Most SME deployments fail in the first fortnight, not the third month. The owner buys into the idea, the IT person nods along, and then nothing happens because nobody decided who owns the box, where it sits, or what data it's allowed to see on day one. The Brain is designed to be installed by a competent generalist — not a vendor team flown in — but that only works if you sequence the work properly. What follows is the actual rollout I run with Irish SMEs of roughly 10 to 80 staff, broken down week by week. No stages skipped, no marketing layer over the top.
Week 1: Decide what the Brain is allowed to know
Before anything is installed, you sit down for two hours and write a one-page data scope document. This is the single most undervalued step in any SME AI implementation. The document lists every shared drive, mailbox, accounting system, CRM, and folder on the network, and against each one you write either "in scope", "out of scope", or "in scope, redacted". It is not a security policy. It is a map of what the model will eventually be able to retrieve.
For a typical Irish SME — say a distributor in Limerick or a services firm in Cork — the in-scope list usually settles around: the shared drive's operational folders, the Outlook shared mailboxes for sales and support, the Sage or Xero export, and the project folders. Out of scope almost always includes HR, payroll, legal correspondence with solicitors, and the directors' personal subfolders. The redacted category is for things like supplier contracts where the commercial terms matter but the counterparty's bank details do not.
You also decide, this week, who the three named humans are: the data owner (usually the MD or operations director), the technical operator (the person who will actually run commands on the box), and the escalation contact (often an external accountant or solicitor). Write the names down. The Brain logs decisions against these roles, and if they're vague, the audit trail is vague.
Week 2: Hardware, network, and the boring infrastructure conversation
The Brain runs on-premise. For an SME that means one machine, sitting in your comms room or under a desk in the server cupboard, with a wired connection and a UPS. The minimum spec I recommend is a workstation-class box with a recent consumer GPU with at least 24GB of VRAM, 64GB of system RAM, and 2TB of NVMe storage. You can go larger, and for firms above about 50 staff you probably should, but this is the floor for a real deployment rather than a toy.
Networking is where week 2 either goes smoothly or eats three days. The Brain needs:
- A static internal IP, reserved on the router or DHCP server.
- An internal DNS entry — something like brain.yourcompany.local — so users aren't typing IP addresses.
- Outbound HTTPS for model and software updates, which can be gated through your firewall to a small allow-list.
- No inbound exposure to the internet. None. If someone suggests publishing it for remote staff, you put a WireGuard or Tailscale tunnel in front of it instead.
Backups are decided this week too. The Brain's index and configuration live in a single directory tree, and that directory should be in your existing backup rotation from day one. If you don't have an existing backup rotation, that's a separate problem you fix before going any further.
Week 3: Install, index a small slice, and break things deliberately
Installation itself is a single afternoon. You bring up the base OS — Ubuntu Server LTS is what I run — install the GPU drivers, pull the Brain bundle, and run the bootstrap script. The bootstrap creates the service user, sets up the local vector store, configures the model runtime, and brings up the web interface on the internal address you reserved last week.
The temptation now is to point it at every shared drive and let it index for forty-eight hours. Don't. You pick one folder — typically the operations or projects folder — and index that alone. A few thousand documents, not a few hundred thousand. This gives you a working system within an hour and lets you see, with real internal content, whether the retrieval is finding what staff would actually expect it to find.
This is also when you deliberately try to break it. Ask it questions about a folder you didn't index — it should refuse cleanly and say it has no source. Ask it about a document you know contains a specific clause and check the citation points to that document and not a similar one. Ask it the same question three times and confirm the answers are consistent. If any of these fail, you fix them now, on a small index, not after you've ingested the whole estate.
Week 4: Expand the index and onboard the first five users
Once the small slice behaves, you widen the indexer. The right order is: operational folders first, then shared mailboxes, then the accounting export, then the CRM. Each one is added as a separate source so you can see in the logs which corpus a given answer came from. Indexing a typical Irish SME's full document estate takes somewhere between a few hours and a couple of days depending on volume and how messy the file structure is.
Users come on in a deliberately small group. Five people, chosen because they represent different parts of the business — one from sales, one from operations, one from finance, one from the management team, and one person who is openly sceptical. The sceptic is essential. They will find the failure modes the enthusiasts skip past.
You give them three things: the internal URL, a one-page guide on how to phrase questions and read citations, and a shared channel where they post anything weird. You don't give them training videos. You don't run a workshop. The interface is a chat box; if it needs a workshop, the rollout is already in trouble.
Week 5: Tune retrieval and start measuring
By week 5 you have real usage data. The logs show what people are actually asking, which sources are being hit, where the model is refusing because it has no grounding, and where the citations look thin. This is where the work shifts from installation to tuning, and it's the part most vendors skip because it's unglamorous.
Three things to look at:
- Chunking. If answers are missing context that's clearly in the source document, your chunk size is too small or your overlap is too low. Adjust and re-index the affected source.
- Source weighting. If the model keeps citing an outdated procedure document over a current one, the older file probably has more keyword density. Either archive the old file out of the index or set a recency weight.
- Refusal patterns. If the model is refusing on questions it should be able to answer, check whether the relevant folder was actually indexed and whether the user's permissions allow that source.
You also define, this week, the two or three metrics you'll review monthly. For most SMEs these are: number of queries per week, percentage of queries that returned a cited answer versus a refusal, and a small sample of answers reviewed by the data owner for accuracy. You don't need a dashboard. A spreadsheet is fine.
Week 6 onwards: Wider rollout and the things that will go wrong
From week 6 you open it up to the rest of the staff in groups — typically by team, not all at once. The most common problems in this phase are predictable. Someone will ask why it can't see their personal OneDrive (because you scoped it out, correctly). Someone will paste in confidential client information and ask a question, and you'll need a short policy on what's appropriate to put in the prompt. Someone will discover that two procedure documents contradict each other, and the Brain will surface the contradiction, and you'll have to decide which one is right. That last one is a feature, not a bug.
The longer-term work is keeping the index honest. Documents move, folders get reorganised, people leave and their mailboxes get archived. The Brain re-indexes on a schedule, but someone needs to own the question of whether the scope document from week 1 still reflects reality. I review it quarterly with the data owner. If you skip that review, within a year the Brain is answering from a corpus that no longer matches the business, and trust degrades quickly.
If you want the broader picture of how this fits together across different kinds of organisation, the Intelligence Brain overview covers the architecture and the verticals it's been built for.
Where to start this week
If you're an Irish SME owner reading this and wondering whether to begin: don't order hardware yet. This week, write the one-page data scope document from week 1. Walk the building, list every system, decide what's in and out. If you can't get that document agreed in a fortnight, the rest of the rollout will stall regardless of how good the technology is. If you can, you're already ahead of most of the firms I see, and the next six weeks become a sequence of straightforward engineering decisions rather than a project that quietly drifts.