WAJ Voice Agent

Overview

The WAJ Voice Agent answers a phone or web call and holds a real spoken conversation — it listens while you talk, can be interrupted mid-sentence, looks things up while the caller waits, and replies out loud in English or Arabic. It runs on the same WAJ AI Engine that powers the AI Assistant inside the app, so a voice agent and a chat agent share one brain: the same knowledge base, the same tools, the same conversation history.

It is a separate product from the in-app AI Assistant. Voice agents are set up for your business by the WAJ team rather than switched on from Settings — see "Getting one for your business" below.

Try it before you talk to anyone

There is a live demo you can call from a browser, no booking and no install:

1. Go to waj.ai/waj-voice-agent.

2. Pick an Accent from the picker.

3. Tap Start the call and allow microphone access when the browser asks.

4. Talk normally. Interrupt it mid-sentence — that is the point of the demo.

The panel beside the call lists topics you can ask about. They are suggestions, not buttons — say them in your own words. The Arabic version of the page runs the same demo with an Arabic voice.

This demo agent is a sales assistant. It answers questions about WAJ from the WAJ Help Center, and it can take your details and book a demo call with the team. It deliberately has no access to any real business's bookings or data, and it will tell you so if you ask it to change an appointment.

What a voice agent can do

• Answer questions from a knowledge base. Documents you supply are indexed and searched mid-call, so answers come from your own material rather than the model's guesswork, and the agent can say where an answer came from.

• Read live business data. Sales summaries, branch and staff performance, bookings over a period, inventory — pulled through connected tools while the caller waits. Exact figures are pushed to the caller's screen; the voice gives the rounded, speakable version.

• Walk a caller through a process. Step-by-step troubleshooting or setup guidance from your own guides, with an escalation path when the steps don't resolve it.

• Take real actions, where you have configured action tools. Opening a ticket, submitting a request, capturing a lead. Sensitive actions can be put behind an OTP check that is held and verified on the server, so nobody can talk the agent past it.

• Hold a natural conversation. The caller can cut in at any point and the agent stops talking and listens. Background noise and short "mm-hm" backchannels are distinguished from a real interruption before it yields the floor.

• Say numbers safely. Spoken output is rendered by rule, not improvised: account numbers are masked to their last digits, long decimals are rounded, tables go to the screen — and results are read back verbatim rather than paraphrased.

Languages and accents

The agent is engineered per language rather than translated. Holding phrases, apology lines, "the figures are on your screen", even hesitation sounds exist separately for each language, so English does not leak into an Arabic call.

• Arabic — the accent is a setting, not a fixed default. Gulf (Saudi, UAE, Kuwait, Qatar, Bahrain, Oman), Egyptian, Levantine (Jordan, Lebanon, Syria), Iraqi and Moroccan are all selectable, and Modern Standard Arabic is available where a formal register is wanted.

• English — American, British, Australian, Indian and Irish accents.

The call language is pinned from the start, so a foreign brand or product name cannot drag the agent into another language halfway through.

How it connects to your systems

The voice agent binds the same integration surface as every other WAJ agent:

• Tool servers — connect a tool server and choose which of its tools this agent may use. Credentials are injected on the server side and stripped from what the model sees, so the agent never handles a secret.

• Action tools — point-and-click HTTP actions configured once, optionally OTP-gated.

• Knowledge bases — documents are ingested, chunked and embedded into a searchable index the agent queries during a call.

• Shared history and memory — voice and text conversations sit on the same session history, and long-term user memory can be turned on or off per agent.

What a call feels like

• The agent starts speaking about three seconds after the caller finishes a question, because answers are streamed sentence by sentence rather than waiting for the whole reply.

• An interruption is acted on in roughly a second, and the audio already playing stops within milliseconds.

• End-of-turn is detected by meaning rather than by a fixed silence timer, so a short thinking pause mid-sentence does not cut the caller off.

• Audio time is metered and capped per business per day, with limits on how many calls can run at once.

Getting one for your business

Voice agents are configured per business rather than self-serve. Book a live demo call from the Voice Agent page — the demo agent itself can take your name, business name and phone number on the call, or you can use the form on the page.

Once a voice agent is set up, its behaviour lives in configuration rather than in code: provider, model, the voice and accent per language, how long a pause ends a turn, whether interruption is allowed, and the daily audio budget are all settings the WAJ team can change without redeploying anything. Changes are tested against recorded call scenarios and scored before they reach your callers.

Notes & limitations

• The public demo on the website is a sales agent. It cannot see, create, change or cancel appointments for any business — that is by design, not a missing feature.

• A live call needs a modern browser and microphone permission. Chrome, Edge and Safari are supported.

• Each business has a daily cap on audio minutes and a limit on simultaneous calls; heavy call volumes should be sized with the WAJ team up front.

• The voice agent is not part of the standard WAJ subscription and is not turned on from Settings — it is provisioned by the WAJ team.