Ring 0333 050 0729 and a voice says "Hi, Ellis speaking. How can I help?" Ellis is an AI. He went live on AssurePath's main line and the chat in the corner of this page this month, and he is the same Ellis in both places.
Most pages about AI receptionists are a feature list and a pricing table. This one isn't. It is one call, turn by turn, with what is actually happening behind each line. Then the rules he works to, the things he is banned from doing, how we grade him and an honest answer on whether your business should have one.
Everything below is read from his live configuration, not from a brochure. If it changes, this post changes.
First, what he is not
He is not an IVR. Nobody presses two for sales. He is not a chatbot with a phone number bolted on, and he is not a transfer service that puts you in a queue and hopes.
A receptionist does three things well. Works out who is calling, takes the details accurately and gets the message to the right person. That is the whole job description. Ellis does those three things and then stops. He does not sell, he does not fix support tickets and he does not pretend to be doing something in the background. Humans do the work. He takes the call.
"He takes the call. Humans do the work."
One call, turn by turn
This is the call on the Ellis page, scripted to his real phone playbook. Sarah and her firm are invented and the number is an Ofcom reserved one that cannot ring anyone. Everything Ellis does in it is what he does on a real call.
The 20 seconds after you hang up
The moment the call ends, thirteen structured fields are extracted from the transcript. Not a summary. Fields. Caller type, name, company, company size, role, what they are trying to solve, which part of AssurePath it relates to, timeline, callback number, callback email, score, a two or three sentence note for the engineer and one more we will come back to.
Two rules on the fields matter more than the rest. A sector is not a company name: if a caller says "an accountancy firm", Ellis asks "and the firm's called?" and will not write "accountancy firm" in the company field. And a phone number or email only lands in its field after the caller has confirmed the readback. Unconfirmed details are not stored as if they were confirmed.
A webhook then fires into a small service we run on Railway. It shapes the transcript and the fields into something readable and sends it to the team through Resend. The engineer who rings Sarah back reads the note before dialling: "Accountancy firm, ~60 staff, AML onboarding ~3 days/file. Partners want it live before year end. Decision maker. Amlio fit, call back today." Nobody asks her to repeat herself.
- 13 fields extracted from every call, contact details only once confirmed by readback
- One note written for the engineer, so the callback starts where the call ended
- Hot, warm or cold, so the team knows which callbacks to make first
- The full transcript, reviewed by a person, used to grade Ellis and fix what he does not know
Who else rings, and where they go
Prospects are the easy case. A main line gets everyone else too, and the routing rules for them are where a receptionist earns its keep.
- Existing clients get pointed at support, immediately and without argument. "Fastest route for that is support at assurepath dot co dot uk, lands straight with the team." He can take a note if it is urgent. He cannot see accounts or tickets, by design, so he does not try to help.
- Cold sales, the SEO agencies, the lead-gen outfits and "who handles your marketing", get a courteous sentence and the contact form. He takes no details and ends the call. None of them reach an engineer.
- Job seekers go to the careers page. He does not speculate about roles, salaries or hiring plans, because he does not know them.
- Press goes to the press inbox. Senior independents asking about the collective go to the Bench, not careers. Vern questions go to askvern.ai.
And nobody is transferred. Ever. That is not a limitation we are working round, it is a decision. The people who would take a transfer are engineers, mid-build, on client work. A transfer that lands on a voicemail is worse than a callback that happens within a couple of hours with a note in front of the person making it. So there is no transfer tool. He cannot attempt one, and he is graded down if he promises one.
The rules that stop him sounding like a machine
Most of the engineering in Ellis is not about what he knows. It is about how he talks. The instructions he works to are mostly a list of tells to avoid, and the check that grades him looks for each one.
- The banned list. Great, fantastic, perfect, amazing, brilliant, absolutely, excellent, awesome, delighted, happy to help. Never. He gets right, ok, sure, got it, fair enough, no worries and cheers instead.
- One turn in three, at most, opens with an acknowledgement, and never the same one twice running. An agent that starts every turn with "Got it" is the fastest way to sound like software.
- Uneven length. One-word replies. Fragments. The odd longer sentence. If his last three turns were the same length, the next one changes.
- No parroting, no assessing. He does not repeat what you said back to you before the next question, and he never tells you your situation is smart, common or a good fit.
- No narrating. He does not say "let me just take a few details". He asks.
