AI voice agents for support: where they pay off, by industry

In a mystery-shopping study of 2,105 car dealerships and 200 independent service shops in the US, published by Pied Piper in July 2025, dealerships using an AI voice agent on the service line successfully handled 91% of calls with no employee involved. The rest is the interesting part: in 9% of calls the agent could not resolve the request, and when it tried to transfer the call to a person, the transfer failed 56% of the time — on average 88 seconds into the conversation.

An AI voice agent for customer support is not won on the easy calls. It is won or lost at the border between what it can do and what has to reach a person — and that border does not sit in the same place in a restaurant, a dental practice or an online shop. Below we take six types of business in turn and answer the same four questions for each: which calls the agent takes, which system it writes the outcome into, what you measure and when it hands over to a person. The worked example is a restaurant — a fictitious one, built by us, which you can call today.

Why “is a voice agent worth it?” has no general answer

According to Eurostat, only 5.2% of Romanian companies with at least 10 employees used some form of AI in 2025 — the lowest share in the EU, where the average rose to 20.0%, from 13.5% a year earlier. A Romanian business owner asking today whether they need a voice agent is among the first — and has nobody in their own sector to copy.

In customer service departments, the money has already moved. A Gartner survey of 199 customer service leaders, run in April–May 2026, shows that their AI spending grew by 38% while the function’s total budget grew by 2%. When AI money grows faster than the budget, the risk is automating around the technology rather than around the call. At a small business, the mistake looks like this:

• The agent gets “the whole phone”. It handles bookings well, stumbles on complaints, and the customer remembers the second experience, not the first.
• The agent writes nowhere. It takes the request, but someone has to copy it from a transcript into the calendar the next day — you have paid for a more expensive voicemail.
• There is no way out to a person. In the Netherlands, Consumentenbond reported in August 2026, based on more than 10,000 respondents, that 7 in 10 consumers who had used a chatbot got no answer at all, and almost 90% say it has become harder to reach a person. Chatbots, not voice — but the mechanism is the same.
• Nobody measured anything beforehand, so nobody can say what the agent delivered.

The limits of the technology itself — how well it understands spontaneous spoken Romanian, the pause the caller hears, the obligation to say it is AI, call recording and the cost per minute — we covered separately, in what an AI voice agent on your business phone can and cannot do. The question here is a different one: which call, and in which business.

Four signs a call is right for an AI voice agent

It is not the type of business that decides, but the type of call. A call is a good candidate when it ticks as many of the signs below as possible — and the last one is mandatory:

• It is frequent and the calls look alike. The same few questions, in the same order, day after day.
• It comes in peaks. Calls pile up exactly when your staff are busy with the customers in front of them.
• It has value outside business hours. A call at 10 pm that today reaches nobody is a request that, as far as the business is concerned, does not exist.
• It ends with a structured outcome written into a system: a reservation, an appointment, a status read from a database.

Customers seem to draw the same line. A survey by Avaya (a contact centre vendor) of 510 US consumers, from January 2026, shows that for a fraud or security issue 70% prefer a person, but for simply tracking an order only 19% do. And 56% say they prefer an AI assistant if it solves their problem quickly.

There is one more condition, shown by the figures at the start: the way out to a person has to work. In the Pied Piper study, the AI agent “beat” people on conversation quality but secured the appointment slightly less often — 86% versus 90% at dealerships with people on the phone. The agent is good on the predictable path and weak precisely where the call leaves it.

You do not buy a voice agent for the phone. You buy it for one type of call that ends with a row written into a system.

The worked example: a restaurant, from the call to the table on the floor plan

A restaurant ticks all four signs: bookings look alike, calls come exactly when the dining room is full, people call late in the evening, and the outcome is a table blocked on a floor plan. A Harris Poll survey of 2,065 US adults from 2025 — commissioned by Hostie, which sells AI agents for restaurants — shows that 63% prefer to call when contacting a restaurant and that 69% would probably give up on a restaurant if nobody answers.

To avoid discussing this in the abstract, we built a complete example: Magic AI Rezervări. It is our product and we present it as a demonstration. The “Magic AI” restaurant is fictitious, the voice agent is called Ioana, and on the test number +40 373 806 836 (a Romanian number) you can make, change and cancel reservations in the test system. Against the grid above, it looks like this:

• Which calls it takes: new reservations, changes, cancellations and questions about the menu, in Romanian or English, 24/7, including when the restaurant is closed. Returning customers are recognised: greeted by name, offered their usual booking and their favourite table, with allergies and occasions remembered — the name is confirmed, not asked for again.
• Which system it writes to: the restaurant’s real floor plan. The agent checks free tables in real time, knows which tables can be joined for a group and which one is “by the window”, and the reservation appears instantly in the management app, with a live floor plan, list and timeline. The customer gets a booking code, plus an SMS and an email with a management link, in their own language.
• What you measure: how many reservations the agent created, especially outside opening hours, and how no-shows and cancellations change before and after — in your own system, not against someone else’s average.
• When it hands over to a person: if the voice service does not answer, the customer is automatically transferred to the restaurant. Beyond that, a complaint or an event that has to be negotiated deserves a person from the restaurant anyway.

