Calls turned into records and tasks by voice and conversation systems.

The value is in what happens after the call.

Demos show the conversation. The saving comes after it: the call attached to the right customer, the promise pulled out as a task, the delivery date changed in the system, the follow-up scheduled because somebody said "next week" out loud.

We build that whole path. Audio becomes a transcript and a summary. Names, dates and amounts are matched against your records, and changes reach the ERP or CRM through the same narrow, permissioned tools an agent uses. Messages to customers are drafted and held for a person by default, and refunds, discounts and cancellations always belong to a person.

How well does speech recognition handle Arabic?

Less well than English, and worse on dialect. The Open Universal Arabic ASR Leaderboard (Wang, Alhmoud and Alqurishi, presented at Interspeech 2025) tested open-source speech models on five multi-dialect Arabic test sets. The best of them averaged a word error rate of 25.71%, roughly one error for every four words. Every model did best on Modern Standard Arabic and declined significantly on dialects such as Egyptian and Khaliji, which the authors put down to training data weighted towards MSA.

A call in the Gulf can mix Gulf Arabic, MSA and English in one sentence, so no published benchmark, that one included, predicts how your calls will read. We measure word error rate on your own recordings, per language, before launch and after it. A glossary of your product, place and customer names goes in, and low-confidence passages are marked in the transcript for a person to check.

The speech models run self-hosted, Whisper-class models among them. Whisper was trained on 680,000 hours of labelled multilingual audio (Radford and colleagues, 2022). Recordings are sensitive and call volumes are predictable, which makes running the models on your side the usual choice.

It answers a narrow band of calls.

Questions with one right answer held in your systems suit it: opening hours, the status of an appointment, where an order is. Everything else goes to a person with the transcript and a suggested reply already loaded.

At the handover the caller does not have to repeat anything, and the person taking over sees what was said and what the system found.

On a call or in a chat, five things always go to a person:

• Quoting a price that is off your price list
• Agreeing a refund, a discount or a change to what was charged
• Promising a date the system has not confirmed
• Giving medical, legal or financial advice
• Sending the first message to someone who has not contacted you

Money, dates and first messages are blocked in the tools and the systems behind them, so no caller can talk the assistant past them. Advice is a topic and cannot be blocked like an action, so advice questions are routed to a person and tested with questions written to tempt it.

Every channel answers to the same rules.

The same retrieval and the same rules sit behind WhatsApp, web chat, phone and email, so an answer stays the same whichever way the question arrived. Chat answers come from your own documents with the source attached, and a question the documents cannot answer goes to a person. Phone runs through your existing provider.

WhatsApp sets rules of its own. Its Business Messaging Policy lets a business message only people who have opted in, allows free-form replies within 24 hours of the customer's last message and approved templates only outside that window, and requires any automation to offer prompt, clear and direct escalation paths, such as a transfer to a person in the chat, a phone number or email. The in-chat handover to a person and the template flow are built into the system.

Recordings stay where you decide.

Speech models run on your own hardware, in your own cloud tenancy or on our servers in Germany. No recording trains a model outside your company.

The rules on recording calls differ by jurisdiction and by what the caller is told. The consent prompt and the retention schedule are built into the system, worded by your legal adviser.

Built and tuned on your own recordings.

• Audio and channel survey. Where conversations happen, their audio quality and languages, and which ones go unrecorded today.
• Transcription pipeline. Models chosen against your recordings, with speaker separation, punctuation, Arabic handling and your glossary.
• Understanding layer. Summaries, entities and actions tuned against transcripts your team has marked up.
• Write-back and rules. Into the ERP or CRM, with the gate on anything outbound or financial.
• Pilot on one team. One desk for two weeks, side by side with how it is done today. Then widen.

Priced once, and yours on full payment.

One fixed price covers the build, paid milestone by milestone. Running costs, speech and model inference plus hosting, are estimated from your call and message volume before you commit. On full payment the source code passes to you with the glossary and the evaluation set built on your audio, and warranty terms are in pricing and engagement.

Tell us which channel carries most of your conversations, on a free 30-minute call or through the contact form.