by Serguey Shinder
Early in 2026 we put an assistant on our trade website to answer product questions, and we watched it for mistakes. Every week somebody read a sample of its answers against the product data, and the error rate was low enough that after two months we stopped worrying.
It was one of our sales representatives who noticed the problem, and it was not a mistake. A customer of hers, a builder, had asked the assistant how long a resin anchor took to cure at five degrees, because he was fixing a balustrade on a cold morning and needed to know when it could take weight. The assistant told him it could not advise on the use of hazardous chemical products and suggested he consult a qualified professional. He was the qualified professional. He rang the branch instead, and mentioned while he was on the phone that he had given up asking it anything with a warning label.
We pulled three months of conversations about our chemical products: adhesives, sealants, resins, solvents, treatments. About one in seven ended in a refusal like his. For nearly all of them the answer was printed on the safety data sheet or the label, documents we publish on the product page and are required to provide. Curing times, ventilation, protective equipment, how long before a treated surface could be painted.
None of this appeared anywhere we looked. Our review sample checked answers for accuracy, and a refusal contains nothing inaccurate. Our dashboard counted every conversation that ended without the customer asking for a person as handled, and a refusal usually ended exactly that way, because the customer left. From the inside, the assistant was careful and correct. From the outside, it declined to help with a large share of the questions our customers most needed answered, and the people it failed were the ones who knew what they were doing.
We changed what the assistant reads and what we count. For every product it now has the full text of the data sheet and the label, and on a safety question it quotes the relevant section and links to the document rather than paraphrasing. Refusals are recorded as their own outcome, separate from answers and separate from handovers, and fifty of them are read by a person every week. Refusals on chemical products fell to about one in forty. The ones left are mostly questions about mixing two products together, where the data sheet itself says not to, and saying so is the answer.
A wrong answer at least leaves a trace somebody can find. A refusal leaves nothing but a customer who has quietly decided the tool is not for people like him, and we had built our monitoring to find only the first kind of failure.
– Serguey Asael Shinder
Leave a Reply