Setting Up an AI Phone Agent: A Practical Guide Using fonio

The phone is the last analogue bottleneck

A trade business with eight employees receives between 20 and 40 calls on an ordinary working day. Some get answered. The rest hit a voicemail nobody enjoys speaking to, or simply ring out. Whoever fails to get through dials the next number. That lost enquiry appears in no statistic, because it never reached a system in the first place.

The same pattern shows up in medical practices where the line is permanently busy between 8 and 9 a.m. In property management firms with three hours of phone availability per day. In car dealerships where the service advisor is standing at a vehicle, not at a handset. In law firms whose front desk is unstaffed over lunch.

The problem is not a lack of digitalisation. These businesses have a CRM, a calendar, a ticketing system and an ERP. The problem is that the most common first touchpoint – the phone call – is the only channel without a queue you can work through. An email waits until someone reads it. A call is simply gone.

An AI phone agent addresses exactly that: it picks up when nobody else can, understands free-form speech, answers standard questions, books appointments and writes the outcome into the systems you already run. This article uses fonio as a worked example to show how that works technically, how setup actually runs, and which legal obligations come with it.

Disclosure: innFactory AI Consulting is a fonio partner. We implement the platform for clients and support selection, integration and compliance. The facts in this article come from publicly available sources; statements fonio makes about itself are labelled as such.

What an AI phone agent is – and what it is not

Not a traditional IVR

A classic IVR (“For appointments, press 1”) walks the caller through a predefined decision tree. Every path has to be modelled in advance. Whatever is not in the tree does not exist. Which is why callers reflexively press 0 to reach a human.

An AI phone agent has no decision tree. The caller states their request in their own words and the agent classifies it. Instead of maintaining paths, you maintain knowledge and rules.

Not a chatbot with a phone number

The difference between a chatbot and a voice agent is not the language model – it can be the same one. The difference is the physics of a live conversation.

In chat, a reply may take three seconds. On the phone, one second of silence is already awkward and two seconds feel like a dropped call. So the agent has to listen, transcribe, understand, possibly look something up in a knowledge base, compose a reply and speak it – all below the threshold at which a human says “Hello?”

Harder still is turn-taking: how does a machine know the caller has finished speaking? Answer after every pause and you interrupt people who were merely thinking. Wait too long and the agent feels sluggish. These two properties – latency and turn-taking – decide whether a caller finds the conversation usable or hangs up. Not the intelligence of the model.

Reading the vendor landscape

The voice agent market has diversified considerably since 2024. Three broad categories exist – stated without judgement, because each has its place:

Developer and infrastructure platforms. They provide building blocks: telephony connectivity, real-time media transport, speech-to-text, text-to-speech, orchestration frameworks. Maximum capability and freely chosen integration depth – in exchange you build state handling, error handling and the user interface yourself. A sensible choice for product teams making voice part of their own product.

Turnkey solutions for small and mid-sized companies. They ship a finished application with a web interface, number management, calendar integration, a knowledge base and reporting. You configure rather than program. The trade-off is predictability against flexibility. fonio sits in this category.

Enterprise CX suites. Rooted in the contact centre world, they bring workforce management, quality assurance, omnichannel routing and deep integration into existing service processes. Appropriate once the contact centre is an organisational unit in its own right.

The right category follows from a single question: who will look after the agent going forward? A business without an in-house engineering team is better served by a turnkey solution, even if a platform could theoretically do more.

fonio in profile

fonio GmbH is based at Neustiftgasse 73–75 in 1070 Vienna, Austria (commercial register FN 644725a, Vienna Commercial Court, registered 28 Jan 2025). The managing directors are Daniel Keinrath (co-founder and CEO) and Matthias Gruber (co-founder and CTO); operations started in autumn 2024.

Funding moved quickly: an angel round of 3 million euros in December 2025, then a seed round of 17 million US dollars (roughly 14.6 million euros) on 9 June 2026 at a 140 million dollar valuation, led by 20VC. Total funding now exceeds 20 million US dollars. In September 2025 fonio acquired the Linz-based provider fluently.

