An AI voice receptionist is software that answers your business phone line, talks with the caller in natural speech, works out what they need and then acts on it. Call routing is the decision step: the system sorts each call into a category, such as sales, support, billing or urgent, and then transfers it to a person, takes a message, books an appointment or answers the question itself. It works at any hour and follows the routing rules you set.
This guide covers how a call reaches the AI, how the AI understands the caller, how the routing decision and the transfer work, the rules that apply in the EU and US, and what to check before you buy or build one.
What an AI voice receptionist actually does
A human receptionist does four things on every call: answers, listens, decides and hands over. An AI receptionist follows the same pattern.
- Answers every call straight away, including out of hours and when several calls come in together.
- Listens and asks. It turns the caller's speech into something it can reason about and asks follow-up questions when the request is unclear.
- Classifies the call into one of the categories you defined.
- Acts on that category. It may transfer the call, record a message, create a ticket, book a slot or answer a simple question such as opening hours.
The value comes from the last two steps. Answering a phone is easy. Sending each caller to the right place, and dealing properly with the callers who fit no category, is where these systems succeed or fail.
How a call reaches the AI
Most businesses want to keep the phone number they already have. The usual way to do that is a SIP trunk, a service that turns a phone call into internet traffic that software can handle. OpenAI's documentation, for example, describes buying a number from a SIP trunking provider, pointing the trunk at its Realtime API endpoint, and setting up a webhook that tells your application when a call comes in. Another option is to forward your existing number to a number on the trunk, either for all calls or only when nobody answers or the office is closed.
This choice matters more than it seems. It decides whether the AI answers every call or only overflow and out-of-hours calls. It also decides how transfers to staff will work and which provider holds your call data.
How the AI understands the caller
There are two main designs, and OpenAI's voice agent guidance describes both.
Speech-to-speech
A single model interprets the caller's audio, decides what to do, uses tools and replies in speech, all in one session. This keeps the conversation simple to run. The trade-off, in our view, is that you have fewer points at which to inspect or change what happens between hearing the caller and replying.
Chained pipeline
The system turns speech into text, a language model reasons over that text, and a text-to-speech engine speaks the reply. OpenAI suggests this path when each stage needs to be visible or replaceable. You can store the transcript, run policy checks and call internal systems before the agent says anything. For routing, that control is useful: the classification is a readable decision on a text transcript, and you can log it, test it and audit it.
Whichever design you pick, measure delay. OpenAI advises tracking backend time separately and comparing the median and the 95th percentile across similar calls. A receptionist that responds quickly on average but sometimes leaves several seconds of silence will still make callers hang up.
How call routing works, step by step
- Greeting and disclosure. The AI answers, names the business and states that the caller is speaking to an automated assistant (see the rules section below).
- Intent capture. It asks an open question such as "How can I help?" and listens to the answer. A keypad menu forces callers to fit your options. Here the caller uses their own words.
- Classification. The model maps what the caller said to one of your categories. A good design also returns a confidence level and has an "unclear" category, so the system can ask again instead of guessing.
- Detail collection. Depending on the category, it collects the caller's name, callback number, account or order reference, and how urgent the matter is.
- Routing decision. Rules combine the category with context: time of day, which staff are available, whether the caller is already a customer, and any priority flags.
- Action. It transfers the call, takes a message, creates a record in your CRM or helpdesk, books a time, or resolves the request and ends the call.
- Record. It logs the transcript, the category, the action taken and the outcome so you can review them.
Blind transfers and warm transfers
When the action is "put the caller through to a person", there are two ways to do it.
- Blind (cold) transfer. The call goes straight to the destination with no briefing. On SIP systems this is usually done with a SIP REFER message, which hands the call over. OpenAI's SIP documentation includes a refer endpoint that sends an active call to a phone number or SIP address. On Twilio, Ultravox's documentation describes blind transfers that redirect the call with the Dial verb.
