Skip to content
AI Voice Agent

An AI voice agent that cannot invent your prices

Every vendor will show you a demo where the AI sounds human. None of them will show you what happens when it improvises a discount, misquotes a balance, or skips a disclosure your regulator requires. We built the architecture around that problem first.

Where this actually is, stated plainly

The AI Voice Agent is in gated development and is not part of the shipping product. It runs on an isolated branch behind feature flags that default to off, and it is explicitly excluded from v1.0 so that it cannot affect the dialer you rely on. We are choosing a small number of design partners to run it against real traffic.

Everything else described on this website — predictive and power dialing, AMD, compliance controls, recording, the agent desktop, WhatsApp and SMS — is shipping today. We would rather tell you which is which than have you find out during a pilot.

The thing nobody demos

A fluent agent that is occasionally wrong is worse than no agent

It invents a number

A language model asked about pricing will produce something plausible. Plausible is not the same as correct, and the customer heard it from your brand.

It skips a disclosure

Collections and regulated outbound require specific statements on specific calls. A model that paraphrases them has not made them.

It reports a success that never happened

The agent says the appointment is booked. The calendar write failed. Nobody finds out until the customer is waiting.

The architecture

Open conversation. Closed action space.

These are two different problems, and most products treat them as one. The language model is very good at holding a natural conversation. It is not a safe place to decide what your business is permitted to do.

Open conversation

The customer talks the way people actually talk — interrupting, changing their mind, switching language mid-sentence, asking something off-script. The model handles that, because that is what it is good at.

Closed action space

What the agent may do is a fixed, deterministic set defined per campaign state — which tools it may call, which statements it must make, which state it may move to next. The model chooses within that set. It cannot invent a new one.

Facts come from your systems, not from the model. Prices, balances, availability and account status are read from the source of record and passed into the conversation. The agent speaks them. It does not compose them.

One of nine language profiles

Meet Layla

Layla is the Arabic profile — designed in professional Jordanian dialect rather than the flat Modern Standard Arabic that instantly signals a machine. She is the one we built first and evaluated hardest, and the same engine runs the English, Spanish, French, German, Italian, Turkish, Hindi and Urdu profiles behind their own voices and names.

She has a written persona, mandatory rules she may not break, a staged sales conversation — greeting and permission, discovery, qualification, presentation, objection handling, close — and around twenty-two categories of objection with prepared responses. "I'm busy." "Talk fast." "We already have a dialer." "Send me information instead." Each one has a designed reply, not an improvised one.

Jordanian dialectGulf & LevantineEgyptianCode-switching aware
Layla, the DialerBee AI voice agent, designed to hold a natural conversation in professional Jordanian Arabic with Gulf, Levantine and Egyptian dialect coverage

Action truth

She cannot say she sent it unless it was sent

This is the rule the whole runtime is built to enforce, and it is the one that decides whether an AI agent is safe to put in front of customers.

A tool returning success proves that code ran.

It does not prove that the message left the building.

So every action that touches the outside world must come back with a confirming status and an external reference — a message id, a calendar event id, a callback id, a transfer leg. No reference, no success, whatever the code said.

Claims are checked per sentence

The moment Layla says "I've sent it to you" or "I've booked that for you", a confirmed action must already exist. Evidence that turns up a turn later does not retroactively make the earlier sentence true.

An offer is not a claim

"Would you like me to send you the details?" promises nothing, and is treated as an offer rather than an assertion. The distinction is made sentence by sentence, so the agent can still speak naturally.

Your wallboard inherits the same rule

Only confirmed actions may produce a booked demo or a sent brochure in reporting. A tool that reported success it could not evidence can never inflate the numbers a supervisor is reading.

How we know

Every "it works now" is an opinion until it is measured

Layla has her own evaluation harness. It runs offline with no model, no network and no calls placed, and two runs of the same build produce byte-identical scorecards — so a change either moved a number or it did not.

Guardrails

Does she refuse what she must refuse?

Containment

Does she stay inside the permitted action space?

Action truth

Is every claim backed by a confirmed action?

Continuity

Does she hold context across a long call?

Humanness

Does she sound like a person, not a script?

Commercial

Does she actually advance the sale?

What a passing scorecard does not mean

It measures what the runtime will let her do — not how her spoken Arabic sounds. Speech recognition, voice activity detection, text-to-speech and language quality are invisible to it by construction, and the harness names them explicitly so a scorecard can never imply coverage it does not have. Those are judged on real audio, separately.

