Your phone, answered.
Your inbox, answered.
Describe the assistant to the Claude or ChatGPT you already use. It answers on a real number you own — or a mailbox address — and it can only touch the tools you handed it.
This is the AI receptionist for small business — an AI answering service and AI voice agent in one: it answers the calls, texts and email you couldn't, in the words you gave it, with only the tools you handed it. It picks up on a number from Parsons Phone or a mailbox from Parsons Email. Every AI receptionist we priced runs $29 to $150 a month and answers before you can; this one is $15 a month for every Parsons app, and only speaks when you didn't.
It rang. You were up a ladder.
Nine calls came in yesterday. You took three. The other six met a voicemail greeting you recorded a long time ago and have not listened to since.
Two of those six were new work. You will never know which two, because people who need a job done today call the next name on the list.
The four emails are still sitting there. So are the two texts. None of it is hard — it is just that all of it arrives while your hands are full.
And the four emails. And the two texts. Every one of them is a person who was ready to buy something.
You looked into this once. You couldn't get a number.
Every published price in this category is either an addition problem or a sales call. Here are all three shapes, side by side, with ours in the same units.
The developer platforms
The answering services
This one
Retell's published list prices, checked August 2026. Their standard-tier model line runs from $0.003 to $0.16 a minute depending which one you pick; we have taken a mid one — $0.080, the GPT-5.4 and Claude-Sonnet rate. A faster tier roughly doubles it. Vapi's published list, same date: $0.05/min orchestration only, 10 concurrent calls included then $10/line/mo, HIPAA $2,000/mo, zero data retention $1,000/mo. Bland publishes $0.14/min with no platform fee, $0.12/min on a $299/mo plan and $0.11/min on a $499/mo plan; enterprise is a quote.
Cheapest paid plan on each vendor's own pricing page, published list price, checked August 2026. Rosie $49 is Professional, 250 minutes. Goodcall $79 is Starter and is per agent — $66 a seat if you pay annually. Smith.ai $150 is Pro on the AI receptionist — 75 calls at $2.00 a call, with a free 25-call tier below it and a per-call charge above the plan.
Everyone in this category sells an AI that answers first, because they are paid by the minute. Yours answers last — after your hours, your regulars, your cell phone and your team have all had their turn.
Configured nothing? Then it behaves exactly like theirs — the assistant answers every call. Zero rules is the default, not the ceiling. Everything below is what you get to add on top.
A real phone system. The AI is one step in it.
Rules match on your business hours, on specific callers, and on an expiry you can set for a single day. Each rule runs a chain of up to six steps. Change the clock and the caller below — you are running the same evaluation the line runs.
then forward → +1 415 ••• 7781 · whisper "press 1 to accept" · ring 20s
then forward → ring_group → menu → assistant → voicemail → reject
then voicemail · greeting · ≤600s
One honest limit: the assistant cannot transfer a live call. Call transfer is not offered in v1. The forward and ring_group steps run before the AI picks up, not after.
The end of the chain is not a dead end.
Your cell is ringing with a whisper prompt on it. You are holding a wrench. Press one of these.
That is the honest list. Live calls are not recorded and not transcribed — there is no transcript of this conversation anywhere, and we are not going to pretend otherwise.
The chain resolved before the assistant step, so no model was invoked. The talking is still metered, though: a forwarded call costs us a carrier leg at each end, so its minutes are counted exactly like assistant minutes — same $0.03, same monthly ceiling, counted once. One metric, one price, no asterisk.
The missed-call text only fires when the whole chain goes unanswered. You picked up, so the caller gets you instead.
"Hi, calling about the water heater — it's leaking at the base. My number is…"
Recorded, transcribed by a one-shot pass, stored on the call row and emailed to you with the duration.
"Sorry we missed you — reply here and we'll pick it up."
Sent to the caller when the whole chain goes unanswered. Requires an SMS-capable number with A2P registration approved.
Platform notifications to you go by email, not text. An SMS summary to your own phone is possible, but only if you grant the assistant the phone send_sms tool and ask it to.
One assistant, one channel. On purpose.
A voice that reads a URL aloud is a bad voice. An email that answers in two sentences is a bad email. So the channel is fixed when you create the assistant, and the writing rules arrive with it. Pick one to see what it does to the same sentence.
The channel picks the model. There is no model dropdown, and no model bill to assemble: on voice, the model is inside the $0.03 minute. On email and SMS the assistant's thinking is drawn from your prepaid wallet at provider cost × 1.5 — one line, one multiplier, printed in the receipt below. Want it on the phone and in your inbox? That is two assistants, and they both live on the same page.
It answers your inbox. In drafts, until you say otherwise.
