Walk into any mid-sized Irish accounting practice and you'll find the same picture: a partner with twenty years of client knowledge in their head, three managers carrying the weight of compliance season, a shared drive that nobody fully understands, and a working folder structure that made sense to someone in 2017. The intelligence of the firm is real — it's just not retrievable. When the partner is on holidays, or retires, or a senior leaves for industry, a measurable chunk of that firm walks out the door. That's the problem an intelligence brain is built to solve, and it's why I think the Irish profession is going to adopt this faster than most people expect.
The shape of an Irish practice is unusual, and generic AI doesn't fit it
Most of what gets sold as "AI for accountants" is built around the US or UK market. It assumes a tax code that isn't ours, a regulator that isn't ours, and a client mix that isn't ours. Irish practices sit in a specific shape: a long tail of owner-managed companies, a heavy concentration of agricultural and construction clients in rural firms, sole traders crossing the VAT threshold every year, and an increasing number of cross-border structures driven by inbound FDI and the post-Brexit reshuffle. Add the Revenue Online Service, iXBRL filing, the Companies Registration Office, and the specific quirks of how Irish SMEs actually keep their books — half on Sage, a quarter on Xero, and the rest on Excel and goodwill — and you have an environment that off-the-shelf tooling handles badly.
An intelligence brain isn't a chatbot bolted onto your practice management system. It's a layer that sits across the documents, emails, working papers, prior-year files, and partner notes that already exist, and makes them queryable in the way a senior manager's memory is queryable. The difference is that the brain doesn't leave at five o'clock and doesn't take a job in industry next March.
Why on-premise matters for accounting specifically
I'll be blunt: if your AI tool is sending client trial balances, draft accounts, or payroll data to a cloud endpoint owned by a US hyperscaler, you have a problem under your engagement letters and probably under your professional body's confidentiality rules. CCAB-I, ACCA, and Chartered Accountants Ireland all expect you to know where client data is processed. "It's in the cloud somewhere" is not an answer that survives a regulatory inspection.
This is the technical reason I built the brain to run on-premise or in a tenant the firm controls. The model weights, the vector index, the document store, and the inference all sit inside the practice's boundary. Nothing leaves. When a manager asks "what did we conclude on the R&D claim for client X in the last cycle?", the question is answered by a model running on hardware the firm owns or rents directly, against an index built from the firm's own files. There's no third-party training, no data exhaust, no surprise terms-of-service change that suddenly puts client confidentiality at risk.
The engineering trade-off is real. On-premise means you don't get the absolute frontier model. What you get instead is a model that's good enough for the work — extraction, summarisation, retrieval, drafting — and complete control over the data path. For accounting, that's the right trade.
What the brain actually does inside a practice
Strip away the marketing language and the technical work breaks into four layers.
Ingestion. The brain reads the firm's existing document estate — working paper files, signed accounts, correspondence, engagement letters, tax computations, bank statements, payroll registers — and builds a structured index. For most Irish firms, this is a mix of PDF, Word, Excel, and Outlook .msg files going back a decade or more. The ingestion layer has to handle scanned documents (a lot of older audit files are just scans of scans), handle Irish-specific document types like Form 11s and CT1s, and preserve the relationship between a document and the client, the year, and the engagement.
Extraction. Once a document is in, the brain pulls structured data out. From a set of accounts: turnover, gross profit, key ratios, directors, auditor's report wording. From a tax computation: adjustments, balancing allowances, losses carried forward. From an engagement letter: scope, fee basis, partner. This is where domain knowledge matters — a generic extractor will miss that "Section 110" means something specific, or that a "close company surcharge" calculation needs particular fields.
Retrieval. The interesting part. When someone asks a question, the brain has to find the right context across potentially millions of documents and assemble it into something the model can reason over. This is retrieval-augmented generation, but tuned for accounting work: the retrieval has to respect client boundaries (you never want client A's data leaking into a query about client B), respect time (last year's treatment may not apply this year), and respect confidence (if the brain isn't sure, it should say so and cite the source).
Drafting. The output layer. Drafting a first-pass response to a Revenue query, summarising prior-year audit findings before fieldwork, producing a client briefing note, drafting a section of a report. The brain doesn't sign anything — a partner does. But it removes the cold-start problem from every piece of written work in the firm.
The compliance season problem, and why retrieval beats memory
Every Irish practice has the same November bottleneck: the income tax filing deadline, the rolling CT deadlines, and the run-up to the December and March year-ends. During those weeks, the limiting factor isn't software — it's how fast a manager can recall what was done last year for a particular client.
"Did we claim the Employment Investment Incentive for them?" "What was the rationale for the directors' loan treatment?" "Why did we change the depreciation policy in the 2022 accounts?" These questions get answered today by someone scrolling through a client folder for fifteen minutes, or by interrupting the partner.
A retrieval system answers them in seconds, with citations. The technical reason this works is that accounting work is overwhelmingly about precedent — what did we do before, and is the situation materially the same? That's a near-perfect match for vector retrieval over a well-indexed corpus. The brain doesn't need to be smarter than your senior manager; it needs to remember more reliably than your senior manager, and surface the right prior-year working paper with the right context.
If you want to see how this is structured for a practice specifically, I've written more about the architecture for accounting firms and what a deployment actually looks like.
Risks I'd want a partner to ask about before buying anything
I'd be suspicious of anyone selling AI into accounting who doesn't volunteer the failure modes. Here are the ones that matter.
- Hallucination on numbers. A model will confidently produce a turnover figure that doesn't exist. The brain has to be designed so that any numeric output is either retrieved verbatim from a source document with a citation, or flagged as a calculation. Don't accept a tool that lets the model freestyle on figures.
- Client data segmentation. If the index isn't partitioned correctly, a query about one client can pull context from another. This is a confidentiality breach by design. Test it on day one.
- Audit trail. Every query, every response, every source cited needs to be logged. If a partner relies on a brain-generated summary and it turns out to be wrong, you need to be able to reconstruct what the brain saw and what it said.
- Model drift. When the model is updated, behaviour changes. The firm needs a way to test that updates don't degrade quality on the kinds of queries it actually runs.
- Off-boarding. If you stop paying, what happens to the index? Can you export it? Can the vendor still read it? On-premise solves most of this, but you should still ask.
Where this fits in the bigger picture
The intelligence brain isn't accounting-specific. It's a general pattern — an organisational memory layer that runs inside the boundary of a regulated firm — and the same architecture applies to legal practices, medical groups, and parts of the public sector. The accounting application is one of the cleanest because the document estate is already structured, the questions are already precedent-driven, and the regulatory pressure to keep data in-country is already there. If you want the broader picture of how the platform is built and what it's for, the overview is here.
The reason I think Irish practices will adopt this ahead of larger UK or US firms is unglamorous: the firms are smaller, the partner has the authority to make the decision in a single conversation, and the pain of losing institutional knowledge when a senior leaves is felt directly by the people who own the business. The economics are immediate.
Where to start this week
If you run or sit on the management team of an Irish practice and any of this resonates, the useful thing to do this week is small and concrete. Pick one client — ideally a long-standing, complex one — and write down every question a new manager would need answered to take that client over cleanly. Then walk through your own systems and time how long it takes to answer each question from existing documents. That exercise tells you, more honestly than any sales pitch, where the institutional memory of your firm actually lives, where it's fragile, and how much of it is in one person's head. Once you've seen the gap, the conversation about what to do about it becomes much easier. Email me directly if you want to talk through what a deployment would look like for your firm — I'd rather have a thirty-minute call than send a brochure.