Audit, Blueprint, Build, Run. Scope and price are fixed up front, and everything sits in your name.
Start from the team that does the work, or the shape of your business.
Two-way syncs with reconciliation, retries and alerts, in accounts you own.
Systems in production, with their numbers and screens.
Customer stories
AI outbound sales platformA multi-tenant SaaS platform 
Property automations and agent fleetA 120-property rental portfolio 
Part finder for a cable manufacturer6,000+ article numbers decoded 
Production dashboard45% faster order processing 
Lead enrichment and sales automationFirst replies in under three minutes See the work
Portfolio Work we have shipped Client builds, prototypes and recorded walkthroughs, with real product screens, grouped by the tools they run on. Open the portfolio Field notes from building automation and AI systems, with sources.
AI voice agents
We build AI agents that take inbound calls, call new enquiries back and handle bookings, working inside your calendar, CRM and order system. Every call is written up, and anything the agent cannot handle goes to a person.
Hi, I have had toothache since last night. Can someone see me today?
Sorry to hear that. Dr Okafor has a 3:40 this afternoon. Shall I book it?
Yes please. Is it the usual price?
An emergency appointment is £65, and treatment is quoted on the day. You are booked for 3:40, and I will text you the details.
What the agent did
OutcomeBooked · 2 min 10 s · no hand-over needed
In short
An AI voice agent answers or makes calls in a natural voice and takes actions in your systems while it talks: booking appointments, qualifying leads, checking orders. SBD builds them on platforms such as Retell, ElevenLabs and LiveKit, connected to your calendar, CRM and helpdesk, with an AI disclosure, a hand-over to a person, and every call logged in accounts you own. Running costs are usually $0.06 to $0.20 a minute.
Where it usually starts
Use cases
Most clients start with one call type, usually the most frequent or the one that happens out of hours, and add more once the transcripts show it working.
Reception
Answer every inbound callBookings, opening hours, prices, directions and order status, answered from your own policies. Anything else goes to a person with a written summary.
Reception
After-hours and overflow coverTakes the calls your team cannot: evenings, weekends, and the moments every line is busy. Urgent cases are sent to the on-call person straight away.
Sales
Call new enquiries back within a minuteWhen a form comes in, the agent calls the person back, asks what your sales team would ask, and books a meeting with the right rep. Only for people who asked to be called.
Sales
Qualify inbound leadsA normal conversation that fills in your scorecard as it goes: volume, timing, budget, current supplier. Scores and notes land in the CRM.
Scheduling
Book, move and confirm appointmentsWorks against your real calendar or booking system, handles reschedules and cancellations, and sends confirmations by SMS or email.
Scheduling
Reminders and no-show follow-upCalls or messages the day before, confirms or rebooks, and offers a new slot to anyone who missed theirs.
Support
Order and delivery statusChecks the caller against the order record, reads out where the order is, and logs a ticket if something has gone wrong.
Support
Triage and hand overWorks out what the caller needs, collects the details, and transfers to the right person with the summary already on screen.
Operations
Supplier and driver check-insShort outbound calls to confirm a delivery window, a collection or a stock count, with the answer written back to your system.
Operations
Feedback calls after a visitA two-minute call after a job or appointment that asks three questions and flags poor scores to a manager the same day.
Website
A concierge on your siteLike Ava, bottom right of this page: answers questions from your own content, opens the right page and fills forms for the visitor to send.
Website
Callers in their own languageDetects the language and replies in it, so one agent covers customers your team cannot speak to.
How one call works
They repeat for each turn of the conversation. Each one is a part we choose, configure and test.
The call arrives on a number we set up, your existing line, a WhatsApp number or a button on your site.
What the caller says is transcribed as they speak. A turn detector listens for when they have actually finished, so the agent does not cut in.
A fast language model reads the transcript, your rules and your knowledge, and decides what to say or which tool to use.
Checks the calendar, looks up the order, writes the CRM record. The agent only says it is done once the system confirms it.
The answer is spoken back in the voice you chose, and the caller can interrupt at any point.
Transcript, recording where allowed, actions and outcome are saved and checked against your criteria.
Architecture
Cascaded: speech to text, model, voiceBest forMost business calls. It is our default.
Architecture
Speech to speech: one realtime modelBest forConversations where sounding natural matters more than control, or many languages on one line.