Point a mailbox address at an email assistant and mail arriving there is answered by it. New bindings always start as drafts. Flip the switch and watch where the reply goes.
| address | assistant | mode | since |
|---|---|---|---|
| hello@example.com | Front Desk | send | Aug 14 |
| quotes@example.com | Quote Desk | draft | Aug 19 |
The one place that answers "what is running right now, and on whose behalf."
A stranger writes to it. That is the whole threat model.
An email assistant reads text written by people you have never met. Two mechanisms ship in the composed prompt itself — here is what the model is actually handed.
Every inbound message is wrapped this way before the model reads a word of it — and the marker id changes every time, so nothing inside can guess it and close the fence early.
{
"action": "none",
"reason": "asked for a price I don't have"
}
It is allowed to not reply. Most systems in this category structurally cannot choose silence.
It can do exactly what you handed it. Nothing adjacent.
An assistant points at a scoped tool profile in the MCP Hub. Deny-by-default by construction — the profile is the list, and the list is the whole reach. Toggle a tool and watch the hub tell you who else is affected.
This is PLAT-63 — the hub can be asked which assistants consume a profile, and which of them sit behind a line a stranger can dial, before you grant anything. It is not only advice: a grant onto a profile that answers a public number or address comes back confirmation_required, with a link, rather than going live.
The profile is re-resolved on every single tool call, never trusted from when the call started. Revoking lands mid-conversation. There is nothing to redeploy.
And the honest half: you can still point a public line at your all-tools admin profile if you insist. The confirmation is a link, not a refusal — an agent that simply cannot act teaches you to route around it. So we make you look at what is exposed, and then we let you decide.
Thirteen assistant tools. Twenty-three phone tools. No builder to learn.
There is no flow-chart canvas here, because the whole surface is MCP. You drive it from Claude, ChatGPT, or your own client — the same place you already work. The console exists so you can check.
You never have to open it. It exists so you can check. Every create, prompt edit, voice change, profile swap and archive also reads back from chat through recent_activity — no console required.
The paperwork nobody warns you about.
Before your assistant can text anyone in the US, a carrier consortium has to approve your business. It is tedious, it is slow, and it is not optional. You can do the whole thing by answering questions in chat — and here is what it costs, in advance.
Every project gets its own isolated telephony subaccount. Your numbers, your registration, your call logs — not pooled with anyone else's, and yours to release whenever you want.
Four things that stop this becoming a bill you didn't expect.
The page that headlines "never miss a call" and never mentions a ceiling is the page that produces an angry customer in month two. So: the line stops at 120 voice minutes a month by default. That is two hours of talking, and the ceiling is yours to adjust from the console. It is counted across every project on your account, and it is a stop, not an overage.
A per-call ceiling, on by default at nine minutes, settable anywhere from 30 seconds to 10 minutes. A stuck call cannot run all night.
Every minute is $0.03 from your prepaid wallet, with the model already inside that number. The ceiling is the hard stop on top of it — reached, the line stops answering until the calendar month rolls over.
Each number is a recurring committed charge — $3.99 local, $4.99 toll-free. You cannot accidentally accumulate fifty of them.
A per-project ceiling in dollars on the assistant's AI, with 0 as a hard lock. Voice has its own ceiling in the card above and outbound texts have theirs — 200 segments a day. Lowering a cap is one sentence to your AI; raising one is deliberately not, so nothing you delegate can widen its own limit.
Drawn from a prepaid wallet you top up. Voice $0.03/min, outbound SMS $0.02/segment, inbound SMS free. AI on the email and SMS channels bills at provider cost × 1.5.
These are not wallet items — they ride your subscription as recurring charges. Buying a number, releasing one, or filing a carrier registration all refuse the first time and quote you the charge: you call again with confirm before any money moves. If a number's rent goes unfunded for seven days it is released and the rental stops — you are emailed, and a released number usually cannot be recovered.
An archived assistant cannot strand a paid number. Archiving refuses while the assistant still answers a live channel — unbind first. It is a small thing that stops you paying $3.99 a month for a line nothing picks up.
One month, itemised.
A one-person business with one local number, an hour of calls, and forty missed-call text-backs. Every line below is a real rate from the price sheet — and the one line that is not here is named underneath, rather than left for you to find.
- Voice ceiling 120 minutes a month — adjustable in the console
- Up to 25 phone numbers, account-wide
- Every channel: phone, SMS, email
- Every other Parsons product
The trial covers the subscription, not purchases. A number or a carrier registration bought during the trial is charged to your card that day — the tool says so and asks you to confirm before it buys.
The $15 isn't the price of this product. It's the price of the account.
One subscription opens every Parsons product — you do not buy them one at a time. Most people arrive already paying more than $15 for a single one of them.
Unlimited projects. No extra cost.