On traction, fonio states the following about itself: more than 10 million US dollars ARR in under twelve months (as of 11 Aug 2026), around two million automated calls per month, and more than 7,000 customers – the website now says “10,000+”. The team is described as roughly 60 people. Active markets are Austria, Germany, Switzerland, Italy, Poland, France, the UK, the Netherlands, Spain, Brazil and the US. Publicly named reference customers include Volkswagen, Brita and Storebox.

The product covers inbound and outbound telephony plus a WhatsApp chatbot as a separate product; email, chat and an AI-native CRM are announced. Features include 24/7 call answering, transcription and summarisation delivered by email or SMS to recipients that vary by topic, live transfer with if-then rules and business hours, and appointment booking via Google Calendar, Microsoft 365, cal.com and Calendly (Calendly requires a paid plan), plus fonio’s own scheduler with a booking URL and SMS fallback.

The knowledge base is filled from Q&A pairs, text PDFs (scans do not work), an automatic website crawl and live web search during the call. Questions the agent could not answer land in an inbox; the system proposes answers that must be confirmed before they take effect.

fonio PBX adds a full phone system: team routing individually, in parallel or sequentially across up to ten stages, voicemail, a browser softphone, blind transfer, a caller allowlist that bypasses the AI, and SIP access for desk phones. Outbound campaigns run via CSV upload (Team tier and above) or the outbound API; KYC verification is mandatory and additional carrier costs apply per destination country.

Pricing (fonio pricing page, as of 17 Aug 2026, net excl. VAT)

Solo includes one phone number and one user. Team adds three numbers, unlimited users, a dedicated SIP trunk plus outbound and campaigns. Scale adds a custom voice, zero data retention, an individual SLA, SSO and individual DPAs.

Additional minutes cost 15 euros per 100 minutes (12 euros annually), an extra number 7 euros per month, importing a number 5 euros per month. fonio offers a 30-day money-back guarantee but no conventional free trial. fonio also sells setup packages: Basic at a one-off 1,000 euros (8 hours per year), Extended at a one-off 2,500 euros (20 hours, one to two API integrations), and a care package from 99 euros per month.

When choosing a tier, watch concurrency rather than minutes. In practice the bottleneck is simultaneous conversations, not monthly volume. A practice with a call peak at eight in the morning needs more than one channel even if its total volume is modest. Your existing phone system statistics answer that question in five minutes.

How fonio works technically

Precision matters here, because a lot of imprecise material circulates online.

fonio describes its approach as a self-developed orchestration layer, built so that data stays in-house. As genuinely in-house development, fonio specifically names:

1. speech recognition tuned for difficult dialects,
2. its own turn-detection model that distinguishes between someone having finished and someone merely thinking – according to a CEO interview this reduced the churn caused by that problem from around 50 percent to under 1 percent,
3. real-time emotion recognition that adapts tone and pace,
4. the orchestration layer that coordinates these components with business systems in real time,
5. a new generation of self-produced voices.

As figures, fonio cites latency consistently below 800 milliseconds, RAG retrieval in 0.2 seconds even across 300-page PDFs, and – explicitly as its own claim – that “in around 90 percent of cases the agent resolves the request completely.”

On top of that, fonio’s own privacy policy for the web app names established subprocessors: OpenAI for the language model, Deepgram for speech-to-text, ElevenLabs for text-to-speech, LiveKit for real-time media and WebRTC, Twilio Ireland for telephony and Hetzner for hosting. Additionally New Relic, PostHog, Stripe, Attio and Microsoft Ireland.

The accurate description is therefore: a proprietary orchestration and speech-quality layer built on established speech and language model providers. The blanket claim “fonio has its own AI models” would be wrong. Equally, the circulating assertion that European-hosted language models are used is unsubstantiated and contradicts fonio’s own legal documents.

Why that layering is the actual lever

This is not a weakness but sensible division of labour. Language models such as GPT and speech synthesis models such as those from ElevenLabs are interchangeable components improving on their own rapid schedule. No 60-person company will run foundational research there – nor would it need to.

The quality of a phone agent is decided elsewhere. When does it start speaking? What does it do when a circular saw runs in the background? How does it handle a heavy regional accent? How quickly does it find the right passage in a 300-page document without creating a pause? How does it bridge the wait while an API responds? Those questions are answered by the orchestration layer. That is the part you cannot buy off the shelf.

