Custom AI Agents
Every business has a sheet by the phone.
Your phone script, built into an agent that follows it
We sit down with the people who answer your phone and write down what they actually do: the questions they ask first, the rules they never break, the calls they pass to someone else. Then we set it up as a voice agent, connect it to what it needs to reach, and ring it until it holds — before a customer hears it.
+40 774 566 367 · No form: a build starts with a phone call.
Ring a sample caller through it
Traced on the sample sheet · not a live call
Pick any line to see what it became.
A fork in the diary: Saturday for plan customers, the first weekday slot for everyone else.
Every pin is a decision somebody at your desk already makes. We set them with you, one at a time.
Written with your team
The rules nobody wrote down, written until they hold
Your office knows plan customers get Saturdays and nobody else does. It’s never been written down, because it never had to be. An agent needs it written — and the first way a rule like that gets written often isn’t the way it holds. So we write it with you, ring it, and write it again.
The rule, in the brief
Only offer Saturday appointments to callers on a service plan. deleted: Ask whether they’re on one before offering times. added: Ask for the plan number on their renewal letter before offering times. With a number, offer Saturday and write the number down so the office can check it. Without one, offer the first weekday slot, or take a message for the office to confirm the plan.
Rewritten after draft 2’s test call.
A made-up rule for Quillmoor Heating, a made-up firm. This one took three drafts; plenty take one. What doesn’t change is that each draft is rung, not argued about.
Heard right
It listens for your names, because they’re on its list
Speech recognition is good at ordinary words and guesses at proper ones: your street names, your plan codes, your people. So a build includes a list of the words your callers say and a dictionary doesn’t. We put it together with your team — up to 100 terms — and every one gets said out loud in testing.
What the caller said
“Hi, is that Quillmoor? It’s the combi at 9 Wrenfield Close — plan QM-20417. Mira said Dan could come out.”
Written down without the list
Hi, is that quill more? It’s the calm bee at 9 Renfield clothes — plan cue em-20417. Myra said Dan could come out.
Written down with it
Hi, is that quill more? It’s the calm bee at 9 Renfield clothes — plan cue em-20417. Myra said Dan could come out.
Hi, is that Quillmoor? It’s the calm bee at 9 Renfield clothes — plan cue em-20417. Myra said Dan could come out.
Hi, is that Quillmoor? It’s the combi at 9 Renfield clothes — plan cue em-20417. Myra said Dan could come out.
Hi, is that Quillmoor? It’s the combi at 9 Wrenfield Close — plan cue em-20417. Myra said Dan could come out.
Hi, is that Quillmoor? It’s the combi at 9 Wrenfield Close — plan QM-20417. Myra said Dan could come out.
Hi, is that Quillmoor? It’s the combi at 9 Wrenfield Close — plan QM-20417. Mira said Dan could come out.
Hi, is that Quillmoor? It’s the combi at 9 Wrenfield Close — plan QM-20417. Mira said Dan could come out.
The list · sample
8 of 100 terms · 56 of 1,200 characters
Illustrative, read by an AI-generated voice: the kind of mishearing a list prevents, not a real caller’s recording. A term on the list is expected, not guaranteed. The list is used for agents speaking English, Spanish, French, Japanese and Hindi; in the other nine languages the recogniser doesn’t take one, so it reads names back and waits for a yes instead.
Wired to your systems
What it reaches during the call, what only after, and what we’d build for you
While the caller is on the line, the platform reaches your Google Calendar, texts the caller, warns your team and puts calls through. After the hang-up, a workflow takes it on: a Slack post, an email, a row in Google Sheets — or a signed delivery of the call to any other system. The end that receives that delivery is work we do on the build — a small receiver written for your system, or a Zapier, Make or n8n scenario set up for you.
Sample call · 2:14
Quillmoor Heating · a made-up firmSolid — on the platform todayDotted — built for you, as part of the build
- 0:02 · get_call_context · Thursday, 10:12, and the number calling
- 0:31 · save_lead_details · Plan number QM-20417, written down
- 0:44 · check_availability · Thursday 8:30 is free · Google Calendar
- 0:58 · book_appointment · Booked · Thu 8:30 · texted
- 1:06 · send_sms · What to have ready, texted to them
- 1:12 · notify_team · Mira texted: a plan to check before Thursday
- Plan check in your system · Built for you · scoped on the call · Not on the platform today. Code we’d write if your system has an API — we’ll tell you on the call whether it can be done, and it’s quoted on its own.