- Warm (attended) transfer. The AI puts the caller on hold, calls the staff member, briefs them and connects everyone once they accept. Ultravox describes doing this on Twilio with the Conference API: hold music for the caller, a private "whisper" message giving the staff member context, and a key press to join. The AI can then drop off once everyone is connected.
Blind transfers are simpler and cheaper. Warm transfers mean the caller does not have to repeat themselves, and nobody is left ringing an empty desk. For sales enquiries and complaints, the warm transfer usually justifies the extra complexity.
Plan for the transfer that fails
The most common gap in these systems is what happens when nobody picks up. Decide in advance whether the AI should try a second person, return to the caller and take a message, or offer a callback time. Check the details of your chosen platform too. Ultravox warns that with its built-in cold transfer tool for Twilio, the AI call ends as soon as the tool is invoked, whether or not the transfer succeeds. A detail like that can decide whether a caller gets help or gets cut off.
Where the routing logic should live
Managed voice platforms make a quick start easy, but the platform then controls how your calls behave. In a voice receptionist we built, the system answers every call, day or night, on the client's existing phone number, then classifies each call and routes it. Speech-to-text and text-to-speech run through a single provider. Because a managed voice platform decides how your calls behave, we ran classification and routing on our own server, so the routing logic is ours to change. When a new department opens or a rule changes, nobody has to wait for a platform feature or work around its limits. You can read more in the AI voice receptionist case study.
When you assess any option, ask: who can change the categories and routing rules, how quickly, and is that logic stored somewhere you can export it?
The rules you need to follow
This is not legal advice, but these points come up on almost every deployment.
- EU: tell callers they are speaking to AI. Article 50 of the EU AI Act requires that people be informed when they are interacting with an AI system, unless that is obvious. The European Commission states that Article 50 applies from 2 August 2026 and that the notice must be clear and given from the start of the first interaction. A spoken disclosure in the greeting is the practical way to meet this.
- US: call recording consent. Federal law sets a one-party consent baseline, but 12 states, including California and Florida, require every party to consent to recording. A short recording notice at the start of the call covers callers from those states.
- US: outbound calls are a separate matter. In February 2024 the FCC confirmed that AI-generated voices count as "artificial" under the Telephone Consumer Protection Act. That mainly affects outbound calling. If your receptionist also places callbacks, get advice on consent before you switch that feature on.
- Everywhere: data handling. Transcripts and recordings contain personal data. Know which providers process them, where the data is stored and how long you keep it, and take local advice on data protection and recording rules in the UK and EU.
How to test it before it takes real calls
OpenAI recommends testing in stages: first controlled single-turn requests using synthetic speech, then replayed recordings of real people to test different voices, microphones and background conditions, and finally a simulated caller for full multi-turn conversations. It adds human listening to judge naturalness and pacing. For a receptionist, that means:
- Building a test set from real call reasons, including vague, angry and off-topic callers, and checking that each one lands in the right category.
- Testing every transfer path, including no answer, a busy line and a declined warm transfer.
- Testing out of hours, background noise, strong accents and callers who interrupt.
- Measuring the median and 95th percentile delay before the AI starts speaking.
- Reviewing a sample of live transcripts every week after launch and adjusting the categories as needed.
Questions to ask before you buy or build
- Can we keep our existing number, and will the AI take all calls or only overflow and out-of-hours calls?
- Which architecture does it use, and can we see a transcript and the classification for every call?
- Does it support warm transfers, and what exactly happens when a transfer fails?
- Who can change routing rules, and does that logic live in a platform or in code we control?
- Which systems will it write to (CRM, helpdesk, calendar), and how?
- How are AI disclosure and recording notices handled for callers in different places?
- Where are recordings and transcripts stored, and for how long?
If the answers point to custom routing, integrations with your own systems or code you own outright, a built system is likely to suit you better than an off-the-shelf product. Khive builds AI voice agents in code and hands them over, so the client owns the system.
If you are considering an AI receptionist, send us a brief covering your call volumes, the categories of call you receive and where each one should go, and we will tell you how we would approach it.