A call your bank can defend.
Telephony-grade voice agents that answer, verify, resolve and hand over — in Arabic dialects, Swiss German, French and English — with recording consent, retention and human escalation designed in before the first call is taken. Replacing an IVR is the floor. A service line that improves every month is the work.
ما هذه الممارسة
- A production system on a telephone number: carrier integration, call routing, speech recognition tuned to the languages your customers actually speak, a governed language model behind it, and the integrations that turn a conversation into a completed transaction.
- A regulated service channel. Recording notices, consent capture, retention and deletion, identity verification and the handover to a human are designed with your compliance function and tested before go-live, not patched after a complaint.
- Multilingual by construction. The agent recognises Levantine, Gulf and Egyptian Arabic alongside Modern Standard Arabic; it understands Swiss German dialects and replies in Standard German; it works in French and English, and it follows a caller who switches language mid-sentence.
- Measured against an evaluation set your team owns. Every intent, every dialect and every failure mode has test calls, and the release process runs them before any change reaches a customer.
وما ليست عليه
- A chatbot with a text-to-speech voice. Telephony has latency, noise, interruptions and no screen; an agent that was not designed for those conditions fails on the first real call.
- A replacement for your people. The agent handles what can be resolved within policy and hands everything else to a person with the full context of the call — never a cold transfer, never a caller repeating their story.
- A vendor's model behind an API you cannot see. Where sovereignty or residency demands it, speech and language models are deployed inside your own data plane, and the transcripts never leave it.
- Finished at go-live. Calls change, products change and callers find the edges; the practice runs the improvement cycle that keeps the agent accurate as they do.
القدرات
- Telephony and carrier integration
- SIP trunking with your existing carriers, number porting, call routing, queue integration and failover to your current IVR or contact centre, so the agent is a new front door on the number customers already know.
- Arabic dialect recognition
- Recognition and understanding across Levantine, Gulf, Egyptian and Modern Standard Arabic, with code-switching to English handled as the norm rather than an error, and numbers, dates and names captured correctly in both directions.
- Swiss German, French and English
- Recognition of Swiss German dialects with replies in Standard German or the caller's chosen language; French and English at native quality; explicit language selection or detection, at the caller's preference.
- Identity verification
- Knowledge-based checks, one-time codes over a second channel, and — where your policy and your regulator permit and the caller has consented — voice verification, each matched to the risk of the transaction it permits.
- Recording notice, consent and retention
- The notice a caller hears, how consent is captured and stored, what is retained, for how long and where, and how a caller exercises their rights over the recording — designed against the revised Swiss FADP, the GDPR, Jordan's Personal Data Protection Law and US state consent rules.
- Policy-bound resolution
- The agent acts only within written policy: which transactions it may complete, which limits apply, when it must verify again and when it must stop. Policy is versioned and reviewed like code.
- Human handover with context
- A warm transfer that carries the transcript, the verified identity and the resolved and unresolved intents to the person who takes the call, so the caller never starts again.
- Core-system integration
- Read and write integration with core banking, policy administration, scheduling, ticketing and customer-record systems, through your integration layer where one exists and through a governed adapter where it does not.
- Latency and conversation design
- Sub-second turn-taking, interruption handling, confirmation of critical values and a conversational register written for each language rather than translated from one.
- Evaluation and red-teaming
- An evaluation set of recorded and synthetic calls per intent and language; adversarial testing for prompt injection over voice, social engineering and out-of-policy requests; release gates that run both.
- Observability and quality review
- Transcripts, intent outcomes, handover rates and latency per call; sampled human review with a scoring rubric; drift alerts when recognition or resolution quality moves.
- Sovereign and on-premises deployment
- Speech and language models deployed inside your data plane — Swiss, EU, US, on-premises or a national cloud — with the control plane holding policy and observability but never audio or transcripts.
- 01Inbound callCarrier, number, queue
- 02Recording noticeConsent captured or routed unrecorded
- 03Language and dialectDetected or chosen by the caller
- 04IdentityVerification matched to risk
- 05IntentUnderstood and confirmed
- 06Policy checkMay the agent act, and how far
- 07ActionCore-system transaction
Available from every step, carrying transcript, identity and intents.
Human handover
كيف ننفّذ
Five phases, each closed by a gate your team signs. A voice agent is never switched on for all callers at once; it earns traffic one intent, one language and one caller segment at a time.
- 01Discover
- Listen to real calls, map intents by volume and risk, read the recording and consent obligations for each jurisdiction, and inventory the systems a resolution must touch.
- بوابة: Intent map, compliance requirements and integration inventory approved by the business, compliance and IT sponsors.
- 02Define
- Write the conversation design per language, the policy the agent may act within, the handover rules, the evaluation set and the target architecture, including where audio and transcripts live.
- بوابة: Policy, conversation design and data-plane decision signed; evaluation set agreed with the quality function.
- 03Build
- Integrate telephony, deploy speech and language models in the chosen plane, build the integrations, and run the evaluation set on every change.
- بوابة: Evaluation set passing at the agreed thresholds in every language; penetration and red-team findings closed.
- 04Prove
- Route a defined slice of real traffic — one intent, one language, one segment — with human review of every call, then widen the slice as the numbers hold.
- بوابة: Resolution, handover and quality scores at target across the pilot slice for the agreed period; compliance sign-off on recordings and consent capture.
- 05Operate
- Run the improvement cycle: weekly quality review, monthly policy and evaluation-set updates, new intents and languages added through the same gates.
- بوابة: Monthly steering review of quality, cost and the backlog of new intents; annual re-test of consent and retention.
