I've now watched the same movie four times. A founder sits through a gorgeous demo: an AI tool that forecasts demand, triages support tickets, and probably files your taxes if you ask nicely. They sign. Six weeks later the pilot is dead, the vendor is muttering about "data hygiene," and the founder is out five figures plus a chunk of their faith in technology. An AI readiness assessment of their business data would have predicted the whole thing.

The model was never the problem. The model is almost never the problem. The demo ran on clean sample data, and the pilot ran on the company's actual data, which lives in eleven spreadsheets, two of them named FINAL_v7_reallyfinal.xlsx.

Before you buy AI anything, you need an honest answer to one question: is your data good enough? That's what an AI readiness assessment of your business data is for, and you can run a decent version of it yourself in an afternoon.

An AI readiness assessment for business data is a short audit that predicts whether your data can support the AI project you're about to fund. It comes down to five questions about truth, volume, history, freshness, and documentation. Answer them honestly and you'll know whether to green-light a prototype, fix your plumbing first, or politely decline the next demo.

Why AI Projects Die in the Data Layer

There's a dirty secret in the AI industry: the models are fine. GPT, Claude, Gemini, pick your favorite, they're all astonishingly capable at reading messy text, extracting structure, and drafting answers. What they can't do is invent data you don't have, or untangle data that contradicts itself.

Ask anyone who's shipped AI agents in production and you'll hear the same ratio: roughly 20% of the effort is model work and 80% is data plumbing, definitions, and edge cases. The demo hides this because the demo is allowed to cheat. Production isn't.

So the audit below isn't about your ambitions or your budget. It's about the raw material. Five questions, scored 0, 1, or 2 each. Ten points total. Brutal honesty required.

The Five-Question AI Readiness Assessment for Business Data

First up: can you find the truth? If a customer calls and asks what they owe you, either one screen answers it, or the answer depends on whether you consult the CRM, the invoicing tool, or Sandra's spreadsheet.

Score 2 if there's a clear system of record for the entities that matter (customers, orders, invoices). Score 1 if one exists but half the team works around it. Score 0 if "the truth" is a negotiated settlement between three exports. Most AI projects quietly require a 2 here. A model reading three contradictory sources doesn't average them; it picks one, confidently, at random.

If this question stings, that's normal. It's also fixable, and the fix usually starts with the boring work of consolidating one system of record rather than anything involving a neural network.

Questions Two Through Four: Volume, History, Freshness

Next: do you have enough rows? Say you want churn prediction. Lovely. Count your customers, then count how many have actually churned. If the answers are 340 and 11, you don't have a machine learning problem; you have a statistics homework problem with insufficient data. Score 2 if the thing you want to predict has happened hundreds of times in your records, 1 if dozens, 0 if it's happened twice and one of those was a data entry error.

Then there's history: how far back does it go? Seasonal patterns need at least two full cycles. If you want an AI to forecast holiday demand and your order system was migrated last March, the model will learn a year that never repeats. Score 2 for two-plus years of clean history, 1 for a year, 0 for "we switched systems and the old data is in a CSV nobody trusts."

Finally, freshness: does data arrive faster than it rots? An AI that answers questions about this morning's operations is useless if the underlying export runs every Friday. Score 2 if relevant data lands within an hour of the event, 1 if it's daily, 0 if someone manually assembles it before meetings. This is the question that separates a useful assistant from a very confident historian.

Question Five: Can a Human Explain the Fields?

Open your main system and look at the columns. If you see amt2, status_flag_old, or a code table where 7 means "pending" except in 2019 when it meant "archived," you have a documentation problem. I call it the Denise test: if understanding a field requires asking Denise, and Denise is on vacation, the field doesn't really exist.

Score 2 if a competent stranger could read your schema and guess 90% of the meanings correctly. Score 1 if there's a data dictionary that's merely two years out of date. Score 0 if the knowledge lives entirely in people's heads.

This matters more for AI than it ever did for reporting, because a language model will happily interpret amt2 as whatever seems plausible in context. Tools that enforce structured outputs from LLMs help, but they can't rescue a field whose meaning is folklore.

Scoring the Audit and What Each Tier Means

Add up your five scores. Now the uncomfortable part:

One warning: teams consistently overscore themselves by about two points. The fix is to make someone who didn't build the spreadsheets do the scoring. Preferably someone slightly grumpy.

The Two-Week Data Fix That Pays for Itself

If you land in the middle tier, resist the urge to launch a "data transformation initiative." Those have kickoff decks and a 40% survival rate. Instead, pick the single workflow you most want AI to touch and fix just its data, end to end.

A typical two-week cleanup: dedupe the customer list to one canonical ID, write one-paragraph definitions for the ten fields that matter, and automate the Friday export so it runs every hour. That's it. No new platform, no data lake, no committee.

The payoff shows up in weird places before the AI even launches. Reporting stops contradicting itself. Meetings get shorter, because the numbers on the dashboards executives actually open finally agree with each other. And then, when you do run the next demo, you can ask the only question that matters: great, now show me that working on my data.