The AI receptionist answering a 175-year-old newspaper

The Weekly Register-Call has been publishing in Colorado since 1851. Its phone still rings, and for a long stretch nobody was there to answer it. Two people cannot staff a switchboard and also put out a newspaper. So calls went to voicemail, and voicemail — as anyone who has run a small business knows — is where intent goes to die.

The brief was narrow: answer every call, understand what the caller wants, and either resolve it or hand it to a human who is actually available. No personality, no small talk, no pretending to be a person.

The shape of it

Twilio takes the inbound call and streams audio. Claude classifies intent against a short, fixed list — subscriptions, obituaries, legal notices, advertising, back issues, editorial, everything else. Once intent is clear, the call is either answered from a small knowledge base or bridged to the matching 3CX SIP extension. That is the entire architecture, and its simplicity is the reason it works.

• Twilio — inbound number, media streaming, call control
• Claude — intent classification and short factual answers only
• 3CX — SIP extensions for the desk, with real availability
• Supabase — call log, transcript, intent, outcome

The three rules that made it usable

One: never guess a price or a deadline. Legal notice rates and obituary deadlines are the two things callers ask about most and the two things that must not be wrong. Those answers come from a table, not the model. If the table has no entry, the call goes to a person.

Two: escalate on the second failure, not the fifth. Early versions kept trying to clarify. Callers hated it. Now, two consecutive low-confidence turns and it transfers with a summary attached.

Three: log everything and read it weekly. The transcripts are the product roadmap. Half the features in the current version came from reading what people actually called about.

The value was not the AI answering questions. It was that a 175-year-old paper stopped dropping calls, and I got a written record of what its readers want.

What surprised me

The biggest win was not resolution rate. It was the transcript archive. Within a month it was obvious that back-issue requests were a real, recurring, monetisable demand nobody had ever measured, because it had always vanished into an unlistened voicemail box. That is now a product.

The second surprise: callers do not mind talking to a machine if it is fast, honest about being a machine, and gets them to a human quickly when it should. What they mind is waiting.

What I would do differently

I would build the log and the intent table before touching the voice layer. The voice plumbing took two days. Getting the intent taxonomy right — and keeping the rate table authoritative — took two months, and that is where all the actual quality lives.