Each project gets its own data, its own members, its own spend caps and its own telephony subaccount. Your business, your side business, and the one you help your brother with.
Metered, but not everything is.
Some things are measured and simply never billed. They are not a trial allowance — there is no cap row for them at all.
Price it against one thing at a time, and take the unflattering half first. The cheapest AI answering service in this category is $49 a month for 250 minutes — phone only, no inbox, no tools of its own, one business per account, and Goodcall charges its $79 per agent. At the receipt's volume we come in at little over half of that. At theirs the gap narrows: the $15 buys access, not talk time, so 250 minutes of calls lands near $44 with the number — and past roughly five hours of talk a month the per-minute line makes theirs the cheaper phone line. If one phone line at real volume is genuinely all you want, run that arithmetic; we would rather you read it here.
What the $15 buys instead is every other product on the list above, across unlimited projects, with no per-seat charge and no per-workspace fee: add your brother's shop and the weekend rentals and the subscription line does not move. And it is not a loss-leader with a metered bill hiding behind it — minutes, texts and AI are prepaid, itemised and capped, printed in the receipt above. The catch is not the price, it is the ceiling: the line stops at two hours of talking a month by default, and raising it is a console setting, not a plan change. That is the whole of it, and it is in the receipt.
Where we win, and where we lose.
A table that only flatters us would be worthless to you. The losses below are set in the same weight as the wins, because they are the same kind of fact.
| Parsons Assistants | Voice-AI dev platforms Vapi · Retell · Bland |
AI answering services Rosie · Goodcall · Smith.ai |
|
|---|---|---|---|
| Routing that runs before the AI | yesHours, VIP list, forward with whisper, ring group, menu, voicemail, one-day override | not the modelThe minute is the revenue; answering less is not their business | partialHours and forwarding, no tool chain |
| Email as an answered channel | yesAn assistant answers a mailbox address, draft-first | noVoice and chat widgets | noPhone only |
| Approval before anything is sent | yesNew email bindings always start as drafts | no | no |
| A revocable, deny-by-default tool profile | yesRe-resolved on every tool call; revoking lands mid-conversation | webhooksFunctions with no permission model | noFixed integrations |
| "Which public lines can reach this tool?" | yesThe hub answers before you grant it | no | no |
| Built from your own AI over MCP | yes13 assistant tools + 23 phone tools; the console is for checking | noTheir builder is the product | noTheir web form is the product |
| A price you can read without a sales call | yes$15 a month or $150 a year, plus a receipt | mixedPublished as an addition problem, or as a rate plus a monthly platform fee | yes$49–$150 a month, published |
| One subscription across every product | yesAnd unlimited projects at no extra cost | no | no |
| Where we lose — all of it real | |||
| Outbound calling campaigns | noInbound only. We do not dial out. | yesThis is their core business — buy theirs | no |
| Live-call transcripts and call analytics | noDuration, status and tool-call counts. Voicemail is transcribed; the conversation is not recorded. | yesFull transcripts and dashboards | yesUsually included |
| Mid-call transfer by the assistant | noNot in v1. Forwarding happens before the AI, not after. | yes | yesOften to a human |
| Porting your existing number | noNo port-in and no import. You buy the number in-app. | yes | yes |
| Choosing the model | noThe channel picks it. Deliberate — and still a no. | yesAnd you price it yourself | no |
| Healthcare and PHI | noNot eligible, no BAAs, and we will not pretend otherwise. If you handle PHI, buy from someone who sells into it. | someBland's own hero targets regulated industries; Vapi prices HIPAA as a $2,000/mo add-on | varies |
| Compliance certifications | noWe do not hold SOC 2 or ISO and will not imply one. | yesBland publishes SOC 2 Type II, a HIPAA BAA and PCI DSS; Vapi prices HIPAA as a $2,000/mo add-on | varies |
| Maturity | previewShipping, in preview, with a 14-day trial and a monthly or yearly plan you can cancel. | GA | GA |
Competitor rows describe their published positioning as of Aug 2026, not a benchmark we ran.
The questions people actually ask.
What happens when I use up my 120 minutes?
Can I keep my existing number?
search_available_numbers and buy_number, it lands in your project's own telephony subaccount, and you can release it whenever you like.Will it answer everything, even when I'm free?
Does it record my calls?
Can it take a card over the phone?
Will it invent a price to be helpful?
{"action":"none"} is a legal answer. It is allowed to say nothing rather than guess.Do I have to use the dashboard?
/app has six sections — Assistants, Channels, Numbers, Calls, Email and Settings — and it exists so you can verify what is running, not so you can operate it.Can my assistant text me a summary after every call?
send_sms tool, prompt it to text you, and have your A2P registration approved. We would rather tell you that now than after you have paid for a number.