Stop Throwing Spreadsheets at AI.

Here's a sentence I hear at least once a week: "We just export everything and let AI analyze it."

I understand the appeal. AI is the most impressive technology of our lifetime, the export button is right there, and the demo is intoxicating — paste in a spreadsheet, ask a question in plain English, get a confident answer in seconds.

The problem is the word confident. Not the word answer.

The evidence is in, and it's brutal

MIT researchers looked at enterprise AI adoption in 2025 and found that about 95% of generative AI pilots produced no measurable P&L impact. Ninety-five percent. Companies with data teams, budgets, and consultants couldn't make "point AI at our data" pay off.

The failures weren't the model's IQ. The report's core finding was about context — the tools didn't know the business, didn't retain what mattered, and were fed inputs that didn't line up with reality on the ground.

If Fortune 500 companies fail at this with dedicated staff, what do we think happens when a five-location business pastes twelve exports into a chat window?

I'll tell you what happens, because I've watched it. The AI produces a beautiful analysis of the data it was given. The data it was given says one location has double the labor cost — because that location's payroll export includes a manager who covers three stores. The AI doesn't know that. It can't know that. It writes five convincing paragraphs about a staffing problem that doesn't exist, and it never once says "are you sure these numbers mean what you think they mean?"

A human analyst gives you a wrong answer nervously. AI gives you a wrong answer in perfect prose. The second one is more dangerous.

Confidently wrong is a legal category now

Air Canada learned this publicly. Its support chatbot invented a bereavement refund policy that didn't exist, a customer relied on it, and a tribunal ruled the airline liable for what its AI made up. The company's defense — essentially, "the chatbot is its own entity" — did not go well.

Now internalize that lesson for your own operations. When AI misreads your dirty exports and you cut staff, drop a product line, or sign a lease based on the analysis, there's no tribunal to appeal to. You're both the airline and the passenger.

AI isn't the problem. The pipeline is.

Let me be clear, because this is not an anti-AI essay: AI absolutely belongs in a modern small business. I use it daily. It should be drafting your job posts, summarizing your contracts, answering the customer questions you've answered a thousand times.

But there's a hierarchy that can't be skipped:

1. Clean, connected data — one canonical version of your locations, people, vendors, and accounts, with every system mapped to it.
2. Numbers with lineage — every figure traceable to a source record, so "why does this say $48K?" has an answer.
3. Then AI on top — asking questions of data that's already true.

Most businesses are trying to run step 3 without steps 1 and 2, and hoping the model sorts it out. It won't. A large language model is a reasoning engine, not a reconciliation engine. Feed it three systems that disagree about reality and it will not detect the disagreement — it will average it, smooth it, and narrate it.

Garbage in, garbage out is fifty years old. AI updated it: garbage in, gospel out.

The takeaway

The businesses that will actually win with AI over the next decade aren't the ones adopting it fastest. They're the ones whose data is clean enough that AI's answers can be trusted — and checked. That's the unglamorous work, and it's exactly what we built Sourcebook for: the clean, connected, traceable layer that makes everything you stack on top of it — AI included — actually worth listening to.

Do the boring part first. The magic part only works after.