Nine steps to your own phone agent

The sequence below follows fonio’s official documentation at fonio.info. Registration runs via app.fonio.ai/register; fonio advertises being “ready to test in 60 seconds”. Realistically, plan half a day to a full day for a production-ready setup – the effort sits in the prompt and the knowledge, not in the technology.

1. Review the dashboard. It shows call history, minutes consumed, numbers, users and concurrent calls. Use this view later as a control instrument: it tells you whether your tier matches reality.

2. Create the assistant. Choose the greeting and the voice. Important for multilingual operation: select the “Multi” voice and “Multi” language.

3. Generate company knowledge. Enter your own website URL and click “Generate” – fonio crawls the site and extracts the core information automatically. The result is a usable first draft, not finished knowledge. Then add Q&A pairs for your ten most common questions and upload text PDFs (scans will not work).

4. Connect a phone number. Three options: book a fonio number (carrier per the pricing page is Twilio), import your own number for 5 euros per month, or connect your existing system by SIP trunk. In practice the most common route is call forwarding at your existing provider – fonio documents the GSM codes *21* (all calls), *61* (no answer) and *67* (busy). That makes the rollout low-risk: the agent initially handles only what would otherwise have rung out. The SIP trunk is designed as a static IP-based peer trunk without registration and is documented for 3CX, STARFACE, easybell, Zadarma, Twilio, Telnyx and Yeastar; registrar-based systems require a conversation with support.

5. Configure call routing. If-then rules, business hours or round-the-clock operation. This is where you decide which requests the agent handles itself and which it hands to a human.

6. Set up the post-call email. Fixed recipients, or – considerably more useful – recipients that depend on the topic: appointment requests to reception, complaints to the service lead, quote requests to sales.

7. Write the prompt. The heart of the matter. fonio provides templates for answering machine, reception, appointment booking and first-level support. More on this in the next section.

8. Enable post-call SMS and appointment booking. Connect a calendar (Google Calendar, Microsoft 365, cal.com, Calendly) or use fonio’s own scheduler with a booking URL and SMS fallback.

9. Wire up the API and webhooks. Three hooks are available: the inbound webhook before the conversation starts, which passes caller data into the prompt as {{variable}} and persists it as {{inboundContext.fieldname}}; the API request during the call, which can be triggered conditionally; and the post-call webhook after the conversation. API documentation lives at app.fonio.ai/api/docs, API keys in the workspace.

Test before going live. Voices can be previewed as audio free of charge; real test calls consume minutes from your allowance. Plan a handful of realistic test conversations – including the uncomfortable ones: callers with strong accents, callers with background noise, callers asking for something that is not in the knowledge base.

Prompt practice: where phone agents actually fail

The prompt is not the agent’s manual, it is its job description. Six points determine quality:

Role and context. Who is the agent, which company do they work for, and what is the setting? “You are Anna from Müller Property Management. You take calls outside office hours.” That is more precise than any personality description.

Conversation goal. What should exist at the end of a successful call? A booked appointment? A qualified callback request with name, property and issue? Without a defined goal the agent chats pleasantly and produces nothing usable.

Escalation rules. When does it transfer, when does it end the call? Write explicit triggers: water damage, heating failure, an angry caller, a legal question. An agent without an emergency exit is more dangerous than no agent at all.

Handling the unknown. The most important rule in the entire prompt: what does the agent do when it does not know something? The useful instruction is to say so openly, capture the request and promise a callback – not to improvise something that merely sounds plausible.

Pronunciation and glossary. fonio supports a glossary for technical terms and proper nouns. Use it. An agent that mispronounces your company name or flagship product loses credibility in its first sentence. The same applies to place names, model designations and medical terminology.

Data capture. fonio can capture email addresses, phone numbers and customer numbers precisely and extract variables into structured fields. Define those fields up front – they are the interface between the conversation and your CRM.

Beyond that, speaking rate, sensitivity, creativity (temperature) and background noise can be tuned. The agent can also send DTMF tones and therefore navigate third-party IVR menus – useful in outbound scenarios.

An honest note on multilingual operation, taken directly from fonio’s documentation: in multi-language mode the agent does not reliably switch languages on its own. The switch has to be explicitly permitted in the prompt, and fonio recommends a neutral voice for it. If you plan bilingual operation, test it deliberately. Overall, more than 20 voices and more than 25 languages are available in every tier; a custom voice is limited to the Scale tier.

