How to choose an AI receptionist: a European buyer's guide

Seven questions that separate a working AI receptionist from a demo that falls apart in production, plus the EU-specific rules, the red flags, and the cases where you should not buy one at all.

01Does it replace your telephony or sit alongside it?

Some products require porting numbers, installing softphones, or replacing your PBX. Others integrate at the carrier or SIP level and leave numbers, phones, and contracts untouched. The second kind fails safer: if you cancel, your telephony is exactly as it was.

02Which languages, and can it switch mid-call?

European call traffic is multilingual as a baseline, not a feature. A caller who starts in one language and drops into another mid-sentence is normal in most EU markets. A receptionist that handles only the demo language will embarrass you in the first week.

03Where does the audio go? (GDPR and the AI Act)

Call audio is personal data. Under GDPR you need a data processing agreement, a clear answer on storage location and retention, and a lawful basis for recording that matches your member state's rules. Under the EU AI Act, callers must be informed they are talking to an AI, so any vendor promising an agent that "passes as human" is selling you a compliance problem.

• Data residency: is audio processed and stored inside the EU, and can the vendor name the region?
• Encryption: is one customer's data isolated from another's, or is it one shared bucket?
• Model training: is your audio used to train models shared with other customers?

04What happens when the AI cannot help?

Every AI receptionist has a ceiling. The product question is what happens at the ceiling: a live transfer to the right employee with context is the good answer; a dead end, a voicemail, or "please call back later" is the bad one. Ask what the employee sees when a call is handed over.

05Does it know the caller, and does the data land anywhere?

A receptionist that starts every call from zero handles each conversation politely and learns nothing. The difference in practice: does it read prior conversations with that number, and does each call produce a transcript, a summary, and a CRM update, or just a notification that someone called?

06Will the vendor show you their real numbers?

Every vendor claims high resolution rates. Ask for the measured version: what share of conversations does the AI resolve end to end, on production traffic, with sample sizes? Vendors who measure will show you. As a reference point, we publish ours: 48.5% of conversations at one client resolved with no human, and 41% of callers hang up in the first exchange, both with methodology, on our data page. Any serious vendor can produce the equivalent.

07What is the pricing shape, and where is the cap?

Common models: per minute, per call, or a subscription sized to team and call volume. The number that matters is not the cheapest quote; it is the comparison against what missed calls already cost you. And ask about the failure case: if something loops or misfires, what caps the bill? A vendor without a runaway cap has not thought about production.

Red flags, any vendor

• No live line you can call before signing. The product is the call; a vendor hiding the call is hiding the product.
• Promises that callers "won't be able to tell it's an AI". Illegal to operate that way in the EU, and callers resent it when they find out.
• No straight answer on data residency within one business day.
• No transfer-to-human path, or one that dumps the caller into voicemail.
• No transcripts. If you cannot audit what the AI said, you cannot manage it.
• Claimed resolution rates with no sample size and no methodology.

When not to buy one

The one test that beats this whole guide

Call the vendor's own line, unannounced, in your customers' languages. Interrupt it. Ask something weird. Ask for a human. Ten minutes of that tells you more than any feature table, this page included.