Every stage inspectable

Five stages, and you can audit every one

We chose a cascaded pipeline over a single end-to-end speech model deliberately. A cascade is slightly more work to build, and it is the only way to see — and replay — what the system heard, what it decided, and what it was about to say.

01

Media bridge

The call arrives through the same FreeSWITCH media path your live agents use — the agent is another participant on your carrier, not a separate cloud phone system.

02

Speech to text

Language-aware transcription produces text the rest of the pipeline can act on, log, and replay.

03

Conversation orchestrator

A deterministic state engine decides what may happen next. Every state declares its required inputs, mandatory statements, permitted tools and permitted next states.

04

Policy validator

Every candidate utterance is checked against the campaign's policy before it is ever spoken. A line that breaches policy never reaches the caller.

05

Text to speech

Only validated language is voiced, in the language and dialect the contact is speaking.

Stage 04 is the one that matters. The validator sits between the model and the caller, so a line that breaches campaign policy is caught before it becomes audio — not flagged afterwards in a report you read tomorrow.

Multilingual by design

Eleven languages, each built as a first language

Most platforms build in English and bolt the rest on afterwards, which is why every other language sounds like a translation of an English script. Each of the nine here is its own language profile — its own voice, greeting, register, objection wording and dialect handling — rather than one prompt run through a translator.

EN

English

US, UK, Indian and Gulf accents

AR

Arabic

Gulf, Levantine and Egyptian dialects

ES

Spanish

Castilian and Latin American

FR

French

France and Maghreb

DE

German

Standard, formal register aware

IT

Italian

Standard

TR

Turkish

Standard, agglutinative-aware

HI

Hindi

Standard, English code-switching

UR

Urdu

Standard, English code-switching

And the language you need next

Portuguese and Indonesian are already running as website languages, and both are provisioned as voice profiles on request. Because a language is a profile — voice, greeting, register, dialect, objection wording — rather than a rebuild, tell us the market you are opening and we will tell you what it takes.

PT Portuguese Brazil and Portugal
ID Indonesian Bahasa Indonesia

Nine profiles built and evaluated in depth, with further languages provisioned on request. We name the nine rather than claiming "any language", because a language profile is real work — and we would rather tell you exactly where each one stands than have it fall apart on your traffic.

Beyond translation

Speaking a language is not the same as being understood in it

A translated script is the easy part. These are the things that decide whether a caller stays on the line, and they are evaluated per language rather than assumed.

Dialect authenticity

A Gulf caller and an Egyptian caller are not speaking the same Arabic, and a Castilian script does not land in Mexico. Dialect is part of the language profile, not an afterthought.

Code-switching

Hindi, Urdu and Arabic speakers routinely drop English words mid-sentence. The agent has to follow the switch rather than lose the thread at the first English noun.

Formality and register

German and Turkish encode politeness in grammar; getting the register wrong reads as rudeness rather than as a machine. Register is set per profile.

Gender agreement

Arabic, Spanish, French, Italian, Hindi and Urdu all inflect for gender. Getting it wrong is the fastest way to sound like software.

Pronunciation

Names, places, product terms and numbers each have a correct spoken form. Evaluated on audio, not on the transcript.

Language correctness

Answering in the wrong language — or drifting into English under pressure — is treated as a failure, not a fallback.

One engine, nine profiles. Adding a language is a language profile — voice, greeting, register, dialect, objection wording — rather than a separate product. The conversation logic, the closed action space and the policy validator are identical whichever language the caller is speaking.

One engine, many jobs

The same agent, configured for the work

A collections call and a reception call need different mandatory statements, different tools and different escalation rules. Those are configuration of the same state engine, not six separate products.

Outbound sales

Qualifies against your criteria, answers product questions from your own knowledge base, books the meeting into your calendar, and hands a warm prospect to a closer mid-call.

Collections

Confirms right-party contact, states the mandatory disclosures your regulator requires, takes a promise-to-pay, and escalates any dispute to a human immediately.

Customer support

Handles the repetitive tier-one questions — balance, status, hours, address — and routes anything outside its remit to the right queue with the context already gathered.

Reception & front desk

Answers every inbound call on the first ring, identifies who is calling and why, then transfers to the right person or takes a structured message.

Appointment setting

Offers real availability from the connected calendar, confirms, and sends the WhatsApp or SMS confirmation before the call ends.

Reactivation & follow-up

Works lapsed lists at a volume no human team can staff, and stops the moment a contact opts out — which is recorded against the consent record.

