What the Vendors Cannot Answer
A week of vendor pitches clarified the questions worth asking — and the ones that separate a working deployment from an expensive demonstration.
2 July 2026 — Most of this week's AI news concerned companies with hundreds of employees. For everyone else, the useful signal was in what vendors could not answer.
Three questions came up repeatedly in European SME procurement conversations this week, and the pattern in the responses is worth recording.
"Where does our data go?" The good vendors answer in a sentence and can show the subprocessor list. The difficult ones answer with a security page. If the answer requires a hyperlink, it is not an answer.
"What happens when it gets something wrong?" This one separates the field sharply. Vendors who have run deployments have a documented error path: what the system does when uncertain, who reviews the output, what the escalation is. Vendors who have run demos have a story about accuracy percentages, which is a different question entirely.
"Can we leave?" The best answer this week was from a mid-sized firm that keeps client data in formats the client owns, and can export everything in a day. That answer costs the vendor something, which is exactly why it is credible.
The regulatory context kept the same shape: high-risk obligations arriving in August, no SME exemption, and a classification question most businesses cannot answer without an inventory they have not built.
What this means if you run an SME. Ask those three questions before any AI purchase, and treat an inability to answer as the answer. A vendor who cannot describe their failure path has not thought about it, and you will be the one who discovers that in production.
Weekly review from the mid-week signal archive.