Your AI Tool Is Only as Good as Your Worst Spreadsheet

Your CRM says you have 847 active customers. Your accounting software invoices 612. Your sales spreadsheet, last updated six weeks ago by someone who no longer works there, shows 923. Which number do you feed the AI?

This is not a technology problem. The OECD documented it across countries in 2019: small businesses rarely lack data outright. The barrier is scattered systems, weak ownership, and no agreement on which copy of a record is the one that counts. The AI tool you're evaluating cannot resolve that conflict. It will pick a number and proceed, and you won't know which one it picked.

The tool choice is the wrong decision to make first

Founders spend hours comparing AI tools — pricing pages, feature matrices, demo calls. The research on AI adoption in SMEs points to a different failure mode entirely: projects stall because of fragmented infrastructure and missing data governance, not because the model underperformed. The model often works fine. The data it ran on was the problem.

Begg and Caira, studying actual small businesses against established governance frameworks, found the frameworks fail SMEs for a specific reason: they assume dedicated roles, formal ownership structures, and committees that small firms cannot staff. The gap isn't ambition. It's that the governance tools were designed for organizations with ten data analysts, not one founder who also handles sales calls on Thursdays.

Okoro's 2021 research on SME governance frameworks identified what the minimum actually looks like: named ownership and defined mastering rules. Not a data warehouse. Not a migration project. A decision about which record is authoritative and who keeps it current.

What 90 minutes actually produces

The audit has one output: a map. For each critical data type — customers, revenue, inventory, whatever your business runs on — you write down where the record lives, who last touched it, and when. Nothing more.

Start with customers. Open every tool your business uses and count how many places store a customer record. Your CRM, your accounting package, your email marketing tool, your invoicing system. Write them all down. Then ask: if these disagree, which one wins? If no one has ever answered that question out loud, you don't have a single source of truth. You have several competing ones, and your AI tool will inherit the conflict.

Do the same for revenue figures, product or service data, and any operational numbers you'd want an AI to reason about. The Hamburg CoreGovernance research built its SME governance model around exactly this exercise: defined responsibilities and secure management, not centralized infrastructure.

When the tool's tolerance for messy data becomes the founder's excuse not to fix it

A reasonable objection: modern AI tools are built for messy data. They deduplicate, they reconcile, they surface conflicts. Why spend ninety minutes auditing when the tool handles it?

The objection misidentifies the failure point. A tool tolerant of inconsistent formatting is not the same as a tool that knows which of your three customer lists is current. Those are different problems. The first is a technical capability. The second requires a decision only you can make — which record is authoritative — and no model resolves that for you. It guesses. The OECD's finding is precise here: the SME analytics barrier is trust in data quality, not tool capability. Skipping the audit doesn't address the trust problem. It defers it until the AI produces an output you can't verify.

Okoro and the Hamburg thesis both land on the same answer to this: ownership decisions made with minimal structure beat ownership decisions never made. The audit is the minimal structure.

What you're actually looking for

After ninety minutes, you should have a list with three columns: data type, authoritative source, owner. If any row has a blank in the owner column, that data type is ungoverned. Feed it to an AI and you're feeding it a record no one is responsible for keeping accurate.

The list doesn't fix the fragmentation. It tells you which fragmentation matters. A customer list owned by your operations manager, updated weekly in your CRM, is ready for AI use. A revenue figure that lives in three spreadsheets with no designated owner is not, regardless of which AI tool you're running.

That's the decision the audit forces: not which tool to buy, but which data you actually trust.