If you are asking how to bring AI into your business, there is a better first question: what will it be pulling from? Most AI tools that do anything commercially useful (summarising customers, flagging churn risk, drafting proposals from account history, forecasting revenue) work by reading your data. Your CRM, your invoices, your pipeline, your support tickets. If those are in good shape, the tools can do useful work from day one. If they are a mess, AI becomes a machine for producing confident nonsense at speed.
What happens when AI meets messy data?
Nothing dramatic, and that's the problem. AI tools don't stop and tell you your data is bad. They answer anyway, fluently.
Ask for your top twenty customers and the tool will rank them, even when one customer's revenue sits across three differently spelled account records and lands them at number thirty. Ask which accounts look like churn risks and it will summarise notes an account manager last updated fourteen months ago, presented as if they were current. Point it at your pipeline for a forecast and it inherits every stale deal your team never got round to closing out. Each answer arrives polished, plausible and specific. A broken spreadsheet at least looks broken. A broken AI answer looks like insight.
The commercial risk lands where the decisions do: which customers get priority, where you hire, what you promise the bank. Wrong inputs at that level cost real money, and the fluency of the output means nobody questions it until something visibly fails to add up.
Why growing businesses get caught
Because until now, people carried the context. In a business of ten, everyone knows that Smith Ltd and Smith Limited are the same customer, that the pipeline number is soft, and that "active customer" means whatever Sarah means by it. The data was messy then too; it just didn't matter as much, because the humans reading it filled the gaps.
AI has no context beyond what's written down. It can't know your conventions, your workarounds or the caveats everyone knows to apply. So the businesses most excited about AI (growing, stretched, no data team, CRM bought three years ago and half adopted) are often the ones where it has the least reliable material to work from.
What should you get right first?
Three things, in this order.
- One trustworthy record per customer. Duplicates merged, ownership assigned, key fields (status, segment, contact) current. Uncomfortable to start, quick to maintain once done.
- Definitions written down and agreed. What counts as an active customer, a qualified lead, churn, a closed deal. If two people give different answers, every report and every AI answer built on that data has already picked a side.
- A process that keeps both true. Who or what updates each field, and when. Increasingly this is a job for AI itself: an overnight sweep that merges duplicates, flags deals nobody has touched, and reconciles the CRM against what's actually happening in email and Slack. You still need an owner checking the sweep does its job, but the upkeep stops depending on anyone's discipline.
For most growing businesses this is weeks of focused work rather than months, and it pays for itself even if you never switch an AI tool on. Forecasts firm up, reports stop contradicting each other, and handoffs between sales and service stop dropping things.
Does that mean waiting to use AI at all?
No. It depends on what the tool reads.
Work that draws on the outside world rather than your systems is safe to start today: research, drafting, summarising meetings, first-pass marketing copy. Your data quality has no bearing on any of it, and the productivity gain is real.
Anything that answers questions about your customers, your revenue or your pipeline is a different category. Those tools are only ever as good as the systems underneath them, so run the check below before you connect one.
A ten-minute readiness self-check
- Pick ten customers at random. Does your CRM show one record for each, with a current owner, contact and status?
- Ask two people what an "active customer" is. Do you get the same answer?
- How many deals in your pipeline has nobody touched in ninety days?
- Do the sales numbers and the finance numbers agree without someone reconciling them in a spreadsheet first?
- If a regular customer halved their orders last month, would anything flag it?
- Can whoever looks after an account see what sales promised when they won it?
- Do prices and discounts live in a system, or in someone's inbox?
- If your best salesperson left tomorrow, how much of their pipeline could the team pick up from the CRM alone?
Comfortable answers to most of these mean you can point AI at your data with reasonable confidence. Three or more wobbles mean the tools will amplify the mess rather than cut through it (they always amplify; the only question is what).
If the self-check turned up more wobbles than you expected, that's normal, and it's fixable in weeks rather than months. Tidying these foundations is the work I do at The Camberwell, with or without the AI step after it. If you'd like a second pair of eyes on what you found, get in touch.