Platforms
Platform fees have fallen to 3 to 8 cents a minute, so the model and the voice now set most of the bill. The choice comes down to your phone system, languages, data location and volume. Published rates, checked 30 October 2026.
Not on the list: Synthflow now sells annual enterprise contracts only, which does not suit small builds, and Hume is closing its voice API on 13 November 2026. Enterprise contact-centre vendors such as Sierra, Decagon, Parloa and PolyAI sell to much larger companies than the teams we work with.
Models
New versions arrive every few months. Because each layer can be swapped on its own, a better or cheaper model is a configuration change and a test run, not a rebuild.
Speech to text
All stream for under a cent a minute. Deepgram Flux also decides when the caller has finished. Speechmatics runs on-premises.
Language models
Speed of the first word matters more than raw intelligence. A fast model runs the call; a cheaper one classifies; heavier reasoning sits behind tools.
Voices
Quality has converged. On a phone line the gap between premium and mid-priced voices mostly disappears, so premium voices earn their price on web calls.
Speech to speech
Gemini Live lists at roughly a tenth of OpenAI’s flagship realtime price. We keep a cascaded fallback in every build that uses one.
Turn detection
Models that listen to the audio, not just silence, so slow speakers are not cut off. Smart Turn is fully open source and runs on a CPU.
Phone systems and channels
Phone minutes are the cheapest part of the bill. Connecting to the phone system you already run is usually the hardest part of the build, so we price it as project work.
Twilio or Telnyx numbers in the countries you need, live in a day. Call costs are about 1 to 3 US cents a minute in the US and UK, more for mobiles abroad.
Forward calls, or connect over SIP. Microsoft Teams needs Direct Routing. 3CX cannot connect straight to most hosted platforms, so we bridge it through LiveKit’s SIP gateway.
Customers can call your WhatsApp Business number and reach the agent. Businesses cannot start calls from US or Canadian numbers, and the account needs a messaging limit of at least 2,000 recipients a day.
A call button in the browser, with no phone line at all. This is where premium voices sound best.
Running cost
All-in per-minute ranges from published prices: platform, transcription, model, voice and phone. They exclude our build and care plan. We confirm the figure from your own call data during the audit.
Latency
An independent test placed real phone calls to five hosted platforms and measured the gap from the recordings. None came in under a second at the median, and the platforms’ own dashboards reported about half a second less. We promise about a second, and measure it on your calls.
Testing and monitoring
Voice testing is now its own category of tools: simulated callers, regression suites and per-call scoring. We use the platforms’ built-in tests, open-source LangWatch Scenario, and dedicated tools such as Cekura or Hamming where call volume justifies them.
A suite of scripted calls covers every call type plus interruptions, background noise, accents, long numbers and attempts to push the agent off script. We use the platform’s own simulation tests or LangWatch Scenario in our pipeline.
A new prompt, model or voice reruns the suite before it reaches callers. Model providers ship new versions every few months, so this runs often.
After each change we read the calls where the agent was interrupted most. Average latency can look fine while callers are being cut off.
We measure the gap callers hear, from call audio. Dashboards report the platform’s side and miss the phone network.
Outcome, hand-overs, tool failures and complaints are scored per call, with an alert when a rate moves.
Per-call and daily spend limits and a cap on concurrent calls, so a loop or a flood of calls cannot run up a bill.
Where voice agents go wrong, and what we do about it
Rules by region
Say it is an AI in the first sentence, get consent before calling anyone, honour opt-outs however they are worded, and keep card numbers away from the agent. This is how we build; it is not legal advice, and some calls need a lawyer.
The rules every agent we build follows
The first sentence tells callers they are speaking to an AI assistant, and that they can ask for a person at any point.
Transfers go to a named person or team with a written summary, so the caller does not repeat themselves.
Any system that hears card numbers falls under PCI DSS. The agent sends a secure payment link instead.
We only build outbound calling to people who agreed to be contacted, with a consent record and an opt-out that works whatever words the caller uses.
No medical, legal or financial advice unless a qualified person has written the answer. Topics outside its job are handed over.
We use licensed stock voices by default, and only clone a real person’s voice with written consent covering scope, duration and what happens if they leave.
It can book a slot but not delete the calendar, read an order but not refund it. Every action is logged.
Transcripts and recordings are kept in systems you own, for as long as you decide, and deleted on request.