Taking automation further: from the call into your systems

An agent that answers calls and sends an email has solved the phone problem but no process problem. It gets interesting when the conversation outcome lands where work already happens.

The route there is the post-call webhook. After every call fonio sends the transcript, summary and extracted variables to a configured URL. What happens next is up to you. In practice an automation platform handles distribution – and fonio explicitly recommends Make.com or n8n as middleware in its own documentation, providing fixed backend IPs for whitelisting. Alongside email and SMS, native integrations include HTTP request (GET, POST, PUT, PATCH, DELETE against arbitrary APIs), Google Sheets, web search, WhatsApp, plus Make and n8n.

Typical flows we build for clients:

• New enquiry → CRM lead. The agent captures name, company, request and budget indication; n8n creates the lead, assigns the right salesperson and sets a follow-up.
• Fault report → ticket. Property, unit, fault type and urgency become a structured ticket in the service system; critical cases additionally trigger a notification.
• Appointment request → calendar plus confirmation. Booking in the calendar, SMS confirmation to the caller, entry in the patient or customer record.
• Status enquiry → live ERP lookup. Using the API request during the call, the agent queries order status live and states it on the phone. This is the moment a caller notices the difference from a voicemail box.
• Known caller → personalised opening. The inbound webhook pulls customer data by phone number before the conversation starts and passes it into the prompt.

We describe how such automation flows are built in detail on our workflow automation with n8n page and in the article AI agents for SMEs: why n8n is the ideal launchpad.

GDPR, recording and the legal side

This section is why many phone agent projects in mid-sized companies stall after the pilot. The technology works – the sign-off is missing.

What fonio says about data processing

fonio’s FAQ states verbatim: “Our servers are located in Nuremberg, Germany.” The privacy policy names Hetzner with a data centre location in Germany. Put robustly: hosting with Hetzner in Germany (fonio names Nuremberg).

In its footer fonio carries the badges “made in Austria”, “GDPR compliant” and “EU AI Act compliant” – these are vendor self-declarations, not external certifications. Publicly linked documents include the terms of service, a DPA at fonio.ai/de/avv, privacy policies for the website and the web app, confidentiality declarations for Germany and Austria, and a Switzerland addendum. For enterprise, the pricing page mentions individual DPAs and NDAs.

Three points belong in an honest assessment:

Zero data retention exists only in the Scale tier – for telephony and WhatsApp. If ZDR is a precondition for your internal approval, you are budgeting from 499 euros per month upwards.

Third-country transfers do occur. OpenAI, Deepgram, ElevenLabs and LiveKit are US providers. Transfers are covered by the EU-US Data Privacy Framework and standard contractual clauses. That is a viable and widely used route – but a point your data protection officer needs to see and assess rather than overlook.

Not substantiated and therefore not claimed here: ISO 27001 certification, regular penetration testing, or a contractual no-training clause. Such statements circulate in third-party sources but cannot be confirmed from the public documents. A 30-day retention period for recordings also appears only in technical documentation and should be settled contractually if it matters to you.

The practical sticking point: consent to recording

This is where law and product design collide directly.

Our assessment – explicitly not legal advice: recording and transcribing phone calls in Germany requires the caller’s consent. Without it you risk violating Section 201 of the German Criminal Code (violation of the confidentiality of the spoken word), on top of the information duties under Art. 13 GDPR. In practice this is handled with an announcement at the start of the call.

fonio supports this: automatic deletion of call data is configurable (Assistant → Technical → Recording). If the caller declines recording, the recording and transcript are deleted. At that point, however, no post-processing and no extraction are possible either – no transcript, no summary, no structured fields, no CRM handover. fonio recommends ending the call or transferring to a human in that case.

This is the point to think through before rollout rather than after: your process needs a defined path for the caller who says no. Build that path or you end up with either a legal violation or a caller stranded in nowhere.

Further obligations

Outbound campaigns are additionally subject to Section 7 of the German Unfair Competition Act (UWG). Marketing calls without prior express consent are unlawful – automation does not change that, it only increases volume and therefore exposure.

The EU AI Act requires that users can recognise they are speaking with an AI. The enforceable transparency obligations have applied since 2 August 2026. In practice: the opening announcement must name the AI, not obscure it.