- Hung up · 2:14 · When the call ends
- send_webhook · Signed delivery · https only · tried up to 3 times
- crm.example.com · Your CRM · the receiving end · built for you, as part of the build
- helpdesk.example.com · Your helpdesk · becomes a ticket once the receiving end is built · built for you, as part of the build
- notify_slack · #office · Every call’s summary · on the platform today
After the call — or during it, if we build it
When the call ends, a signed delivery carries its details: the number, the summary, the outcome and the answers to your questions. It doesn’t carry the transcript. We build the end that receives it for your CRM — directly, or through Zapier, Make or n8n.
When should the plan be checked?
Dotted: a connection we’d build, if your system allows it. The caller hears the answer before they hang up.
Sample payload, trimmed · signed with X-NTV-Signature
- event
- workflow_triggered
- trigger
- call_ended
- call
- caller_number
- +15555550142
- duration_seconds
- 134
- outcome
- booked
- summary
- Plan customer, noisy boiler, booked Thursday 8:30.
- extracted
- plan_number
- QM-20417
Sent after the call by a workflow: signed, tried up to 3 times, https only. It carries the call’s details and summary, not the transcript.
A made-up call for Quillmoor Heating, a made-up firm.
Slack, Zapier, Make and n8n are trademarks of their respective owners. Neuro Tech Voice sends data to them through their standard webhook features and is not affiliated with or endorsed by them.
Gmail™, Google Sheets™, Google Calendar™ and Google Docs™ are trademarks of Google LLC. Neuro Tech Voice is not endorsed by Google.
Tested before launch
It takes real calls before it takes your customers’
We write a test sheet with your team: the regulars, the awkward ones, the one you’re dreading. The agent rings one of you from its own number, down the same route live calls take, and you play the caller. Every test call is transcribed. We read them with you, change what needs changing, and ring again.
On this sample sheet: six test calls, and two changes they caused — both rung again before launch.
Transcript excerpt · test call
Gives the plan number fast, on a bad line
First call
Caller
“Plan’s QM-20417.”
Agent
Thanks — QM-20471. And what’s the problem today?
What changed
added: Read the plan number back, a character at a time, before writing it down.
Rung again · 0:55
HeldAgent
That’s Q, M, 2, 0, 4, 1, 7 — have I got that right?
Caller
“Yes, that’s it.”
A sample test sheet from a made-up build. On a real one the scenarios come from your calls, and there are as many as they need. Test calls stop at three minutes, are marked as tests in your call log, and aren’t billed.
Before it starts
What we’ll ask, what you’ll get, and what we say up front
There’s no price or date on this page. Both depend on what your calls need, so both go in the quote.
What we’ll need from you
- A conversation with whoever answers your phone now: how the calls really go, and what they always say
- The sheet by your phone, or whatever you’d hand a new starter
- Three or four calls you remember — the awkward ones
- The systems a call should reach, and whether they have an API
- Your own test calls, and a yes before it answers a customer
What you get
- A brief written with your team, kept in the dashboard where it can be read and changed
- Your documents set up, so it answers from them
- A list of the names and codes it should expect
- Hours, booking rules, contacts and hand-offs
- Workflows to your systems, and the end that receives them
- The test sheet, with every test call’s transcript
Said up front
- It answers the phone. A chat version would be a separate build, scoped on its own.
- It gets its own local number, alongside your current line. Porting a number in isn’t supported yet.
- It books into Google Calendar. Another calendar is a question for the call.
- Documents are refreshed by hand. It doesn’t learn from calls by itself.
- It always says it’s an AI assistant, and that line can’t be removed.
- The build is quoted on its own. The plans still have no setup fee.
Who makes changes after launch — your team, us, or both — is agreed when the build is quoted.
What you just watched
Five claims. You operated the ones that are filled in.
