statements condensed into a first-pass summary saves real time, provided the summary is treated as an index into the source material, not a replacement for reading it.

Drafting internal process documentation. Month-end close procedures, VAT return checklists, client onboarding steps — the kind of process knowledge that currently lives in one senior person's head and nowhere else. A recorded walkthrough turned into a written SOP is a genuinely undervalued use of an afternoon.

First-pass anomaly flagging, provided a qualified person reviews every flag before anything is actioned. AI is good at "this looks unusual, check it" and bad at deciding whether unusual means wrong.

What still needs a qualified person, full stop

Final judgement on anything that goes to Revenue, a client, or an auditor. AI drafts, checks, and flags. It doesn't sign off. This isn't caution for its own sake — the professional and regulatory liability sits with the practitioner, not the tool, and treating AI output as finished work rather than a first draft is where firms get into genuine trouble.

Client-specific tax or advisory judgement. General AI tools are trained on general information, not on the specific facts of a client's situation, and they will produce confident, plausible, sometimes wrong answers to specific technical questions. Useful for a first orientation on a topic, not for the actual advice.

Anything involving client PII pasted into a consumer AI account with no data processing agreement. This is the point worth taking seriously rather than treating as boilerplate: a personal ChatGPT or ad-hoc AI account has no enterprise data agreement, no assurance about training-data use, and no audit trail. If your team is already pasting client financials into personal accounts — and in a lot of practices, they are — that's a live GDPR exposure worth addressing before anything else on this page.

The data protection question, properly

This is where a lot of general AI advice goes vague, so it's worth being specific. Under GDPR, client financial data is personal data, and a practice has an obligation to know where it's going once someone pastes it somewhere. The practical fix isn't "ban AI" — that just pushes the behaviour underground — it's a short written policy: which tools are approved for client data, which tier of subscription (enterprise/business tiers of most major AI tools carry proper data agreements; free consumer tiers generally don't), and a plain instruction that anything containing client PII goes through the approved tool, not whatever's fastest in the moment.

Four rules, written down, that a team can actually remember, is worth more than a fifteen-page policy nobody reads. If you want the fuller version of this, the GDPR and AI tools piece on the Courses hub sets out the same four-rule structure in more depth.

A realistic first 30 days

For most practices, the sequence that actually works is: audit what AI features are already switched on (or off) inside your existing software, write the four-rule data policy before anyone does anything else, pick one recurring task — client onboarding, month-end close, a specific correspondence type — and document it properly with AI assistance, then measure whether it actually saved time before deciding what's next.

That's a smaller first step than most AI-for-accountants content suggests, deliberately. The firms that get real value are the ones that fix one thing properly rather than the ones that bought five subscriptions in a week.

If you want a structured read on where your own practice actually sits against this — not a sales call, a genuine diagnostic — the free Scorecard takes five minutes and tells you honestly whether you're at Level 1, 2, or further along, and what the sensible next step actually is.