AI voice and customer communication

Voice is where a great deal of business still happens and where most automation has historically been terrible. The interactive voice menus everyone has suffered through are the reason people press zero immediately. Current speech models make a genuinely conversational agent possible, which changes what is worth attempting — but it also raises the stakes, because a system that sounds human is one people will believe.

We build voice agents for bounded tasks with a clear handoff, and we disclose that the caller is speaking to a machine.

Why this is worth doing properly

The technical constraint that matters is latency. A conversation tolerates about a second of silence before it feels broken, and that budget has to cover speech recognition, model inference, and speech synthesis. Architecture decisions — streaming rather than batching, a smaller model for simple turns — are driven by that budget more than by anything else.

The second constraint is interruption. Humans talk over each other constantly, and an agent that cannot be interrupted mid-sentence feels immediately robotic. Barge-in handling is a core requirement rather than a refinement.

The third is legal. Recording a call and processing a voice both engage privacy law, and consent requirements differ sharply between jurisdictions — in some US states every party must consent to recording, and in the EU a voiceprint can be biometric data with the heightened protections that carries. Outbound calling adds telemarketing rules on top, which in North America are strict and actively enforced.

Where this work usually goes wrong

Hiding that it is a machine

Beyond being dishonest, it is increasingly a legal requirement — EU AI Act transparency obligations and several US state laws point the same way. It also backfires: callers who work it out mid-conversation are more annoyed than those told at the start.

No fast route to a human

An agent that cannot hand off traps the exact callers whose problems most need a person. Handoff should be available on request, immediately, without a negotiation.

Unbounded scope

An agent asked to handle anything will confidently mishandle the unusual cases. Narrow scope with confident escalation outperforms broad scope with occasional invention, particularly on a channel where there is no written record for the caller to check.

Ignoring latency until the end

A pipeline assembled without a latency budget produces an agent with awkward pauses that no amount of prompt work fixes. It is an architecture problem and it has to be designed for from the start.

Recording without consent

Consent rules differ by jurisdiction and by which party is calling. Getting this wrong on an outbound campaign at volume is a compliance incident, not a minor oversight.

What this covers

AI telephone agents

Conversational agents on real phone numbers, scoped to defined tasks with disclosure at the start of the call.

Inbound customer-service agents

First-line handling for common enquiries, with immediate handoff on request or on confusion.

Outbound lead follow-up agents

Follow-up calls on inbound enquiries, operating within telemarketing rules and calling-hour restrictions.

Appointment-booking agents

Booking against live calendar availability, with confirmation sent in writing afterwards.

Lead qualification by phone

Structured qualification against agreed criteria, with a written summary into the CRM.

Customer reminder calls

Appointment and payment reminders, usually the highest-return and lowest-risk voice automation available.

Multilingual voice assistants

Agents operating in several languages, with quality verified per language rather than assumed.

Human handoff systems

Warm transfer with full context passed, so the caller does not repeat themselves.

CRM-connected call workflows

Every call logged against the contact record with transcript, outcome and next action.

Call summaries and transcription

Automatic transcripts and structured summaries, which are frequently valuable independently of the agent itself.

Conversation analysis

Aggregate analysis of what callers actually ask, which is a genuinely useful research output.

Customer sentiment classification

Flagging calls that went badly for review, rather than trusting a sample.

Automated call reporting

Volume, resolution rate, handoff rate and duration reported per period.

How the work runs

1. Pick a bounded task

   Reminders, booking, qualification or first-line triage. We do not start with open-ended customer service, because the failure modes are most visible exactly where the caller is already frustrated.

2. Design the conversation

   Explicit disclosure, a defined path, and a handoff trigger on request, on repeated confusion, or on any topic outside scope. The escalation logic gets more design attention than the happy path.

3. Build to a latency budget

   Streaming speech recognition, streaming synthesis and barge-in handling, measured end to end. If the budget cannot be met, the task is wrong for voice and we say so.

4. Handle consent and recording properly

   Recording notice and consent capture appropriate to each jurisdiction called, with retention periods set deliberately rather than left at a vendor default.

5. Test against real calls

   Evaluation on recordings of genuine calls, including the difficult ones — accents, background noise, interruption, callers who go off-script.

6. Launch narrow and monitor

   A small share of traffic first, with every call transcript reviewed initially, widening only once handoff and resolution rates hold.

