Summary
Why most AI pricing projects stall after the pilot: the real blocker isn’t data quality, it’s that your company’s pricing logic and decision rules were never documented. Learn how to check if this gap exists in your business.
Every leadership team I sit with has an AI data project running. Big budget, eighteen month roadmap, a platform that isn’t live yet.
Then a pilot works and it doesn’t scale. Nobody can say why. In the recent past I would have blamed the data. My guess now is the data was never the constraint.
We build pricing analytics for a living, so I went and counted our own. We’ve distilled how we actually do revenue management into a platform – market share, elasticity, rate-mix bridges, promo evaluation, price and promo scenarios. 18 analyses. Between them they call on 29 data fields, and only 10 are required to run anything at all: period, brand, category, banner, pack size, units, volume, value, price, and a promo flag (that can actually be derived).
Ten columns. I bet all ten can be found on your company’s invoices. Our team ran analyses on that 1,241 times in the last five weeks. Not one of those clients needed a data lake to get there.
The hard part was never the maths. It was writing down how we decide. When a coefficient is too weak to trust. Which price drop counts as a promo event. What we do when a client only has 40 weeks of history.
That took months – well, years, if you consider our experiences preceding the platform. And it was uncomfortable, because although we had hundreds of decks to pull this intelligence from, most of the validation, the correct codification, turned out to sit in three or four people’s heads. We’re a pricing firm and it still took us this long. Assume it takes you longer.
There’s a name for what comes out the other end: the ontology. The objects, the rules, and how they connect. Not the data and not the model. The layer that decides what any of it means.
Also, it doesn’t stop. In those same five weeks we edited that logic 309 times. Edge cases. Missing fields, missing mappings. Granularity that doesn’t line up between two systems that are both right. None of it is glamorous and all of it has to be decided by someone. At least initially, before any “Agent” is born.
This is the part AI doesn’t do for you. It builds fast. What it won’t do is tell you what’s worth building, and it will walk you somewhere wrong with complete confidence if you let it. You can generate 50,000 lines of code in a week now. The code isn’t the problem. The problem is that nobody wrote down what the right answer looks like. There’s nothing to check it against, and you’ll set prices off it anyway. Then it works in the pilot and falls over everywhere else. That’s the part nobody could explain. There’s no shortcut here, not for something as specific as the price of your own product.
Your data is probably fine. But ask five people in your company how a price gets set and you’ll get five answers. Not one of them is written anywhere a machine could read it. If your best pricing person resigned on Friday, what could a machine still do on Monday?
If you hold a portfolio, this is a valuation question now. Two companies, same margin, same systems. One has its commercial logic written down. In the other it sits with a CRO and two regional directors. Three years ago you’d have paid the same for both. The first can run pricing across every add-on in weeks. The second re-hires the judgement each time, or pays a consultant to. Nobody is diligencing for this yet.
Something you can do this week without us
Ask your commercial team for three rules they would say out loud. Our biggest customers get our best price. We don’t discount deeper than 25%. Big packs are cheaper per kilo.
Then have one analyst check each one against the last twelve months. Revenue per kilo by customer, not per unit, or pack mix will fool you. Discount depth by event. Price per kilo by pack size. That’s a day of work, and the ten columns are already in your ERP.
My bet is that at least one of the three is false and nobody knew. If that happens, forget whether your data is AI-ready. Your company can’t currently tell you which of its own rules it follows.
That’s the conversation I’d want to have.