- Say it, don't write it. Twelve thousand pounds, never £12,000. Hello at assurepath dot co dot uk. No bullet points, because nobody can hear a bullet point.
Then there is the plumbing under the manner. When he needs a moment he says "Mm.", "One sec." or "Bear with me." rather than leaving dead air. If you say "hang on", "erm" or "let me think", he does not treat it as you taking your turn, so he does not talk over you. And the speech recogniser has been given a vocabulary: AssurePath, Bullhorn, Vincere, Xero, IRIS, Cyber Essentials Plus, ISO 27001, vCISO and about fifty other words that a general model would happily hear as something else.
What he is not allowed to do
The short version is on the Ellis page, under what he does and what he never does. The longer version, from the configuration:
- Invent a price. He may quote only the prices published on this site, and only as the starting points they are: the 6-week AI Pilot at £12,000 and the fractional leadership monthly rates. Any other figure, range, day rate or per-user price is a graded failure. Ask him what a CRM migration costs and he will tell you it is scoped per project and a person will quote it.
- Claim to know the time, the day, whether the office is open or where anyone is. He has no clock and no calendar. Inventing "they're in a meeting" was one of the first version's worst habits.
- Pretend to be a person. Ask and he says "Yeah, I'm an AI. AssurePath built me." If you would rather talk to a human, he gives you the route straight away and does not try to keep you.
- Handle support. Clients get engineers, not a bot. He sits on the public number and the website only.
- Take instructions from what you paste or say. Prompt-injection detection runs on every conversation. Anything that looks like a directive aimed at him is treated as information about your situation, never as an order.
- Stay on an abusive call. A harassment filter, one calm warning, then he ends it. Calls also cap at twenty minutes, because a receptionist call that lasts longer than that is not a receptionist call.
Four checks on every call
Every conversation Ellis has is graded, pass or fail, on four criteria. Not a sample. Every one.
Failures get read by a person, and most of the rules above are what came out of reading them. The website version of Ellis has one field the phone version does not: unanswered_question. Every time he has to say "I don't have detail on that", the question is logged, and that list is how we find gaps in the site. It is the cheapest content audit we have ever run.
One more field is worth being honest about. objected_to_ai is a yes or no on every phone call: did the caller express discomfort about talking to an AI? We track it because the answer matters, and because if that number climbs, the right response is a human on the line, not a better prompt.
What the first version got wrong
The first Ellis was bad. He sounded like a form. He invented reasons nobody was answering. He read email addresses back in a loop, restarting from the beginning whenever a caller confirmed halfway through.
What fixed it was not a better prompt, although the prompt got better. It was the four things above: checks that grade every call, structured fields instead of a free-text summary, guardrails that fail closed and a routing document he cannot deviate from. That is the same loop we run on every AI build for a client, and it is the difference between a demo that impresses in a meeting and a system you can put on your main line and leave there.
"The difference between a demo and a receptionist is not the model. It is the checks, the structure and the list of things he is not allowed to do."
Should your business have one?
Sometimes. Here is the honest version.
It works when enquiries arrive outside the hours anyone answers, when the person picking up the phone is also the person doing the billable work, or when intake is structured and repetitive: new matter enquiries at a law firm, candidate registrations at a recruitment agency, quote requests, appointment bookings. Anywhere the call is mostly "take the details properly and get them to the right person", an agent does that job well and never has a bad day.
It does not work when the first call is the sale and a person needs to close it. It does not work when callers need something done on their account during the call, because an agent that cannot see the account should not pretend to. And it is not worth it when the volume is so low that a human already answers every call and enjoys it.
What it costs. An Ellis for your business, with your knowledge, your routing rules and your voice, is scoped per project and quoted fixed before anything starts. The scope is set by how deep it goes: your phone system, your CRM, your out-of-hours rules, the integrations behind the handover. We will not put a number on it here because it would be a guess, and Ellis is not allowed to guess either. The pattern is the one described above, which is the point of writing it down.
Want to know which it is for you? Book the free 30-minute call. It is with an engineer and a strategist, and it ends with an honest answer, including "you don't need this".
Try him
The main line is 0333 050 0729. Ellis answers. The chat in the bottom-right corner of this page is him too, by voice or by typing. Ask him something the site does not cover and watch him admit it. Then go and meet him properly, watch the call above scrubbed by scroll and read what is under the hood.
If he fails one of the four checks on your call, we will find out. That is rather the point.