What it is
Neverclosed is an always-on receptionist for small businesses. It lives on a company's website as a chat widget and on their phone line as a voice agent. It answers the questions customers actually ask — hours, services, pricing, "do you handle X?" — captures leads, and books appointments straight into the business's calendar. When a conversation needs a person, it hands off to one. It runs 24/7, because the customer with a burst pipe at 11 p.m. isn't going to wait until morning — they're going to call the next result.
Why we built it
Small businesses lose customers in the gap between someone reached out and someone answered. The owner is on a job, it's after hours, the front desk is slammed — and the message sits. Decades of lead-response research (the MIT/InsideSales work among others) point the same direction: the odds of ever connecting fall sharply within the first few minutes. Most people don't leave a voicemail and wait. They move on.
The usual fixes each have a catch. A human answering service is expensive and still misses the overflow. A generic chatbot answers instantly but will confidently invent a price or a policy it was never given — worse than silence, because now you've misinformed a customer and can't take it back. We wanted the instant-response half without the making-things-up half.
How it works (a little under the hood)
Three design decisions do most of the work.
1. It answers only from the business's own facts. The agent is grounded in one business's hours, services, prices, and policies, and it's built to say "let me take a message and have someone follow up" rather than invent an answer. On top of that we run a second verification pass whose entire job is to catch the first one making a claim the business's facts don't support. Anti-hallucination isn't a slogan for us — it's a second agent checking the first.
2. Actions are verified at the data layer, not in the transcript. The part we care about most as engineers. An agent saying "you're booked Tuesday at 3" is not the same thing as a booking existing. So we don't trust the words — a booking counts only when the record actually lands in the calendar and reads back correctly. We hold our own tests to the same bar: the grader reads the database, not the chat log, and a "success" the agent narrates but never performed is scored as a failure. (There's a whole harness behind that sentence — the subject of our next post.)
3. It's always on, and a human can step in instantly. The system runs on Cloudflare Workers and Durable Objects — fast, always-on, no server to babysit, each business isolated. When an owner opens their console mid-conversation, the AI pauses and the human takes over seamlessly; the visitor just sees a teammate join. Close the console and the AI resumes. Every lead is saved the instant it's captured (a lead you don't persist is a lead you lost), and the owner is alerted the moment someone asks for a callback.
What we refuse to do
- It always says it's AI. No pretending to be a person.
- Flat, transparent pricing — no per-minute meter. You know the bill before the month starts. Much of this category runs on overage surprises; we don't.
- Proof before promises. We'd rather show the agent handling a hard question than tell you it's good. We test ours hard enough to catch it failing — and we fix what we catch.
Who it's for, and where we are
Neverclosed is built for the businesses that feel every missed call personally — home services, clinics, shops, practices — the ones where a single after-hours lead can be the week. It's live, and we're opening it up now.
Open to collaboration — and building the distribution
We're early, and we'll say so plainly: the product is live, and the work in front of us right now is distribution — getting Neverclosed to the businesses that need it, done the right way. We're building that deliberately, from ambition rather than urgency.
We're also open to collaborators. If you're a developer who wants to build on or integrate with this, a partner who can help it reach the businesses it's for, or someone who believes an AI receptionist should be held to prove it, don't promise it and wants to build that with us — we'd like to talk. We're not looking to contract the work out; we're looking for people who want to build it alongside us.
We're building this in public, and the engineering runs deeper than one post can hold: the verification harness, the edge architecture, the honesty guardrails. This is the introduction. The rest is coming.
— The Neverclosed team · tryneverclosed.com
Top comments (0)