The phone rings, an AI receptionist picks up — and what it says in the first few seconds decides whether the caller stays on the line or hangs up feeling like they got nowhere. A long menu with eight options, a system pretending to be human instead of saying plainly what it is, silence that drags on too long before the first word, or an ad instead of a greeting — every one of these mistakes costs the call before the caller's actual question even gets asked, and that cost sits close to what gets counted for missed calls in a company: a caller who hangs up rarely calls back on their own.
This piece covers what a good opening line for an AI receptionist should contain, what the law actually requires here, how to give the caller a fast path to a human, and how to test whether a greeting even works before worrying about whether it sounds "natural."
What the AI Act actually requires
Transparency requirements under the EU's AI Act have applied since 2 August 2026: anyone using an AI system should be clearly informed that they're interacting with a machine, if that isn't already obvious. This covers chatbots, virtual assistants and automated customer-service systems among others — which plainly includes an AI receptionist answering the phone. The information has to be delivered at the latest at first contact, and in a way that's clear, understandable and accessible, including for people with specific needs.
What the rule does not give — and isn't worth pretending it gives — is one ready-made, mandatory phrase. There's no single sentence that has to be recited word for word; there's a requirement of clarity and timing (no later than the start of contact), and the exact wording is left to the business to set in its own tone.
What a good opening line is made of
Four elements, in this order, are enough to satisfy the clarity requirement without putting the caller off.
- 01Business name
- →02clear system disclosure
- →03what it can help with
- →04a path to a human
The order matters: the business name immediately confirms the caller dialed the right number, the disclosure removes uncertainty before it turns into irritation, and the statement of what the system can help with gives the caller a reason to stay on the line instead of asking for a human in the first second.
Three greeting variants — and when each one fits
The table below shows three variants that differ in length and in how quickly they offer a human. The duration is a rough estimate, not a measured value — every business should time its own recording with a stopwatch.
| Variant | Rough duration | When it fits | Risk |
|---|---|---|---|
| Short | approx. 5-7 seconds (estimate) | A simple line handling one type of request | The caller doesn't know what the system even covers |
| With a task list | approx. 10-14 seconds (estimate) | A business with several clearly different services | Too long a list starts sounding like a menu nobody listens through |
| With an immediate human option | approx. 8-10 seconds (estimate) | A business where a share of callers have urgent or complex cases | Offering "human" too early can discourage simpler, faster requests |
None of these is universally better — the choice depends on how varied the requests are and how large a share of callers would rather speak to a person right away.
Offering a human: right away or on request
The second question, beyond the greeting's content itself, is when the option to reach a human appears. It can be offered right in the first sentence, or left as a command available on request — saying "agent" at any point in the call, for instance. The second approach tends to work better where the system resolves most requests on its own, and the first where callers routinely end up with a human anyway.
What happens outside business hours
If the person the "agent" command routes to isn't actually working right then, the command shouldn't end in dead air or the same option repeating in a loop. The right fallback is a callback with a saved summary of the case, not a dead end — the caller leaves knowing someone will follow up, instead of guessing whether the call reached anyone at all.
Call recording: what to say up front, and what goes in the full disclosure
If the call is being recorded, a short mention at the start is enough — one sentence noting the call may be recorded — with the full details delivered separately, for instance via a link in a text message or on the website. The short mention doesn't replace the full disclosure; it just moves the details out of the first few seconds, where clarity about talking to a system matters most.
What belongs in the full disclosure
Among the information that should be provided when collecting data is a notice of the right to lodge a complaint with the supervisory authority — a GDPR obligation independent of whether the conversation is handled by a system or a person. There's no need to include the supervisory authority's address or contact details — a general statement of the right to complain is enough. What happens to the data from calls like these, physically, is covered in automation and GDPR.
Pace and pauses: how to measure whether the reply makes callers wait
OpenAI's own guidance for voice agents recommends repeating the same test scenarios and comparing task completion, audible response latency, interruptions and unwanted silence — keeping the same caller, model configuration, tools and call channel consistent between runs so the comparison actually means something.
How to measure without bothering real customers
The test doesn't need real callers: a handful of people from inside the business call the same line with the same pre-agreed question and note the second at which the business name is said, the second at which the human path appears, and whether the system talks over the caller at any point. Repeating the same scenario matters more here than the number of attempts.
The caller's language: what not to promise upfront
If a caller starts speaking Ukrainian or English, the system should respond according to what it's actually configured to handle — switch to a supported language, or, if that language isn't set up, say so clearly and route to a human. Promising language support in the greeting that the system doesn't actually have configured ends worse than never promising it.
Why customers still aren't used to this
At that share of businesses using generated voice, most callers still don't have the habit of recognizing a system by voice alone. That's why a clear self-introduction at the start matters more than whether the voice sounds "natural" — most callers can't tell on their own anyway, and ambiguity breeds more distrust than an openly stated disclosure ever would.
Do it yourself: testing your own greeting
- Write down 2-3 greeting variants that differ in length and in when the human option appears.
- Ask 5 people you know to call the line with the same, pre-agreed question.
- Note the second at which the business name is said and the second at which the human path appears.
- Flag any moments where the caller tried to interrupt the system before it finished speaking.
- Compare the variants against each other, rather than judging one variant on its own.
What it looks like when a system handles the greeting
Instead of relying on someone remembering the right wording, the whole process can run on one consistent pattern:
- 01call comes in
- →02business name and disclosure
- →03purpose of the call
- →04resolved or handed to a human
- →05summary in CRM
This isn't a promise that a system replaces talking to a human wherever the case calls for it — the principle of keeping a human in the loop on important decisions applies to phone handling too: where the stakes are high, the final call should still belong to a person.
The greeting and the full phone-call script are built by the AI reception and telephony service. The same tone of introducing itself as a system, just in a website chat window instead, is handled by the AI Chatbot service, and the confirmations and reminders sent to the client after the call go out through Automatic messages. If the client also needs to book a specific slot during that same call, the calendar side runs on Booking Systems, and everything gets recorded in one place by CRM and automations — no manual copying from a call note into a calendar. The difference between an assistant like this and a plain chatbot, visible in three process traits, is covered in AI agent vs chatbot, and what can realistically be handed to a system in everyday company processes — and what can't — is the subject of process automation in a company.
Frequently asked questions about the first seconds of an AI receptionist call
Does the opening line have to say the exact words "I am an artificial intelligence"?
The rule requires clear disclosure that the caller is talking to a system, but it doesn't dictate one specific phrase — what matters is that the disclosure comes no later than first contact and is understandable, not that it sounds identical across every business.
What happens if a caller asks for a human and no one's available?
The system should offer a callback with a saved summary of the case, rather than leaving the caller without a response or repeating the same option in a loop.
Is a short greeting always better than a longer one with a task list?
Not always — a short greeting works well for a single type of request, but where a business handles several clearly different matters, a task list helps the caller reach the right topic right away, as long as it doesn't run too long.
Does every sentence in the greeting need to mention call recording?
No, a brief mention at the start plus a link to the full disclosure delivered separately is enough — repeating the full notice in every sentence would overload the greeting without any real benefit to the caller.
What should happen if a caller speaks a language the system doesn't support?
The system should say so clearly and route the call to a human, rather than trying to continue in a language it isn't configured for — promising a language that doesn't actually work ends worse than not promising it at all.
How can you tell if a greeting actually works before rolling it out for good?
The simplest way is to ask a few people to place test calls with the same question and note the second at which each key element of the greeting lands, and whether anyone tries to interrupt the system along the way.
