The Watson Weekly Weekend edition is sponsored by Avalara - the agentic AI platform automating global tax and compliance for leading eCommerce brands.

Lindsay Althouse works at a $4B HVAC distributor with 450 stores and 72 ERPs. She didn't open our conversation talking about cloud migration or AI strategy. She opened it talking about a guy on a roof in July. That reframe was the whole panel.

Last week I moderated a panel with Lindsay from Johnstone Supply, Bryan House from Elastic Path, and Lee Trotter from Data Realm. Three observations are still rattling around in my head.

Most B2B software gets designed for the buyer who signs the contract. Johnstone designed theirs for the contractor on a rooftop in July. That's a different product. It has to load fast, surface the right part on the first try, and respect that the user has dirty hands and 90 seconds before the next call.

Too many B2B teams build for procurement and then wonder why adoption stalls. Lindsay's team interviewed contractors directly. Then they ran an alpha through the summer with real users, and a beta with 100 customers from December to March. That is not how most legacy distributors work. It should be.

Johnstone runs 72 store groups, each on its own ERP. The boardroom instinct is to spend three years and tens of millions ripping and replacing. Bryan House made the case for the opposite move: put a commerce platform on top, abstract the legacy mess, and ship something customers can use this quarter.

This is the part I want every CIO to hear. Modernization doesn't mean replacement. It means giving customers a clean front door while the back of the house keeps doing what it already does. The $4B business doesn't stop while you re-platform.

Johnstone and her team put an AI assistant in the app. Then they did something most companies skip. They thought about the failure mode.

If a contractor asks which capacitor fits a specific furnace, and the model hallucinates, somebody installs the wrong part. That's not a returns issue. That's a safety issue.

Their answer: a private LLM, in a controlled environment, restricted to data already on their site. No open web. No consumer model. No improvising. The trade-off is fewer party tricks. The upside is the answers are right.

In B2C, a bad AI answer costs you a refund. In B2B, it can cost a building. More distributors need to internalize that gap before they ship.