The second half of the work comes after the booking. The agent calls the customer 24 hours and 2 hours beforehand (the intervals are configurable), and the SMS or email goes out only if they do not answer. Why it matters: a meta-analysis published in BMJ Open (Robotham et al., 2016), covering 21 studies (8,345 patients with notifications and 7,731 without), found 25% fewer no-shows among those who received electronic notifications — 15% versus 21% — and voice reminders appeared more effective than text ones. A caveat: the data come from healthcare, not restaurants; we did not find a scientific study measuring the effect on restaurant reservations.

Restaurant operators also put reminders first — as an opinion, not a measured effect: in OpenTable’s October 2026 report for the UK, a no-show at a table for two costs £55.70 on average, and 43% of operators consider confirmation reminders the most effective measure. How often guests fail to turn up varies widely, though: Bloom Intelligence, a restaurant software vendor, measured across its own network of about 50 locations, most likely in the US, 3.0% no-shows on phone bookings and 3.9% on online ones. A “typical rate” does not exist; you only learn yours from your own system.

Two details apply to any agent: nobody should be able to find out about or cancel someone else’s reservation, and the agent must be protected against manipulation and limited against abuse — at Magic AI, all three are built into the product. And a remembered allergy is health information; treat it as such.

Prices shown in October 2026, excluding VAT, are fixed monthly: Small, up to 15 tables, €249/month, setup €490, 250 reservations per month; Medium, 16–40 tables, €490/month, setup €890, 600 reservations per month; Large, over 40 tables, €890/month, setup €1,490, 1,200 reservations per month. A pay-per-result option is also available, from €149/month. Setup includes configuring the floor plan, the agent and the dedicated number. The most useful test is not the feature list, though: call +40 373 806 836, ask for a table for four on Saturday evening, move it to another time, then cancel it. The details are on the Magic AI Rezervări page.

By industry: what it takes, where it writes, what you measure, when it hands over

Medical or dental practice

The right calls for the agent are new appointments, rescheduling, cancellations, reminders and questions about opening hours and the address. The outcome is written into the practice’s calendar, and the main metric is the no-show rate before and after — here the BMJ Open meta-analysis applies directly. Any call about symptoms, treatment or an emergency goes to a person, no discussion.

The difference from a restaurant is legal. Data concerning health are a special category: Article 9 of the GDPR prohibits processing them in principle, with exceptions such as the provision of healthcare by or under the responsibility of a professional subject to professional secrecy. Recital 35 adds that information collected when registering for healthcare services is also data concerning health, and the Court of Justice of the EU (Case C-21/23 Lindenapotheke, 2024) has read the notion broadly: data from which a state of health can merely be inferred count too. Even “I would like an appointment for …” can therefore be special-category data: you need a legal basis under Article 9, a data processing agreement with the agent’s provider, restricted access, care over what gets transcribed and a clear rule on how long transcripts are kept — retention and automatic deletion are the part we work on in data protection.

A benchmark comes from England, with a different health system: in the GP Patient Survey 2026, with 654,714 responses, 56.9% of patients said it was easy to contact their practice by phone — so roughly 4 in 10 did not find it easy.

Car repair shop

Two calls dominate: booking a service or the periodic roadworthiness test (ITP, Romania’s equivalent of the MOT), and “is my car ready?”. The first is written into the workshop calendar; the second requires the agent to read the job status from the work order — without access to it, the only possible answer is “a colleague will call you back”. There are two metrics: how many calls ended with an appointment being offered and how many transfers to a person succeeded. Diagnosis, a disputed quote and any discussion about warranty go to a person.

The data again come from the US. Pied Piper found in 2025 that 9% of customers who called a service department hung up without being offered an appointment, compared with 13% in 2024, and independent shops offered an appointment on 79% of calls, versus 90% at dealer groups. Car Wars, which sells telephony to dealerships, estimates about 19 million missed opportunities in 2025 out of roughly 53 million service calls, with 53% of missed calls ending up in voicemail. We found no comparable data for Romania.

Hotel or guesthouse

Here the strongest sign is the value of the hours outside business hours: availability enquiries come in the evening, from someone still choosing where to stay, and the pre-arrival questions repeat from guest to guest. The agent writes into the reservation system and only makes sense if it can read it. You measure the requests taken outside business hours and how many turn into bookings. Groups, negotiated rates and any complaint during the stay go to a person. We found no recent, independent data on unanswered calls at hotels, so we quote no percentage.

Online shop or deliveries

The “where is my order?” call is almost the ideal candidate — in the Avaya survey, only 19% of customers prefer a person for simply tracking an order. The condition is integration: the agent has to read the status from the order system and from the courier’s system. Without it, all it can say is “check your email”, which is exactly what the customer did before calling. The metric is the share of status calls closed without a person. Disputed returns, a parcel marked as delivered but never received and any dispute about money go to a person.

