Small-business AI adoption in the United States has moved fast: U.S. Chamber of Commerce surveys put usage at 23% in 2023 and 58% by 2025 (national figures). The pressure to “do something with AI” now reaches every owner, usually from several directions at once. The uncomfortable truth is that readiness has nothing to do with ambition and everything to do with order.
Before any AI conversation goes further, we run a readiness check with the business, and it comes down to three questions about data. None of them mention models, vendors or licences. When the answers are solid, the first project is usually obvious and small. When they aren't, the right answer is “not yet”, and this article is about why that's a result and not a failure.
Question one: where does your data live?
Try to list the places. The accounting system, the CRM if there is one, the spreadsheet the sales team keeps because the CRM is slow, the shared drive, the WhatsApp group where the warehouse confirms deliveries, and the inboxes of the three people who've been there longest. Most owners can name the systems. Fewer can say which one holds the version that gets used, and almost none can say how a number travels from one to the other.
This matters because any AI project starts by feeding something a body of data, and it can only be fed what can be found and exported. If a meaningful share of what the business knows lives in inboxes and heads, the model will learn from the part that was easy to reach and be confidently wrong about the rest.
Question two: who owns it?
Ownership has two meanings and both count. The first is a person. For each kind of record (the customer, the product, the price, the supplier), is there someone whose job it is to keep it right? Not the IT provider, and not “everyone”. A name.
The second is a system of record. When a customer's address differs between the invoicing system and the delivery app, which one wins? In many businesses the honest answer is “whichever one the person looked at”. Three systems that disagree amount to a standing argument, and an AI tool trained on all three will take a side at random.
Question three: can you trust it?
The test we use is deliberately blunt. Would you change your prices today based on what the system says about margin by product? Would you send a payment reminder to every customer the report flags as overdue, without someone checking the list first? If the answer is “let me look at it first”, you've just described the data quality problem in your own words.
Distrust is visible in meetings long before anyone runs a query. It sounds like two managers arriving with different numbers for the same month, or a report everyone knows to adjust by hand before reading. Those habits are the business compensating for data it can't rely on. Automation removes the person doing the compensating.
| Question | Ready | Not yet |
|---|---|---|
| Where does the data live? | You can name the systems and say which one holds the version that gets used. | A meaningful share lives in inboxes, spreadsheets and people's heads. |
| Who owns it? | Each kind of record has a named owner and a system of record that wins when two disagree. | “Everyone”, or whichever system the person happened to look at. |
| Can you trust it? | You would change prices or send reminders based on what the system says, without checking first. | “Let me look at it first”, and two managers arrive with different numbers for the same month. |
What “not yet” leads to
If any of the three answers is shaky, AI will automate the disorder, faster and at scale. That isn't a reason to do nothing. It's a reason to sequence, and the sequence is usually shorter than owners fear.
- Name the systems and decide which one is the record for each kind of data. This is a meeting and a page, not a project.
- Give each kind of data an owner and a simple rule for how it gets corrected.
- Close the gaps that matter, usually two or three: the spreadsheet that should be a field in the system, the export someone does by hand every Monday.
- Then pick the first AI project. Small, measurable and reversible. One process, one metric, one person responsible for reading the result, and a date by which the business can compare it with reality.
A good first project looks modest from the outside. Drafting the first reply to routine supplier queries, with a person sending it. Classifying incoming invoices to the right cost centre before someone approves them. Summarising the week's support tickets for the Monday meeting. Each of these can be switched off on a Friday with nothing lost, and each produces a number within weeks.
When the proposal has already arrived
Many owners come to us with a vendor's proposal already on the table. Three questions sort most of them. Which process, exactly, will this change, and who runs that process today? How will we know in eight weeks whether it worked, in a figure we already track? And what happens if we turn it off?
A proposal that can't answer the first is a product in search of a problem. One that can't answer the second is a pilot with no end. One that can't answer the third has quietly become infrastructure before it has proved anything. None of that is a reason to distrust the technology. It's a reason to insist on the order.
People stay responsible throughout. An AI tool doesn't decide for the business and doesn't remove the need to check, and the businesses that get value from it early are the ones where checking was already easy because the data was already in order. We use these tools every day in our own work, for research, code and documentation, so none of this is scepticism about the technology.
We tell businesses “not yet” often. It's cheaper than telling them “start over” later.
