November 10, 2026 · Operations
Thirty days of letting AI answer first-contact emails
312 inbound emails a month, two people, one product line. We let a model draft the first reply and kept a human on the send button. The reply time fell by 90%. So did our margin for error.
We get 312 first-contact emails a month. Valves, fittings, the unglamorous end of industrial supply. Two of us handle them, and for years the median reply time sat somewhere around nine hours — which in practice means the guy who answered at seven in the morning won the order.
In October I stopped writing first replies myself. Not all of them. The ones that arrive overnight, or while I am on a call, or the ones that are obviously a copy-paste blast to twelve suppliers at once.
The workflow is boring on purpose. Classify the email. Draft a reply with a fixed sheet of facts — what we make, what we do not, MOQ, the incoterms we will quote, the three certifications we actually hold. Then I edit. Every time, no exceptions. Then it goes out from my own address with my own name on it.
What thirty days actually produced
Median first response: 47 minutes. Thirty-eight percent of drafts went out with fewer than ten words changed. Twelve percent were unusable and I wrote them from scratch. Response rate to our replies moved from 19% to 24%, though I would be careful about claiming that is the model — October is a busy month and the sample is small.
The number I care about is the fourth one: how many drafts contained something that was not true. One, in roughly four hundred. That is the whole argument, and it is also the reason the workflow has a human in it.
The failure is worth more than the averages
A Turkish buyer asked for a pressure rating in bar. The model converted from psi, got the arithmetic right, and then added "suitable for steam service" because that phrase appears in a great deal of valve copy. We do not sell that. I caught it on the edit pass because it looked too confident for a sentence that specific.
If I had been in a hurry, it would have gone out. He would have ordered against it. We would have found out at the inspection.
So the rules are now written on the wall above my desk. The model never states a price. It never states a lead time. It never states a certification, a standard or a test result. Those three things are the entire commercial content of an inquiry. Everything else — tone, structure, remembering to ask what the application is, not writing "Dear Sir/Madam" in 2026 — the model does better than I do at six in the morning.
Why the tools do not matter much
People ask which platform. It is a general chat model, a saved prompt sheet, and a translation pass for Spanish and Portuguese. No plugins, no CRM integration, no AI sales platform with a seat price.
The expensive part was writing down the facts so the model could not invent them. That document is four pages long and it took two afternoons. It is also the only asset here that compounds, because every new product we add makes the drafts better without changing the software.
Would I do it again? Yes, with the same rules. What I would not do is let it send unsupervised. The entire gain is about forty minutes a day, and it evaporates the first time you ship a wrong spec to a customer who then orders against it.
One note on how this site runs. There are no affiliate links in this post and no vendor paid for it. If it saved you an afternoon, the hosting bill is on this page — six dollars covers a day of the server, and the ledger is public.
If you are testing something similar, the tools we used are in the library under writing and customer care. Start with one lane, not the whole inbox.
We publish the pricing model, minimum commitment and export-readiness notes for every tool in the library, and we flag it when a vendor changes one without changing the headline price.