ما تحصلون عليه
| المخرَج | الشكل | ما هو |
|---|---|---|
| Intent and risk map | Document and register | Every caller intent by volume, risk and permitted resolution, maintained as the agent grows. |
| Conversation design | Per language | Scripts, prompts, confirmations and escalation language written for each language and dialect group. |
| Policy specification | Versioned document | What the agent may do, under which verification, within which limits — the text your compliance function approves. |
| Compliance design | Document and evidence | Recording notice, consent capture, retention schedule, deletion procedure and data-plane mapping per jurisdiction. |
| The agent | Deployed system | Telephony integration, speech and language models, integrations and handover, running in your chosen data plane. |
| Evaluation set and harness | Test suite | Recorded and synthetic calls per intent and language with the tooling that runs them on every release. |
| Quality dashboard | Operational view | Resolution, handover, latency and quality scores per intent, language and week, with drift alerts. |
| Operating runbook | Document | How to add an intent, change policy, roll back a release, respond to an incident and answer a data-subject request. |
أين تكون مهمة
- الخدمات الماليةمميِّز، حيث تتقدّم AltuonClient service lines for private banks and retail banks, where verification, recording and handover are supervised.
- التأمينمميِّز، حيث تتقدّم AltuonClaims first notice of loss, policy queries and renewals, in the caller's language, with policy-bound resolution.
- الحكومة والقطاع العاممميِّز، حيث تتقدّم AltuonCitizen services in Arabic dialects on sovereign infrastructure, with human escalation for anything the agent may not decide.
- الاتصالاتأساسيHigh-volume account and technical support where an agent that resolves within policy changes the cost base.
- الرعاية الصحيةأساسيAppointment scheduling, reminders and triage routing under HIPAA and its European equivalents.
- الضيافة والعقاراتأساسيReservations, concierge requests and tenant services across languages and time zones.
- الطاقة والمرافقداعمOutage reporting and billing queries that spike unpredictably and must stay reachable.
تعاقدات تمثيلية
- مخطط تعاقدVoice agents for a Swiss private bank's client service lineA telephony-grade agent in Swiss German, French and English that answers, verifies, resolves and hands over — with recording consent handled the way FINMA expects.Switzerlandالخدمات المالية
- مخطط تعاقدA sovereign AI programme for a government ministryArabic-first models on national infrastructure, a governance framework a minister can sign, and citizen services moved onto it one at a time.Jordanالحكومة والقطاع العام
رؤى ذات صلة
- الصوت9 سبتمبر 2026Voice agents in Arabic dialects, and where they breakModern Standard Arabic is a written language. Callers speak Levantine, Gulf, Egyptian and Maghrebi, switch to English mid-sentence and read numbers in two directions. Here is where the systems fail, and what a bank or a ministry should demand before it lets one answer the phone.
- المشتريات9 سبتمبر 2026Procurement's questions about AI vendors, answeredWho owns the model weights? Where is the training data? What happens to our prompts? Can we leave? The questions procurement teams ask AI vendors are the right ones. Here are the answers a serious firm should be able to give, and the ones that should end the conversation.
أسئلة تطرحها المشتريات
Where do call recordings and transcripts live?
In the data plane you choose, per system: Switzerland, the European Union, the United States, your own premises or a national cloud. Speech recognition and the language model run inside that plane; the control plane holds policy, identity and observability metrics but never audio or transcripts. The choice is recorded in the architecture register before the first call is made, and the sub-processor list names every party that could touch a recording.
How is recording consent handled across Switzerland, Jordan and the United States?
Each jurisdiction's obligations are written into the compliance design during the Define phase: what the notice says, when it is played, how consent is captured and evidenced, and what happens when a caller declines — including routing to a person without recording where that is the lawful answer. Retention and deletion schedules follow the same design. Your compliance function signs it before the pilot, and it is re-tested annually.
What happens when the agent does not understand or is not allowed to act?
It says so, in the caller's language, and hands over to a person with the transcript, the verified identity and the intents so far. There is no dead end and no repeated story. Handover rates are measured per intent and language and reviewed weekly; a rising rate is a signal to improve the agent, never a reason to trap the caller.
Can the models run on our own infrastructure?
Yes. Speech and language models can be deployed on-premises or in a sovereign cloud, with the consequences made explicit in the proposal: infrastructure the client provides and operates, model options that fit that footprint, and a cost line for it. Where a hosted model is acceptable, the residency and sub-processor terms are written into the agreement.
How do you prove the agent is accurate before it takes real calls?
With an evaluation set built in the Define phase: recorded calls with consent, and synthetic calls per intent, language and dialect, including adversarial ones. Every release runs the set and must pass the agreed thresholds. Real traffic is then introduced by slice — one intent, one language, one segment — with every call reviewed by a person until the numbers hold for the agreed period.
Who owns the conversation design, the policy and the evaluation set?
You do. They are deliverables assigned to you on payment, along with the integrations built for you. Altuon retains its methods and tooling and licenses them for the life of the system. If the engagement ends, the agent keeps running on your infrastructure with your team operating it from the runbook.
How is the agent kept from being manipulated by a caller?
Policy is enforced outside the language model: the model proposes, a policy engine decides what it may execute, and identity verification is required before any transaction the policy marks as sensitive. Adversarial testing — prompt injection over voice, social engineering, out-of-policy requests — is part of every release gate, and findings are closed before the release ships.
Let us listen to your calls.
The Discover phase begins with real calls and your compliance obligations, and ends with an intent map you can defend. Request a proposal, or book a briefing for the executives who will sign off on the first pilot.