Most small businesses evaluating AI start with the wrong question. They ask which model, which vendor, which tool. The question that actually determines whether an AI project works is much less exciting: is your data in a shape anything could use?
Almost every stalled AI project we have looked at stalled here. Not on the model. On the fact that the information the model needed was locked in PDFs, scattered across four systems, or existed only in one person’s email.
Why your data is the part that can’t be bought
Everyone has access to the same models. Your competitor can subscribe to the same API you can, on the same terms, this afternoon. The model is a commodity and it is getting cheaper every year.
What your competitor cannot buy is eleven years of your service records, your pricing history, the notes your team wrote about why a particular customer churned, or the specific way your business handles the edge cases in your industry. That is the asset. AI is the thing that finally makes it usable at scale, which is a real change from the previous decade, when that data mostly just sat there.
The four questions that determine readiness
Is it accessible? Can a system read it without a human retyping it? Data trapped in scanned PDFs, in screenshots, in a desktop application with no export, or in email threads is technically in your possession and practically unavailable. This is the most common blocker and the least discussed.
Is it structured enough? Not perfectly structured. Enough. A spreadsheet with consistent columns is fine. Free-text notes are fine, and modern models handle them well. What breaks things is inconsistency: three formats for the same date, four spellings of the same company name, a status column where half the values are typos.
Is it current? A system trained or grounded on last year’s pricing confidently gives out last year’s prices. The failure mode of stale data with AI is worse than with a static document, because the answer sounds authoritative and nobody thinks to check the date.
Is it in one place, or at least reachable from one place? Data in five systems is not necessarily a problem if those systems have APIs. It is a serious problem if two of them do not, because then a person becomes the integration layer, which is exactly the manual work you were trying to eliminate.
The cleanup that is actually worth doing
Do not try to fix everything. Almost every business has years of accumulated mess, and attempting a comprehensive cleanup before starting is how AI projects die in the planning phase.
Instead, pick the one use case you care about most and clean only the data that use case touches. If you want an assistant that answers customer questions from your documentation, the job is getting your documentation into readable text and removing the outdated pages. That is it. Your accounting history is irrelevant to this and can stay exactly as messy as it currently is.
This is the same principle as replacing one painful spreadsheet rather than all of them at once, which we covered in going from spreadsheet to software. Narrow scope produces finished projects.
What good enough looks like in practice
For most small business use cases, the bar is lower than people expect. Text that a machine can read. Consistent identifiers so records can be matched across systems. A known update path so the data does not silently go stale. Some indication of when each piece was last correct.
That is genuinely most of it. You do not need a data warehouse, a data engineer, or a governance framework to answer customer questions from your own documentation or to summarize a decade of service history. Those things matter at a scale most small businesses will not reach for years.
The two mistakes worth avoiding
Handing a model everything and hoping. Dumping an entire shared drive into a system produces an assistant that confidently cites a superseded 2019 policy document. Retrieval quality is mostly a function of what you let it retrieve from, and pruning is more effective than tuning. We went through the tradeoffs of different approaches in RAG versus fine-tuning.
Ignoring what is in there. Your data may contain customer personal information, employee records, or contract terms you are not free to expose. Before anything is connected to an AI system, someone should know what categories of information are in scope and what the rules are for each. This is not a compliance formality; it is the difference between a useful tool and a disclosure incident.
Where to start this month
Pick one recurring question your team answers over and over from memory or by digging. Find where the information behind it actually lives. Get that information into readable, current, consistent form. Then, and only then, put a model in front of it.
That sequence is boring and it works. The reverse sequence, buying the tool first and discovering the data problem afterward, is how most of the disappointing AI projects of the last two years happened. If you want help figuring out which question is worth starting with, our AI starting points guide is a reasonable next read, or you can just ask us directly.
Not sure if your data is ready?
Tell us the one question your team answers over and over. We’ll tell you where the data behind it lives and what it would take to put AI in front of it.
Get a Free Consultation