More in our AI policy and the agent risk check.
How we build it
The engineering is the shorter part. Most of the time goes into understanding your calls and testing the agent against them. It follows our usual Audit, Blueprint, Build and Run stages.
We listen to recorded calls or read call logs, group them by reason, and write down what good handling looks like for each, including when to hand over.
Platform, transcription, model, voice and phone route, picked for your languages, volume, phone system and where the data has to stay. You get the expected cost per minute in writing.
What the agent may say, what it must never do, which questions it asks and in what order, the opening disclosure, and the words it uses for your products.
Calendar, CRM, order system or helpdesk, each through a tool with the narrowest permission that does the job.
The regression suite runs until the agent passes, then becomes part of every later change.
Usually after-hours or overflow calls first. Your team reads the transcripts for the first fortnight before it takes more.
By industry
Each industry page shows the rest of the operations work we build alongside the agent.
Appointment booking, reminders, and prescription and results queries routed to the right team, with BAAs in place.
Viewing bookings, tenant repair calls logged with photos requested by text, landlord updates.
Reservations, changes and the questions guests ask before they arrive.
New-client intake, conflict questions collected, consultations booked with the right person.
Order status, delivery windows confirmed with drivers, reorders taken by phone.
Course and enrolment questions, open-day bookings, fee reminders.
Site visit bookings, supplier check-ins, client updates on progress.
Enquiry qualification and document reminders. Advice and payments stay with people.
Price
FAQ
Software that answers or makes phone and web calls in a natural voice. It transcribes what the caller says, uses a language model to decide what to do, takes actions in your systems such as booking an appointment, and replies in a generated voice, in real time.
Yes. Our agents say so in the first sentence. In the EU this has been a legal duty since 2 August 2026, and several US states require it too. Anyone who wants a person can ask for one.
Expect about a second between the caller finishing and the agent speaking. Independent tests of hosted platforms on real phone calls measured 1.3 to 1.7 seconds at the median, and platform dashboards understated that by about half a second. We measure it from your call recordings and tune it before launch.
It depends on the job. Retell suits most small-business phone lines, ElevenLabs is strongest for voice quality, languages and WhatsApp, Telnyx is the lowest cost for US lines, and LiveKit or Pipecat suit existing phone systems, data residency and high volume. Our live demos run on ElevenLabs. We choose after the audit and tell you why in writing.
Usually. We can forward calls from your number, connect your phone system over SIP, or set up a new number. Microsoft Teams needs Direct Routing, and 3CX needs a SIP bridge, which we build.
Running costs are per minute and usually come to $0.06 to $0.20 a minute all-in, depending on the platform, model and voice. Speech-to-speech on OpenAI’s flagship costs more, around $0.35 to $0.45. The estimator on this page gives a monthly figure, and we confirm it from your own call volumes before you commit.
It hands over. The call is transferred to a person or team, or a call back is booked, and a written summary goes with it. You decide the rules for when that happens.
Not by voice. Reading card numbers to an AI brings the whole call into PCI DSS scope, so the agent sends a secure payment link by text or email, or hands over to a masked keypad capture.
Only to people who have agreed to be called. In the US, AI voices fall under the TCPA’s robocall consent rules, and in the UK automated marketing calls need prior, specific consent. We build call-backs to enquiries, reminders and follow-ups, with a consent record and opt-out.
It happens: in the last year one voice provider closed after an acquisition and another is ending its voice API. We keep your prompts, knowledge, tests and integrations in your accounts and in a form that moves between platforms, so switching is a project of days, not a rebuild.
Our builds usually go live in 2 to 8 weeks. One call type with one calendar or CRM sits at the short end; several call types, languages and systems take longer. The audit gives a fixed date and price.
Yes. The live demos let you talk to a clinic receptionist and a lead qualifier in your browser, and Ava, our site concierge, is in the corner of this page.
Prices, versions and rules change often. Checked 30 October 2026:
Further reading: How to choose a voice agent platform · How much should an AI agent be allowed to do? · Putting an agent into production safely · What an AI agent costs
Keep exploring
Start here
On a discovery call we go through your call types and volumes, and tell you which one to automate first, which stack fits and what it would cost to run. The call is free.
Book a discovery call
Ava AI concierge
Ready
AI voice, powered by ElevenLabs. Ava can open pages and fill the contact form for you; you always press send. How we handle data