What you receive

• Scoped conversation design with disclosure and escalation logic
• Deployed voice agent on a real number
• Latency benchmarks measured end to end
• Calendar and CRM integration
• Consent, recording notice and retention policy
• Transcript, summary and analytics pipeline
• Evaluation results against real call recordings
• Monitoring and review process

You probably need this if

• Calls go unanswered outside office hours.
• No-shows are a material cost.
• Inbound enquiries wait days for a first call back.
• Nobody knows what callers actually ask about.
• Your existing phone menu is the most complained-about part of the business.

What we build and work with

A voice pipeline is a latency problem before it is anything else, and the components are chosen accordingly.

Telephony provider

Twilio or a regional carrier for numbers and call handling, with number provisioning that satisfies local registration requirements.

Streaming speech recognition

Transcription that emits partial results as the caller speaks, since waiting for an utterance to finish costs most of the latency budget.

Language model with streaming output

Response generation streamed into synthesis rather than produced in full first.

Streaming speech synthesis

Low-latency voice generation with barge-in support so the caller can interrupt naturally.

Calendar and CRM integration

Live availability and contact records, so bookings are real and history is written back.

Transcript and recording store

Recordings and transcripts retained under a defined policy, with access controlled and expiry enforced.

Call analytics

Resolution, handoff and sentiment measured across all calls rather than a sampled few.

What changes once this is in place

Calls answered outside office hours

Inbound handled at times when the alternative was a voicemail nobody returns.

Reminders that reduce no-shows

Automated reminder calls, typically the clearest and most measurable return in this category.

Every call has a written record

Transcripts and structured summaries against the CRM record, valuable even where the agent itself handles little.

Humans get the calls that need them

Fast handoff with context passed, so escalation feels like continuity rather than starting again.

Consent and recording handled correctly

Jurisdiction-appropriate notice and retention, rather than a vendor default applied everywhere.

How this differs by market

The work is the same craft everywhere. What changes is the law, the language and the buying culture — and those change enough to matter.

European Union

The AI Act requires disclosure that a person is interacting with an AI system. Call recordings are personal data under the GDPR, and voice characteristics used for identification are biometric data with heightened protection. Retention periods must be defined and enforced, not left to the vendor default.

Nordics

Consent expectations around recording are stringent and consumer trust is easily lost, so disclosure should be prominent rather than minimal. Multilingual quality genuinely varies here: model performance in Finnish in particular should be tested against real calls rather than assumed from published benchmarks.

United States and Canada

Recording consent varies by state, with several requiring all parties to consent. The TCPA governs automated outbound calling with express written consent requirements and significant class-action exposure, and both countries operate do-not-call registries that must be screened against before any outbound campaign.

United Arab Emirates

Arabic voice support with dialect variation is a real engineering consideration, and Gulf Arabic performance differs from Modern Standard Arabic in most models. Telecommunications regulation governs automated calling, and many customers will prefer WhatsApp to a phone call regardless of how good the agent is.

Not legal advice. Regulatory summaries on this site describe how we scope and build, and are current to our latest review. Verify the operative text with qualified counsel in the relevant jurisdiction before relying on it.

How we know it worked

Resolution rate without handoff, handoff rate, average handling time and end-to-end latency. Latency is reported as a distribution rather than an average, because the slow tail is what callers actually notice.

For reminder use cases the measure is the change in no-show rate against a pre-deployment baseline, and we insist on capturing that baseline before launch. Without it the improvement claim is unverifiable.

We review transcripts of failed and escalated calls continuously rather than sampling, because the aggregate metrics look acceptable long after individual call quality has started to degrade.

Estimates are labelled as estimates. Any figure on this site that describes a range is a planning estimate with its assumptions stated, not a measured client outcome. We do not publish client results without the client's permission and a date.

Related services

Sectors where this is usually the lead engagement

These are the industries where this discipline is typically the first thing a client buys rather than something added later. The link goes to a page written for that sector specifically, with a paragraph on this service and on every other one.

• Healthcare & Clinics — One of the fastest-growing sectors for digital investment, and the one where the constraints on what you may say and where data may sit shape everything.

It appears on all thirty sector pages, because every one of them carries a paragraph on all twenty-two services. This list names only the sectors where it tends to lead.

Where we deliver this

This service is delivered across the European Union, the Nordic countries, North America and the United Arab Emirates. The craft does not change; the law, the language and the buying culture do. Consent regimes, invoicing mandates and payment conventions differ enough between markets that a campaign or a system built for one frequently cannot be used unchanged in another.

Each country page sets out what actually differs there and what it means for scope — all 32 countries and 10 cities are listed here. A few of the markets we work in most:

• Germany
• Netherlands
• France
• Sweden
• United States
• United Arab Emirates