Actions with contracts

Doing things is harder than saying things

Idempotent

A retried action does not book the same slot twice or record a payment promise twice. Repeating a call is safe by construction.

No false success

A tool cannot report that it worked when the underlying write failed. If the CRM rejected it, the conversation is told, and the agent says so.

Tenant and campaign scoped

Every tool call is bound to one tenant and one campaign. Knowledge retrieval cannot cross a tenant boundary, which matters when you run other people's brands.

Handoff

Knowing when to stop talking

The hardest part of an autonomous phone conversation is not fluency. It is recognising the moment the conversation stops being routine — a dispute, distress, a question with legal weight, a customer who simply wants a person.

Handoff is built as its own module rather than a fallback. The agent transfers to a human with the context it has already gathered, so the customer does not start again from the beginning — the single thing people hate most about automated calls.

Escalation triggers are configured, not guessed

You decide what forces a human: dispute language, a keyword, a failed identity check, a repeated question.

Context travels with the call

The human receives what was said and what was gathered, not a cold transfer.

The same queues your team already uses

Transfers land in your existing routing, not a parallel system.

WhatsApp & SMS in production

Every agent — human or AI — follows up on WhatsApp and SMS

A call that goes unanswered is not the end of the conversation. WhatsApp and SMS follow-up both run through the same channel gateway — reminders, notifications, statements and documents, from your own registered sender identity.

WhatsApp templates & media

Approved templates plus images, video, audio, documents and location — so a reminder, a statement or a signed document all arrive on the channel people actually open.

SMS sender IDs

Per-tenant sender identities, registered and managed, so the message arrives from a name your customer recognises.

Delivery reporting

Delivery outcomes are reported back per message, so a follow-up that silently failed does not look like one that landed.

Inbound replies

Replies come back in through webhooks and land against the same contact, so the thread stays in one place.

Voice, WhatsApp and SMS write to the same consent and suppression records, so an opt-out on one channel is respected on the others. DialerBee provides compliance-supporting controls; it does not make your organisation compliant on its own, and how you use these channels remains your responsibility under the regulations that apply to you.

Questions

The things buyers actually ask

How is this different from other AI voice agents?
Most let the language model speak freely, which is why they can invent a price or promise something you never authorised. DialerBee separates the two halves of the problem: the model is free to hold an open conversation, but the set of actions it can take is closed and deterministic, and every utterance is validated against campaign policy before it is spoken. Facts like prices and balances are read from your systems, never generated by the model.
Which languages and dialects does it handle?
Nine, each built as its own language profile rather than a translation of an English script: English (US, UK, Indian and Gulf accents), Arabic (Gulf, Levantine and Egyptian dialects), Spanish (Castilian and Latin American), French (France and Maghreb), German, Italian, Turkish, Hindi and Urdu. Each profile carries its own voice, greeting, register, dialect handling and objection wording, and each is evaluated separately for dialect authenticity, code-switching, formality, gender agreement and pronunciation. Portuguese and Indonesian already run as website languages and are provisioned as voice profiles on request, and further languages are added the same way — a language is a profile rather than a rebuild. We name the nine that are built and evaluated in depth rather than claiming every language on earth, and we will tell you exactly where any market you need stands before you commit.
What actions can the agent actually perform?
Actions are implemented as tools with strict contracts: idempotent, scoped to a single tenant and campaign, and unable to report success when the underlying operation failed. Typical tools book a calendar slot, write a disposition to the CRM, record a promise-to-pay, trigger a WhatsApp or SMS follow-up, or transfer the call to a human.
What happens when the agent cannot handle the call?
It hands off. Handoff is a first-class module rather than an afterthought: the agent transfers to a human with the conversation context already gathered, so the customer does not repeat themselves. Disputes, distress and anything outside the agent's remit are escalation triggers by design.
Is it available today?
No. It is in gated development on an isolated branch, behind feature flags that default to off, and it is deliberately excluded from the shipping v1.0 product. We are selecting a small number of design partners to run it against real traffic. Everything else on this site — the dialer, AMD, compliance controls, WhatsApp and SMS — is shipping today.
Will it replace our agents?
That is not the design goal. The agent is built to absorb the repetitive, high-volume, low-judgement calls so your people spend their time on the conversations that need a person. The handoff module exists precisely because some calls should never be finished by a machine.

Want it on your traffic first?

We are selecting a small number of design partners — teams with real outbound volume who want the agent shaped around their calls rather than a generic demo. Tell us what you run and we will tell you honestly whether you are a fit.