- Your sheet became steps, rules and hand-offs you can point atTry it
- Rules are written with you, and rung until they holdTry it
- It’s given the names your callers say, and you can see whyTry it
- You can see what reaches your systems during the call, what only after, and what we’d buildTry it
- It takes real calls before it takes a customer’sTry it
The one we did not demonstrate
That it fits your business. Everything above was a heating firm we made up, with a sheet we wrote. The only way to see yours is to bring the one by your phone to a call.
Call us about a buildWhen you don’t need us
If your calls are bookings into Google Calendar, answers from your documents and messages for your team, the self-serve setup does all of it. The trial is free; bookings into Google Calendar start on the Pro plan.
Set it up yourself →Questions
About custom builds
Is this a different product from the one I can sign up for?
No. It’s the same agent on the same platform, set up by us instead of by you. What a build adds is the work: the writing, the setup, the testing, and — where your systems need it — the end that receives each call. Anything the finished agent does, the platform does.
Is it trained on our data?
No — no agent on our platform is. It’s briefed: written instructions, your own documents to answer from, and a list of words to expect. That’s why a change means editing a line, not retraining anything. It doesn’t learn from calls by itself.
Which CRMs and helpdesks does it work with?
None natively, and we’d rather say so. After every call, a workflow can send a signed delivery of the call’s details — the number, summary, outcome and the answers to your questions — and we build the end that receives it for your system, directly or through Zapier, Make or n8n. During the call it does what the platform does today: your Google Calendar, texts, messages, alerts and transfers. A lookup in your own system while the caller waits is separate engineering: if your system has an API, we’ll tell you on the call whether it can be done, and quote it on its own.
Can you build a chat agent too?
Not as part of this. The platform answers phone calls; there’s no chat widget in it. If you need chat, say so on the call and we’ll tell you honestly whether a separate build is worth it.
What does a build cost, and how long does it take?
It’s quoted after the call, because both depend on how much writing, testing and connection work your calls need. We won’t put a number here that we’d have to walk back. What’s quoted is our time: the plans still have no setup fee, and connecting the platform’s own integrations still costs nothing.
Is a build the same as the Enterprise plan?
No. The Enterprise plan is a subscription for high volumes, priced with you on a call: a written SLA, and a named contact. A build is the work of setting an agent up with you, and it’s quoted separately. It runs on whichever plan covers what your calls need — bookings into Google Calendar start on Pro.
Will it keep my number?
It gets its own local business number, bought from the dashboard and billed monthly on its own, alongside your current line. Porting an existing number in isn’t supported yet, and mobile and toll-free numbers aren’t offered.
Does it work with a calendar other than Google?
Not today. It books into Google Calendar during the call, with a fresh check before every booking. If your diary lives somewhere else, tell us on the call and we’ll say plainly whether a connection can be built for it.
What happens when it gets something wrong after launch?
Every call is transcribed and summarised, and a workflow can flag the ones worth reading: an unhappy caller, or a word you choose. The fix is a change to the brief or a document, then a test call to check it. Those are settings in the dashboard; who makes them after launch is agreed when the build is quoted.
Does it tell callers it’s an AI?
Always. Its greeting says so, and if a custom greeting leaves that out, the platform adds it back. A build can change every other line, but not that one.
Where is our data kept?
We’re an EU company, based in Romania. We keep transcripts and call summaries in eu-west-1 (Ireland). We don’t copy call recordings there: each stays with the provider that captured it, and ElevenLabs or Cartesia can keep their own copy of a call they handled. Twilio and ElevenLabs keep theirs in the United States; we haven’t confirmed Cartesia’s country. Some of our infrastructure and subprocessors are outside the European Economic Area, including in the United States; those transfers are covered by safeguards such as Standard Contractual Clauses, and the privacy policy says how. Calls are encrypted in transit, your records are kept apart from other customers’ with row-level security, card numbers are never stored, and a data processing agreement is available on request.
Start
Bring the sheet by your phone
A build starts with a phone call, not a form. Tell us how your phone is answered today and what it needs to reach. We’ll tell you what we’d build, what we wouldn’t, and what it would take.
+40 774 566 367 · The agent answers phone calls, in 14 languages
See everything the agent does →