Here’s something I’ve learned after three decades in IT: when an AI project disappoints, look at the data before you blame the AI. Data readiness for AI sounds like an enterprise concern, involving warehouses, pipelines and a team with “data” in their job titles. For a small business it’s much plainer than that, and much cheaper to fix.
This article explains what data readiness actually means for an SME, the three things you need in place before AI can do useful work with your information, and a 90-day roadmap to get there.
The data maturity gap in Irish SMEs
The gap between small and large businesses in Ireland is wide. The CSO’s 2025 enterprise survey found that 17.2% of small enterprises used AI, against 57.7% of large ones. The AI Economy Ireland 2026 report from Trinity College Dublin and Microsoft Ireland found that only 25% of SMEs save 2 or more hours per employee per week through AI, compared with 54% of large organisations.
Neither report blames data directly. That part is my view, from what I see in businesses of this size. Large organisations have spent 20 years putting their information into systems: document management, CRMs, ERPs, intranets. When they switch on an AI assistant, it has something to work with. Small businesses often have the same information spread across email attachments, desktop folders, a shared drive nobody tidies and three people’s heads.
Point an AI tool at that and you get what you’d get from a very bright new hire on day one with no access to the files. Articulate, and mostly guessing.
Gartner sees the same pattern at the large end. In a February 2025 press release it predicted that through 2026, organisations would abandon 60% of AI projects that aren’t supported by AI-ready data. If enterprises with data teams struggle with this, small businesses should take it seriously too.
What data readiness actually means
For a small business, data readiness comes down to four plain questions:
- Can it be found? Is the current version of a document in one known place, or are there five versions across email and desktops?
- Is access sensible? Can the right people see it, and are the wrong people kept out?
- Is it clean enough? Are client names spelled consistently, are templates up to date, is the price list current?
- Do you know what’s sensitive? Can you tell which files contain personal data or client confidential information?
A data lake can wait. Get your files in order and know what’s in them.
The three things you need before AI
1. One home for your documents
Pick one place where business documents live: SharePoint and OneDrive if you’re on Microsoft 365, or Google Drive if you’re on Workspace. Move working documents there. Archive the rest. Email attachments are not a filing system, however much it feels like one.
This matters because the AI assistants built into these platforms work with what’s stored in them. Microsoft 365 Copilot can only draw on files the user already has permission to access, and Gemini in Workspace works the same way with Drive. If the information isn’t in the system, the AI can’t use it.
2. Permissions that you’d be happy to see tested
Here’s the catch with point 1. If an AI assistant can find anything a user is allowed to see, it will find things that user was never meant to see but technically could. Those “Everyone” shares from 2019 suddenly matter. The salary spreadsheet in the wrong folder suddenly matters.
Before you turn on an AI assistant across your files, review who can access what. In a small business this is usually an afternoon’s work, and it’s worth doing whether or not you ever use AI.
3. A simple data classification
Agree four labels and teach everyone what goes where:
| Label | Examples | Can it go into AI? |
|---|---|---|
| Public | Website copy, brochures, published prices | Yes, any approved tool |
| Internal | Templates, internal procedures, meeting notes | Yes, business-grade approved tools only |
| Confidential | Client documents, contracts, financials | Only approved tools with business terms, and only where the client agreement allows it |
| Personal data | Anything identifying a person, especially health or financial details | Only with a clear lawful basis under GDPR, in approved tools, and ideally minimised first |
This table turns a vague worry (“is it OK to put this into ChatGPT?”) into a decision anyone on the team can make in seconds. The data layer of the Trust Stack goes into more depth on each.
A practical 90-day data roadmap
Days 1 to 30: find and tidy
- List where your business information actually lives today, including personal drives and email.
- Choose your one home for documents and agree a simple folder structure.
- Review sharing and permissions. Remove “anyone with the link” shares on anything sensitive.
- Agree your four data labels and write them on one page.
Days 31 to 60: connect and pilot
- Move active documents for one team or one process into the new structure.
- Turn on a business-grade AI assistant for that team, connected to those documents.
- Pick one task to pilot, such as answering questions about your own procedures or drafting proposals from past examples.
- Time the task before you start so you have a baseline.
Days 61 to 90: measure and extend
- Time the task again and compare.
- Write down what broke: missing documents, out-of-date templates, confusing folder names. Fix them.
- Decide whether the pilot earned its place, then move the next team or process across.
At the end of 90 days you’ll have one process working well, a document structure that supports the next one, and a measured result. That’s a better position than most businesses reach after a year of experimenting.
If you’re on Microsoft 365, our Power Up Microsoft 365 with Claude AI course covers connecting AI to the information you already hold. On Google, the Gemini and Google Workspace course does the same.
Common data mistakes that block AI adoption
Buying the AI first. Licences get bought, the assistant gets switched on, and it can’t find anything useful. People decide AI doesn’t work for them, and that impression is hard to undo.
Waiting for perfect data. The opposite mistake. Some businesses decide to tidy everything before trying anything, and two years later they’re still tidying. Get one area in order, use it, and learn from what breaks.
Ignoring permissions. An AI assistant is very good at finding things. Make sure what it finds is what it should.
Pasting personal data into free tools. The quickest way to turn a productivity win into a data protection problem. Your classification table and an approved tool fix most of it. Our shadow AI article covers how to find out where this is already happening.
Treating it as an IT project. The people who know where the information lives and what it means are the people doing the work. Involve them from day 1.
If you’d like help planning your first 90 days, our AI implementation guidance is built around exactly this kind of practical roadmap.