Phone and IVR agents where latency, telephony and accent handling are the real engineering problems. In production on a live telecom network.
Voice AI development is the building of agents that handle spoken conversations on a phone line: converting speech to text, understanding intent, retrieving or acting through your systems, and speaking a response back fast enough that the conversation still feels like a conversation. AMT builds voice agents integrated with enterprise telephony and IVR, in production today on a telecom network handling live customer transactions.
// the_decision_most_buyers_get_to_first
Start here, because for a meaningful share of companies the honest answer is buy, and we would rather say that on this page than three weeks into a paid engagement.
| A platform is right when | A build is right when | |
|---|---|---|
| Integration depth | The agent needs to look things up and book appointments through standard connectors. | The agent has to write to billing, provisioning or a core system with rollback and an audit trail. |
| Telephony | You are happy on the platform's numbers and routing. | You have existing IVR, contact centre infrastructure or carrier arrangements the agent has to live inside. |
| Compliance | General business use, standard terms. | Data residency, recording rules, sector regulation, or a security review that asks where every byte goes. |
| Volume economics | Low to moderate call volume. Per-minute pricing is fine. | High volume, where per-minute platform pricing stops making sense against a fixed engineering cost. |
| Control | You want it live in weeks and can work within the platform's limits. | You need to tune latency, handle your own accents and vocabulary, or the roadmap cannot depend on a vendor. |
If three or more of the left-hand column describe you, buy a platform. You will be live faster and cheaper, and we will tell you the same thing on a call.
Settle build versus buy properly
The comparison with the questions that decide it, plus what to ask a platform vendor before you sign. One page. It will save you a procurement cycle either way.
// the_engineering_problem
Text agents are judged on accuracy. Voice agents are judged on accuracy and on silence. A pause that would be unremarkable in chat feels broken on a phone call, and callers start talking over the agent, which breaks it further.
A single conversational turn is four stages, and every one of them spends time.
| Stage | What happens | Where the time goes |
|---|---|---|
| Speech to text | The caller's audio is transcribed, ideally while they are still speaking. | Streaming transcription costs far less perceived delay than waiting for the caller to finish. Whether the system streams is the single biggest design choice here. |
| Understanding and retrieval | Intent is resolved, context retrieved, systems called. | Usually the largest and most variable slice. A slow backend API is felt directly as silence on the line. |
| Text to speech | The response is synthesised. | Streaming the first syllables while the rest generates is what makes a system feel responsive rather than considered. |
| Telephony transport | Audio moves across the carrier network. | Small but not zero, and largely outside your control. It is the reason the budget has to be tight everywhere else. |
Designing to a latency budget across all four stages, rather than optimising one of them, is what separates a voice agent people finish a call with from one they talk over. Our telecom deployment averages under 3 seconds end to end while calling into live billing and provisioning systems.
// what_we_build
// proof
Deployed on a live telecom network as part of a single agent serving web, WhatsApp and IVR. 70% of customer queries resolved without a human agent, under 3 seconds average response, with bill payments, plan upgrades and SIM management completed on the call through live calls into billing, CRM and provisioning.
// cost
Voice AI costs have two components: implementation cost and ongoing usage cost. Implementation is primarily driven by telephony, integrations and conversation complexity, while running cost depends on call volume and the AI services used.
| Scope | Typical complexity / pricing | What drives the cost |
|---|---|---|
| Voice added to an existing AI agent | Medium | Telephony or voice-channel integration, STT/TTS, latency optimisation, interruption handling and speech quality. The underlying conversational logic already exists. |
| Standalone Voice AI agent | High | End-to-end conversation design, telephony integration and typically 2-3 enterprise system integrations, including authentication, business workflows and error handling. |
| Running cost | Per minute / usage-based | Telephony, speech-to-text, AI model usage and text-to-speech are typically usage-based. At higher call volumes, these costs should be modelled early to validate the business case. |
// the_honest_section
If a platform covers your case, use a platform. The comparison above is there so you can work that out without paying anyone to tell you.
If your call volume is low, the engineering will not pay back against per-minute platform pricing. Revisit when volume changes.
If the calls you want to automate are mostly complaints, voice automation will make them worse. Route those to a person faster instead.
// faqs
Building agents that handle spoken conversations on a phone line: transcribing the caller, resolving intent, retrieving or acting through your systems, and speaking a response back within a latency budget that keeps the conversation natural.
Buy when your integrations are standard, your volume is moderate and you are happy on the platform's telephony. Build when the agent has to write to core systems, live inside existing contact centre infrastructure, meet sector compliance rules, or run at a volume where per-minute pricing stops working. The table above walks through it.
Less than people assume. Callers interpret silence as a fault and start talking over the agent. The budget has to be designed across transcription, understanding, synthesis and transport together, because optimising one stage and ignoring the others changes nothing the caller can hear.
Yes. That is the normal enterprise case and it is the reason many companies cannot use a self-serve platform. Our telecom deployment runs inside live carrier infrastructure.
Technically yes, and our telecom agent completes bill payments in production. The controls matter more than the capability: scoped write access, idempotent operations, rollback, audit logging, and whatever your payment compliance regime requires.
By tuning to your actual caller base rather than accepting a generic model, and by designing confirmation behaviour so a misheard input is caught in conversation rather than executed. Neither is solved once. Both are tuned against real call recordings.
A build cost plus a per-minute running cost. At high call volume the running cost usually decides the business case, so it should be modelled before anything is built.
The channel and, because of it, the engineering. Voice adds transcription, synthesis, telephony and a hard latency budget on top of everything a text agent has to do. See conversational AI and chatbot development for text.
Thirty minutes with an engineer who has put a voice agent onto a live telecom network. We will tell you whether to build or buy a platform, and we will say buy when that is the answer.
Not ready to talk? Read how we deliver, or see what software development costs. No form.