Works council involvement may be required depending on how the system is configured – particularly where the agent also captures employee conversations or enables analysis that permits conclusions about performance and behaviour.

All of these are solvable. You just need someone to raise them before go-live rather than after. How we set that up in a structured way is described under AI compliance; the assessment must always be made case by case.

When it pays off – and when it does not

Good fits

An AI phone agent pays off where many similar calls hit few lines and the content is well documentable: appointment scheduling, status enquiries, callback requests, initial qualification, recurring standard questions. Medical practices, trade businesses, property management, car dealerships and workshops, law and tax firms, cleaning and security services, hospitality and restaurants.

The business case is usually simpler than expected. The benchmark is not the hourly cost of staff but the lost enquiry. A business that fails to answer five calls per day, with a 20 percent close rate and an average order value of 400 euros, loses roughly 400 euros of revenue per day on paper. Against a tier starting at 99 euros per month that is not a difficult decision. Put your own numbers into that calculation, not ours.

Poor fits

Emotional escalations. A caller with water pouring through the ceiling, a relative in a medical emergency, a complaint after the third failed attempt – those belong with a human. The agent must recognise these cases and hand over immediately.

Complex advisory work. Where every conversation runs differently and the answer requires judgement, a phone agent can pre-qualify but not advise.

Poor knowledge hygiene. If opening hours, prices and responsibilities are not documented internally, the agent will expose that mercilessly. That is not a weakness of the technology – a new employee would face the same problem. But it does mean the real effort sits in tidying up your knowledge.

Wrong expectations. The roughly 90 percent resolution rate fonio cites is a vendor claim across its entire customer base. For your use case it only becomes meaningful once you have measured it against your own calls. Plan a pilot phase with evaluation, and define beforehand how you measure success: requests resolved, data captured correctly, transfer rate, dropped calls.

Our role as a fonio partner

innFactory AI Consulting is a fonio partner. You can source fonio through us – but that is not where our work ends. A licence on its own does not solve a process problem; what matters is fitting the agent into your existing landscape. Specifically, we handle:

Selection and assessment. First the question of whether a phone agent is the right tool at all – and if so, at what scale. We look at call volume, peak load, concurrency and process maturity before anyone talks about tiers.

Setup and telephony integration. Numbers, call forwarding or a SIP trunk against your existing system, routing rules, business hours, escalation paths. We know the pitfalls with registrar-based systems and with telephony landscapes that grew over the years.

Prompt engineering and knowledge build-out. The part that determines quality. We write the prompt, structure the knowledge base, define the fields to extract, build the glossary for technical terms and proper nouns – and test against real conversation patterns rather than idealised ones.

Integration into your systems. Post-call webhook, inbound context, API requests during the call, connections to CRM, ticketing, calendar and ERP via n8n. This is the difference between a pleasant answering machine and a genuine process component.

Training your team. Your staff need to be able to adjust the agent themselves – add new Q&A, change transfers, read the reporting. We hand over operational capability, not dependency. Suitable formats are listed under AI training.

Compliance support. Announcement wording and consent path, deletion concept, roles and permissions, works council topics, EU AI Act documentation – and adding fonio as a sub-processor to your own processing agreements. That last point is routinely forgotten in do-it-yourself projects and surfaces later during an audit. For legal questions we work with a specialist lawyer for IT law.

Conclusion

An AI phone agent is no longer future technology. The decisive problems – sub-second latency, reliable turn-taking, dialect recognition – are solved, and providers like fonio deliver that as a configurable product rather than a development project. Starting with a simple call forward is so low-risk that the technical hurdle effectively disappears.

The remaining work sits elsewhere: in the prompt, in the knowledge, in the integration and in the legal setup. Take those four seriously and within a few days you have an agent capturing enquiries that used to be lost. Skip them and you have an expensive answering machine with speech synthesis.

Your next step: pull the statistics from your phone system for the last four weeks. How many calls went unanswered? At what times of day? That number is the most honest basis for the decision.

• Book a conversation – we look at your use case and say openly whether it is worth it
• Workflow automation with n8n – the path from call to CRM
• AI compliance – consent, DPA and EU AI Act done properly
• fonio.ai and fonio.info – vendor website and official documentation