Back to Services

Conversational systems

Voice Agents

Voice agents succeed when they feel fast, interruptible, and aware of context. We design the audio loop, agent logic, tool calls, and escalation behavior as one system — not a chatbot bolted onto a phone tree.

What a voice agent needs before a live call.

Realtime voice UX with barge-in and silence handling

Conversation orchestration with CRM and tool integrations

QA on transcripts, fallback language, and human handoff

Conversation design, latency budget, then handoff.

Each element follows the last with purpose and momentum—like a well-tuned engine where every component has its place and timing.

  1. 01
    Conversation design and success criteria
  2. 02
    Voice pipeline and latency budget
  3. 03
    Tool execution and CRM write-backs
  4. 04
    QA, escalation, and launch metrics

Proof System

Voice launches include transcript review, interruption handling, fallback language, and measurement for completion and escalation — so support leads can trust the channel.

Every claim is backed by observable evidence, every decision leaves a trace, and every result can be traced back to its source—just like notes on a designer’s table.

Realtime voice UX with barge-in and silence handling
Conversation orchestration with CRM and tool integrations
QA on transcripts, fallback language, and human handoff

How bilingual and Arabic voice agents are tested

QuantaloomAI tests a bilingual voice agent before it is put on a live line. The set is built from the client's own call recordings, and only from recordings the client has consented to use. Those calls supply the words, the pauses, the mid-sentence language switches, and the tasks the desk actually handles. A studio sample is not the script. A hospital desk, a clinic line, or a support queue does not sound like one.

Each item is reviewed by dialect and by task. Dialect labels are Gulf, Levantine, Egyptian, and Modern Standard Arabic when those varieties appear in the consented recordings. Tasks are the jobs the client named, such as a booking, a status question, or a request to reach a person. The note is written against the client's criteria for that item. It is not an accuracy figure, and it is not a promise that a dialect will behave the same way on another account.

Code-switching is included when the recordings contain it. The pairs this work prepares for are Arabic-English and Urdu-English. A turn that starts in one language and finishes in the other is one item. The agent is expected to stay with the caller rather than reset the conversation at the switch. A switch that is not in the recordings is not added to the set.

Names, addresses, and numbers are read back before they are written anywhere. The read-back is its own task in the set. If it is wrong, the item fails review even when the rest of the turn sounded fluent. The failure is kept so a later change can be checked against the same item.

When the agent is unsure, it hands the call to a person and passes a transcript of what was said. Unsure means there is no grounded next step: a missing fact, a request outside the tools it was given, or audio it cannot resolve. The person does not start from silence, and the caller is not left repeating the question.

The same set runs again when the model, the prompt, or the vendor changes, so that change is seen before callers are back on the line. Runbooks name who reviews a failed item, how a prompt or tool change is recorded, and what is watched after launch: handoffs, calls that end without a completed task, and transcripts sampled from live traffic. Monitoring continues that review beside the evaluation set.

The architecture of a support agent is described in Voice AI agents for customer support, and the choice between a voice agent and a chatbot is covered in when a voice agent fits better than a chatbot.

Voice agents for clinic and hospital front desks

A clinic or hospital front desk answers the same kinds of calls through the day. A patient calls to book a visit, reschedule one that is already on the book, ask for a reminder, or get directions to the building. A voice agent can take that call, read the details back, and write the result into the schedule the desk already uses. Before the visit, the same line can send a reminder and collect a short digital intake, so the chart is not empty when the patient arrives.

Scheduling stays with the clinic's existing system. The agent does not keep a second calendar. It reads the open slots that system already publishes, writes the booking, and records a reschedule in the same place. Every action is logged: who called, what changed, and when. Front-desk staff can open that log later instead of reconstructing the call from memory. Anything clinical or unclear is handed to front-desk staff, along with a transcript of what was said. The agent does not give medical advice. A missing fact, a request outside the tasks that were agreed, or audio the agent cannot resolve follows the same handoff.

The work is scoped per requirement. A typical engagement runs 8–16 weeks from a signed scope to a line the desk can use. Hospitals that already run HMIS Pro, our hospital management system, can connect the agent to that record. HMIS Pro is live in hospitals, including Modern Hospital Karachi. Teams in Pakistan can read the hospital management software page for Pakistan, and groups in the Gulf can read the hospital management software page for the GCC.

Voice agent questions

Can you build Arabic or bilingual voice agents?
Yes. QuantaloomAI builds voice agents that take calls in more than one language, including Arabic together with English, when that is the conversation the client actually runs. The languages and the tasks are named in discovery. A build does not promise that every dialect will behave the same way.
How do you test dialects?
Evaluation sets are built from the client's own call recordings, and only from recordings the client has consented to use. Items are scored per dialect that appears in those recordings — Gulf, Levantine, Egyptian, and Modern Standard Arabic when they are present — and per task the client named. The notes stay with the engagement. They are not a public accuracy figure.
What happens when the agent is unsure?
The agent hands the call to a person and passes a transcript of what was said. Unsure means it does not have a grounded next step: a missing fact, a request outside the tools it was given, or audio it cannot resolve. The caller is not left repeating the same question.
Do we own the code?
Code and IP ownership transfer to the client on full payment, per the contract.
How is pricing set?
Pricing is a scoped quote after discovery. The quote follows the tasks, the languages, the tools the agent may call, and the evaluation work around them.
Can a voice agent handle clinic appointment booking?
Yes. A voice agent can book a visit, reschedule an existing appointment, send a reminder, give directions, and collect a short digital intake before the visit. It writes those changes into the clinic's existing scheduling system, and every action is logged. If the request is clinical or unclear, the call is handed to front-desk staff.
What happens when a patient asks a clinical question?
The agent does not give medical advice. It hands the call to front-desk staff, with a transcript of what the patient said, so a person can take it from there.

Ready to build

Scope a voice agent.

Case study: VoiceFlow

When story, table, and speed align, that's when remarkable things happen. Let's begin.