Most Irish SMEs I talk to are sitting on more institutional knowledge than they realise — and almost none of it is searchable. It lives in Outlook archives, in a finance manager's head, in a folder on someone's desktop called old_quotes_DO_NOT_DELETE, in WhatsApp threads with the foreman, in a SharePoint nobody trusts. When the person who knows leaves, retires, or takes a fortnight in Lanzarote, the business slows down. That's the actual problem an intelligence brain solves for a small Irish business. Not "AI transformation". Not chatbots. The plain operational problem of a 30-person firm forgetting what it already knows.
What an intelligence brain actually is, in SME terms
Strip away the marketing and an intelligence brain is three things stitched together: a private index of your own documents and data, a retrieval layer that finds the right bits when asked, and a language model that answers in plain English with citations back to the source. That's it. The reason it works for a small business — a builder's merchant in Thurles, an accountancy practice in Galway, a manufacturing SME in Shannon — is that the corpus is bounded. You don't need a model that knows everything on the internet. You need a system that knows your quotes, your supplier terms, your previous jobs, and your compliance files.
The technical pattern is retrieval-augmented generation, but the engineering decisions that matter for an Irish SME are different to the ones that matter for a multinational. You don't have a data team. You don't have a Kubernetes cluster. You probably have Microsoft 365, maybe Sage or BrightPay, possibly a line-of-business app written in 2009 by someone's nephew. The brain has to meet that estate where it lives, not demand a rebuild.
Why generic cloud AI doesn't fit a small Irish business
The default suggestion an SME owner gets is "just use ChatGPT" or "turn on Copilot". Both have their place. Neither is an intelligence brain. Here's the gap.
A general-purpose chatbot doesn't know your business. Ask it about a quote you sent last March and it'll either decline or hallucinate. Copilot is closer — it can read your tenant — but it's a productivity tool wrapped around individual users, not an organisational memory. It answers from your mailbox and your files, not from the firm's collective record. Two staff asking the same question can get different answers because they have different access scopes. That's fine for drafting an email. It's not a system of record.
The second issue is data residency and control. For a regulated SME — and "regulated" in Ireland now reaches further than people think, with NIS2, GDPR enforcement, and sector rules in financial services, healthcare, and legal — sending the contents of client files to a US-hosted consumer endpoint is a problem you don't want to find out about at audit. An on-premise or EU-hosted brain, with the model and the index inside a perimeter you control, removes that argument before it starts.
The third issue is cost shape. Per-seat AI licensing scales with headcount. Knowledge in an SME doesn't. A 25-person firm has roughly the same volume of institutional memory as a 60-person firm in the same trade. Paying per seat to access shared knowledge is the wrong unit of billing for a small business.
The architecture I'd build for a 20–80 person firm
Here's the shape I actually deploy. It's deliberately boring, because boring is what survives a Tuesday morning when the broadband flaps.
- Ingestion layer. Connectors to Microsoft 365 (SharePoint, OneDrive, Exchange), to a local file share if there's still a Windows server in the corner, and to whatever line-of-business systems matter — Sage, Surf Accounts, BrightPay, a CRM, a job-costing tool. Ingestion runs on a schedule, picks up changes, and writes to a staging area.
- Normalisation and chunking. Documents are parsed — Word, PDF, Excel, scanned PDFs through OCR, emails with their attachments, exported reports — and broken into chunks that preserve context. A quote with line items isn't useful if you split it mid-table. This stage is where most cheap implementations fall over.
- Embedding and indexing. Each chunk is embedded using a model that runs locally or on EU infrastructure, and stored in a vector database alongside metadata: source, author, date, sensitivity label, access scope.
- Retrieval and reasoning. When a question comes in, the system retrieves the most relevant chunks, applies the user's permissions, and passes the filtered context to the language model with the original question. The model answers and cites which documents it used.
- Audit and feedback. Every query, every retrieval, every answer is logged. Users can flag wrong answers. That feedback drives index tuning.
For most SMEs this fits on a single beefy server or a small EU-hosted VM. You don't need GPUs the size of a fridge. You need a system that's cheap to run on a Wednesday when nobody's asking it anything.
The use cases that pay for themselves
I've stopped pitching "AI for SMEs" in the abstract because owners glaze over, fairly. What lands is specific, ugly, recognisable problems.
Quote and tender history. "What did we charge Murphy Construction for the last fit-out, and what was our margin?" Right now that question takes someone twenty minutes of digging through a SharePoint and a Sage export. The brain answers in seconds with the source documents attached.
Onboarding and handover. When a senior estimator, bookkeeper, or office manager leaves, the institutional memory walks out with them. With a brain in place, the replacement can ask "how have we historically handled credit terms for new trade customers?" and get an answer drawn from actual past correspondence and policy notes.
Compliance preparation. Audits — GDPR, ISO, sector-specific — turn into evidence-gathering exercises. The brain can pull every policy, every signed acknowledgement, every relevant email thread, and present them with citations. The work that used to consume a week of someone's time becomes an afternoon.
Supplier and contract intelligence. "Which of our supplier contracts auto-renew in the next ninety days, and which have price-review clauses?" That's a question most SMEs can't answer at all today. With contracts indexed and structured, it's a query.
Internal Q&A. "What's our policy on remote working?" "How do I expense mileage?" "Where's the template for a new starter contract?" The brain replaces the Slack message that interrupts the office manager four times a day.
None of this is glamorous. All of it is measurable. If you want to see how this maps to specific verticals — including the SME shape — I've written more on the small-business intelligence brain page.
What you actually need to have ready
The honest precondition for any of this working is that your data isn't completely feral. You don't need it to be perfect. You do need it to be findable and reasonably well-labelled. The discovery exercise I run on day one looks like this:
- Where does each document type live? Quotes, contracts, invoices, HR files, policies, project records. List the systems and the file paths.
- Who owns each system? Not "IT" — the actual person who knows where the bodies are buried. Often the office manager or finance lead.
- What's the access model? Who can see what, and is that written down anywhere or just in someone's head?
- What's the sensitivity gradient? Public, internal, restricted, regulated. Even a rough cut is enough to start.
- What are the top ten questions the business asks itself repeatedly that nobody has a fast answer to? Those are your first use cases.
This is unglamorous, and it's where most AI projects die. Not in the model selection. Not in the vector database. In the bit where someone has to actually look at the SharePoint and decide what's still relevant.
The control question: on-premise, hybrid, or EU cloud
For an Irish SME, the deployment choice comes down to three options and a set of trade-offs that are genuinely about your business, not about technology preference.
On-premise — the brain runs on a server in your building. Maximum control, no data leaves the perimeter. Suits firms with regulatory exposure, sensitive client data, or a cultural preference for owning the kit. Higher upfront commitment, but predictable running costs.
EU-hosted private cloud — the brain runs on infrastructure in Ireland or another EU member state, dedicated to your business. Data residency satisfied, no shared tenancy with strangers, easier to scale, lower upfront cost. The middle path most SMEs land on.
Hybrid — sensitive corpora stay on-premise, less sensitive workloads run in the EU cloud. More moving parts, but appropriate for firms with mixed data classes.
The thing I'd push back on is the assumption that on-premise is automatically harder or more expensive. For a single-site SME with decent connectivity and a competent partner, it's often the simpler answer. The model runs locally, the index sits beside it, and there's no monthly cloud bill that creeps up every quarter. More on the broader architecture choices on the intelligence brain overview.
Where to start this week
If you run a small Irish business and this resonates, don't go shopping for AI products yet. Do this instead,