The $40,000 chatbot nobody uses: an autopsy

A company came to us last year with a dead chatbot. They had paid an agency roughly $40,000 for it. In four months it had handled fourteen real conversations. Not fourteen thousand. Fourteen.

They wanted to know whether to fix it or kill it. The honest answer required an autopsy, and autopsies are useful, so here is this one, anonymized and shared with permission of the general pattern if not the patient.

Cause of death number one: it was trained on the website, and the website was wrong. Pricing had changed. Two services listed no longer existed. The bot faithfully repeated all of it, which meant the first thing staff learned was not to trust it. Trust, once lost inside a company, does not come back with a patch.

Cause of death number two: it had no job. Ask what the bot was for and you got a different answer per person. Leadership said lead capture. The office manager said reducing phone calls. The agency's proposal said, and I am quoting the vibe if not the words, "24/7 intelligent customer engagement." A system with three jobs has no job.

Cause of death number three: nothing connected to anything. When the bot did capture a lead, the lead landed in an email inbox, where it aged alongside the newsletters. No CRM entry, no alert, no follow-up. The bot was a dead end wearing a friendly face.

Cause of death number four, and the fatal one: no owner. Nobody's job included feeding it updated information, reviewing its conversations, or noticing it was dying. Software without an owner is abandonware from the day it ships. It just takes a few months for the body to cool.

Notice what is missing from this list: the AI. The model was fine. The language understanding was fine. Every wound was operational, which is the finding in almost every failed AI project we examine. The industry keeps debating which model is smartest while projects die of things a checklist would catch.

We rebuilt it, and the rebuild was mostly unglamorous. One job: qualify inquiries and book consultations, nothing else. A governed knowledge base as its only source of truth, reviewed on a schedule by a named owner. A hard escalation rule: any question outside scope goes to a human immediately, with context attached. Direct CRM integration with instant alerts. And a weekly twenty-minute review of its conversations, on an actual calendar, attached to an actual person.

Same underlying technology. Completely different outcome, because a chatbot is five percent chatbot and ninety-five percent the operation wrapped around it.

If yours is dying, check the ninety-five percent first.