Your data doesn't need to be perfect - it needs to be tidy enough

If you've ever been told 'you need to sort your data out before you can do anything with AI', you've probably also been told it'll take a year. Or that you need a data warehouse. Or a data engineer. Or a 12-month transformation programme run by a Big Four firm with a name you can't quite remember.

It won't. And it doesn't need to be perfect.

The 'sort your data first' line is often well-meaning but rarely useful. It comes from a world where the data work is the project. For most SMBs, the data work is just the bit that lets the project happen at all - and the bar is much lower than people think.

What 'tidy enough' means

In our experience, you need four things to be true to start getting real value from AI. Notice what's not on this list: a unified data platform, a master data management strategy, a chief data officer, or a single source of truth for everything.

• Your top three data sources are identified. For most SMBs that's the CRM, the accounting system, and the support inbox. Sometimes the operations or scheduling system. Rarely more than four things matter.
• Each one has a clear owner. Not a committee. A name. The person who, when something looks wrong, you ask.
• Customer records aren't duplicated three times across systems. You don't need perfect deduplication - you need enough hygiene that 'how many active customers do we have?' has one answer, not three.
• The fields that matter (name, email, last contact, status) are populated 90%+ of the time. The fields that don't matter can stay as empty as they like.

That's it. That's the bar for most useful AI work in an SMB.

Why perfection is the enemy

Waiting for perfect data means waiting forever. Meanwhile, your competitors are getting useful answers from imperfect data and improving things as they go. AI is surprisingly tolerant of messy inputs - especially for tasks like summarisation, drafting, classification, and triage. It's far less tolerant of teams that never start.

There's also a feedback loop you miss when you wait. Real AI use surfaces the data problems that actually matter. You'll discover that nobody fills in the 'last contact' field, or that two systems disagree about company names, or that your tags are inconsistent. Those are the problems worth fixing - and you wouldn't have found them sitting in a planning meeting.

Where to spend your first fortnight

Two weeks of focused, unglamorous work usually does it for an SMB. Roughly:

• Day 1-2: list every system you actually rely on. Not what you bought - what you use. Pick the top three.
• Day 3-5: walk through each one with its owner. Look at the fields that matter. Note the obvious gaps and duplicates.
• Day 6-8: clean the worst of it. Merge the obvious duplicate customer records. Fill in the missing emails. Standardise the status values.
• Day 9-10: write a one-page note on what's tidy, what isn't, and what the AI pilot can and can't safely use.

Two people, two weeks, no consultants. That's usually enough to clear the runway for a first AI pilot.

The exceptions

There are use cases where data quality really does matter from the start. Anything involving regulated decisions about individuals, anything customer-facing where wrong answers carry real cost, anything where the AI replaces (rather than supports) human judgement. For those, slow down and invest more.

But for the 80% of useful AI work in an SMB - drafting, summarising, classifying, triaging, answering routine questions - 'tidy enough' is genuinely enough. Get to it in a fortnight. Improve as you go. The perfect data strategy is the one your competitors will still be writing while you're already learning from real customers.