Why NappAI for voice agents

Most voice agent platforms on the market are just that: an agent that talks. To do anything real during the call (look up an order, register a customer, book an appointment) they have to reach out to external automation tools that you set up, maintain and pay for separately, and every trip adds waiting time to the conversation. On top of that, they charge a fee for every minute of call, in addition to what the models cost.

In NappAI the voice agent is part of a complete AI and automation platform. Anything a NappAI flow can do, the agent can do while it talks to the caller, without leaving the platform and without paying a middleman by the minute.

Savings on every minute of conversation

With NappAI the cost per minute of conversation is between 70 % and 88 % lower than on voice platforms that charge a per-minute fee. The savings come from several places at once:

No platform fee per minute. You pay the model providers you choose directly, with your own credentials, with no surcharge on each minute of call.
Local ears and voices. NappAI includes ears (STT) and voices (TTS) that run inside the platform, with no per-minute cost. With them you only pay for the model that thinks.
Automation is included. The agent’s tools are NappAI components. You don’t need another automation platform and you don’t pay for each operation the agent performs.
Less consumption on every turn. On every turn the model receives your instructions and your tools again. The agent uses the provider’s cache so that repeated part costs much less. Waiting phrases don’t pile up in the conversation the model rereads, and background voices are discarded before they are transcribed or answered.
It doesn’t say the same thing twice. If the agent already gave an answer before a tool finished, it doesn’t repeat it when the result arrives. You save that speech and that time.
Shorter calls. Tools run inside NappAI, with no round trips to external services. If a tool answers in under two seconds, the agent replies in one go, without a waiting phrase.
More calls at once on the same server. Each ongoing call needs very few resources, so a single server handles many simultaneous calls.

A complete platform, not just an agent that talks

Hundreds of ready-made tools. You can connect any NappAI component in tool mode to the Tools connector: CRM, ERP, databases, spreadsheets, email, calendar, search over your documents, your own code and much more. They run inside the platform during the call. You can also use MCP servers and webhooks.
Teams of agents. A tool can be a specialised agent, or an agent that in turn coordinates other agents with their own tools. While the caller keeps talking, that team solves complex tasks in the background and the result always reaches the conversation.
Other flows and other agents. The agent can run another of your flows as a tool or talk to agents on other platforms through A2A.
An assistant that builds the flows for you. The AI Assistant builds the flow from what you ask for in plain language: it picks the components, connects them, configures them, writes the agent’s instructions and checks that it works.

It only listens to the caller, wherever they are

With Ignore Other Voices turned on, the agent recognises the caller’s voice in the first seconds of the call and stops paying attention to the others:

Someone talking nearby or a TV in the background does not interrupt the agent and gets no answer. The decision is made before the agent stops talking, so there are no cuts.
It remembers several background voices so it can discard them sooner if they come back.
If the first voice it heard was not the caller’s (for example, a TV playing when the call was picked up), it corrects itself and follows the person who is actually calling.
No noise or echo in the browser. In calls from the browser (the widget on your website and the playground) background noise and echo are removed before the voice reaches the agent, and the microphone volume adjusts itself.
It answers on time even when they call from a bar. An AI model that runs inside NappAI tells the caller’s near voice from background chatter, so the agent knows when they have finished speaking and does not sit waiting for the room to go quiet. It adds no wait in quiet calls, has no per-minute cost and also runs on your own infrastructure.
Less noise by phone. Line noise is reduced before transcription when that improves what the agent understands, with no added wait for the answer.
Privacy by design. The voiceprint exists only while the call lasts and is erased when it ends. It is never stored, and what other people say is not transcribed.

A natural conversation

It knows when you have finished. It combines what you say with how you say it to decide whether you are done. It doesn’t cut you off when you pause while dictating a number, and it doesn’t keep you waiting once you have finished.
It gets ahead. It starts preparing the answer during your pause and starts speaking with the first part of the sentence, without waiting for the whole of it. Numbers, codes and spellings are read in one go, not chopped up.
Well-handled waits. It only says a waiting phrase if the tool is slow; those phrases don’t stay with the model as an example, so it doesn’t end up repeating filler.
It respects your turn. If the agent has just asked you a question, a tool result doesn’t talk over you: it waits for your answer.
No result gets lost. If you interrupt it right when it was about to give you a result, it comes back to it afterwards.
You can interrupt it. If you speak while it answers, it stops and listens.

Any model, including your own

Full freedom of provider. The model, the ear and the voice are chosen separately and from any provider in the catalogue, with your own credential.
Your own models. You can use models hosted on your own infrastructure, including models fine-tuned for your business, and the local ears and voices that come with NappAI.
Two kinds of agent. The Voice Agent, the cheapest and most adjustable, and the Voice Agent Realtime, with speech-to-speech models. A single call can move from one to the other without dropping. See Two kinds of agent.
Fallback models. If the main model, ear or voice fails during the call, the agent switches to a fallback without hanging up.

Transfers to people and to other agents

Direct transfer or transfer with a consultation first to a phone number or an extension of your PBX. In a consulted transfer, the agent first calls the person, briefs them on the case in private and only then hands over the call.
Handoff to another NappAI voice agent within the same call and with the whole conversation, also between a Voice Agent and a Voice Agent Realtime.

See Transfers and hanging up.

Control and governance

Guardrails. Your rules are checked before every action of the agent. If an action breaks them, it is not carried out.
Full history. Every call keeps its transcript, its recording and every tool that was used. See History, transcripts and recordings.
Protected secrets. Keys and passwords don’t show up in the execution log.
Teams, roles and audit. You decide who can view, edit or publish each agent. See Share your account.
Phone lines per account. The SIP lines an account registers can only be used by that account.
Your data stays with you. NappAI can be deployed on your own infrastructure, with the models, the recordings and the data inside it.

Where it makes the difference

Case	Why NappAI
High-volume contact centres	Much lower cost per minute, an agent that isn’t confused by background voices or noise and answers on time even when the customer calls from the street or a bar, and consulted transfer to a person.
Back-office, ERP and CRM operations	The agent works directly with your systems and with teams of agents, all within a single platform.
Banking, insurance, healthcare and public sector	Deployment on your infrastructure, your own models, Guardrails and protected secrets.
Teams that want to launch fast	The AI Assistant builds the flow and you can switch model, ear or voice at any time, without depending on one provider.