IT helpdesk or utilities

The agent is good at triage: it takes the problem, classifies it, opens the ticket and answers the repetitive questions about opening hours or the status of an incident. The outcome is written into the ticketing system, and the metric is how many tickets are classified correctly at the first attempt. Romania’s energy regulator, ANRE, has already set the principle of a way out to a person: in the consolidated version of ANRE Order no. 83/2021 that we consulted, call centre pick-up time is measured from the moment the customer chooses the human-operator option until a person takes the call, with a guaranteed level of 8 minutes and compensation of 0.20 lei (Romanian leu) per minute when it is exceeded. It applies only to electricity and gas suppliers, but the principle is clear: the robot can stand at the front, as long as the path to a person is guaranteed.

One thing is never handed to the agent: resetting a password or two-factor authentication on the strength of a phone call. CISA and the FBI warn that the Scattered Spider group calls helpdesks precisely to convince staff to reset passwords or transfer MFA tokens. An access request made by phone goes through strong verification, carried out by a person — a topic we cover under cybersecurity.

Where an AI voice agent does not pay off

• Heated complaints. An angry customer wants to be heard by someone who can make a decision.
• Negotiation. A price, a deadline or an exception to the rule calls for judgement and authority, not a script.
• Complex B2B sales. Someone buying a project asks questions nobody has asked before.
• Rare, highly varied calls. If every call is different, there is nothing to automate.
• Access resets, for the reasons above.

Some authorities are already spelling the rule out. In the Netherlands, the consumer authority (ACM) and the data protection authority (AP) asked, in a joint opinion from October 2025, that companies using AI chatbots also offer a way to contact a person and say clearly when the user is talking to AI. If your project has no way out to a person, it is not finished.

How to start: one type of call, measured

1. 1For a week, log every call by type — reservation, appointment, status, question, complaint — and the time it came in. Meanwhile, call the test number +40 373 806 836 yourself and note the moment you would want to talk to a person.
2. 2Pick a single type: the one that ticks the most of the four signs and ends with an outcome written into a system.
3. 3Decide which system the agent writes to and check whether that system has an integration that allows it. If it does not, sort out the integration first, then talk about the agent.
4. 4Write the handover rule — when, to whom, what happens if nobody answers — and test it yourself, several times, before a customer hears it. This is where the transfers in the Pied Piper study failed.
5. 5Run the pilot alongside the manual process for two to four weeks, the way we work in our AI agency projects, and expand only if the metrics show value.
6. 6Compare before and after, in your own system: calls taken outside business hours, outcomes written correctly, successful transfers, no-shows.
7. 7Only then move on to the second type of call. The Autonomous Voice Agents package in our packages has exactly the pieces you need along the way: a custom agent, integration with Twilio or your existing telephony, a workflow for appointments, FAQ and escalation, plus a dashboard with transcripts.

Conclusion

An AI voice agent does not fit an industry; it fits a type of call — one that repeats, comes when your people are busy, is worth something even after closing time and ends with a row written into a system. Everything else has to reach a person, reliably. If you want to look together at which of your calls is the first candidate and what should be measured, let us talk for 30 minutes, with no obligation.

Sources

• Pied Piper Prospect Studies — PSI Service Telephone Effectiveness 2025, Business Wire press release, 7 July 2025 (silicon.co.uk); WardsAuto, 21 July 2025 (wardsauto.com)
• CBT News — interview with Stephane Ferri, CEO of Car Wars, 12 March 2026 (cbtnews.com)
• Eurostat — use of AI in enterprises, 2025, news item of 11 December 2025 (ec.europa.eu/eurostat)
• Gartner — press release of 26 August 2026, survey of 199 customer service leaders (gartner.com/en/newsroom)
• Avaya — “Signals of Connection”, 17 February 2026 (avaya.com)
• Consumentenbond, as reported by onderneming.nl, 27 August 2026; ACM and AP, joint opinion, 2 October 2025 (acm.nl)
• Robotham D. et al. — “Using digital notifications to improve attendance in clinic: systematic review and meta-analysis”, BMJ Open, 2016, DOI 10.1136/bmjopen-2016-012116
• OpenTable — Restaurant Impact Report, as reported by restaurantonline.co.uk, 6 October 2026
• Bloom Intelligence — data on restaurant no-shows, 6 August 2026 (bloomintelligence.com)
• The Harris Poll for Hostie — survey of 29 May–2 June 2025 (hostie.ai)
• Regulation (EU) 2016/679 (GDPR), Article 9 and Recital 35; CJEU, Case C-21/23 Lindenapotheke, press release no. 159/24, 4 October 2024 (curia.europa.eu)
• Ipsos for NHS England — GP Patient Survey 2026, 9 July 2026 (ipsos.com)
• ANRE (Romanian Energy Regulatory Authority) — Order no. 83/2021, Performance standard for the supply of electricity and natural gas, Article 16, consolidated version of 1 July 2022
• CISA and FBI — Advisory AA23-320A, Scattered Spider, revised 29 July 2025 (cisa.gov)
• CryptoITData — Magic AI Rezervări, /restaurant page, accessed 13 October 2026