AI Voice Agent for Bookings: Lessons from BookKaru
What an AI voice agent for appointment booking can really handle, what it costs to build and run, and the lessons we learned building BookKaru.
Qazi Ehsan
Xenova Solutions

An AI voice agent for bookings is software that answers your phone, understands what the caller wants, checks your calendar and books, moves or cancels the appointment while the caller is still on the line. It works well for simple, repeatable bookings when it has a narrow job, live access to your diary and a quick route to a human. It works badly when it is asked to handle everything, sounds slow, or traps callers who want a person.
Last updated: October 2026
We learned most of this building BookKaru, a voice booking agent for Urdu and Roman-Urdu speakers. Below is what a voice agent actually does on a call, what owners worry about, what it costs, and the lessons we would apply to any clinic, salon or trades business thinking about one.
What does an AI voice booking agent actually do on a call?
Every turn of the conversation runs through the same loop:
- Telephony: the call arrives on a normal phone number and is streamed to the agent.
- Speech-to-text: the caller’s words are transcribed as they speak.
- Language model: it works out the intent (book, reschedule, cancel, question) and the details it still needs, such as service, date, time and name.
- Tools: it queries your booking system for free slots and writes the booking back.
- Text-to-speech: it replies in a natural voice.
- Hand-off: if the caller asks for a person, or the agent is unsure, it transfers the call or takes a message and sends a text.
The part people underestimate is the tools step. A voice agent that cannot see your real availability is just an expensive answering machine. If your diary lives in a booking platform with an API, integration is usually straightforward; if it lives in a paper book or a spreadsheet, fix that first.
Will customers hang up when they realise it’s a bot?
This is a common worry for owners, and the short answer is: some will. Callers mostly want their problem solved quickly, but a slow, robotic or pushy agent loses them. Three things decide whether people stay on the line:
- Speed. Long pauses before each reply feel broken on a phone call. Set a response-time target (for example, roughly a second) and test it on real phone lines, not just on a laptop.
- Transparency. Say up front that the caller is speaking to an automated assistant and that they can ask for a person at any time. Trying to pass as human backfires when it slips.
- A real exit. “Speak to someone” should work at any point, every time. If no one is free, the agent should take a message and promise a call back within a stated time.
Many owners compare an AI agent with a perfect receptionist. The fairer comparison is with what actually happens today: calls going to voicemail while you are with a client or on a job. Measured against that, an agent that books the simple calls and passes on the rest is an improvement for many small businesses.
What can it handle well, and what should stay with a human?
Keep the scope narrow at launch. Trying to automate every type of call tends to frustrate callers; starting with bookings and FAQs and expanding later is safer.
- Good fit: new bookings for standard services, rescheduling and cancellations, opening hours, prices you publish anyway, directions, parking, and “what should I bring?” questions.
- Needs care: bookings that depend on judgement (which practitioner, how long an appointment needs to be), deposits and payments, and callers with strong accents or poor signal.
- Keep with a human: complaints, anything clinical or legal, emergencies, and upset or vulnerable callers. The agent’s job here is to recognise the situation and pass the call on immediately.
If you already use AI in chat or email support, the same principles apply; we cover them in how to automate customer support with AI without losing the human touch.
Our experience: lessons from building BookKaru
Our experience: BookKaru is a voice booking agent we built for Urdu and Roman-Urdu speakers. To handle the way real callers speak, we fine-tuned open-weight models (Llama and Qwen) on around 900 real call-centre recordings. The main lessons we took away:
- Real calls beat scripts. Callers mix languages, interrupt, change their minds and give details in any order. Real recordings capture this in a way written scripts do not.
- Narrow scope wins. A short list of tasks done reliably is better than a long list done sometimes.
- Confirm before you commit. Reading the booking back (“Tuesday the 4th at 3pm, is that right?”) catches transcription mistakes before they reach your diary.
- Plan the failure paths first. What happens when the line drops, the slot vanishes mid-call, or the caller says nothing? These paths decide how the agent feels.
- Fine-tune only when you must. For most English-language businesses, a strong general model with good instructions and access to your data is enough. We explain the trade-off in RAG vs fine-tuning.
You can see more of what we have built on our work page.
How much does an AI voice booking agent cost?
There are two costs: building it and running it. Here is how the main options compare.
| Option | Best for | Time to launch | Cost model |
|---|---|---|---|
| Off-the-shelf AI receptionist | Standard English bookings on a supported booking platform | Days | Monthly subscription, often plus usage |
| Custom voice agent (our AI Agent Pilot) | Other languages or dialects, unusual booking systems, strict data rules | 4–6 weeks | From £6,500 to build, plus running costs |
| Human answering service | Calls that need judgement and warmth | Days | Monthly or per-call fee; may only take messages unless given diary access |
| Voicemail plus missed-call text-back | Very low call volumes | Hours | Low, but nothing is booked during the call |
Running costs are mostly per minute of call time. Telephony is the smallest part: Twilio, for example, lists UK local numbers at $3.50 a month and $0.01 per minute to receive calls at the time of writing. Speech-to-text, the language model and text-to-speech together usually cost more than the phone line, and prices vary widely by provider and change often, so get a per-minute estimate based on your own call volume before you commit.
| Running cost | How it is charged | What to check |
|---|---|---|
| Phone number and call minutes | Monthly fee plus per minute | Example: Twilio UK local number, $3.50/month and $0.01/min inbound |
| Speech-to-text | Per minute of audio | Accuracy with your callers’ accents and languages |
| Language model | Per amount of text processed | Longer calls and bigger models cost more |
| Text-to-speech | Per character or per minute | More natural voices usually cost more |
| Hosting and monitoring | Monthly | Custom builds only; includes call logs and alerts |
What about privacy and the law?
A voice agent processes personal data: names, phone numbers, and often recordings and transcripts. In the UK, the ICO’s right to be informed guidance says you must give people privacy information at the time you collect their data, so a short spoken notice at the start of the call, plus an updated privacy notice on your website, is the sensible baseline. The ICO also publishes guidance on AI and data protection (currently under review following the Data (Use and Access) Act) that is worth reading before you start.
In the US, the rules bite hardest on outbound calls. In February 2024 the FCC ruled that AI-generated voices count as “artificial” under the Telephone Consumer Protection Act, so automated reminder or follow-up calls that use an AI voice face the same consent rules as other robocalls. An agent that only answers inbound calls is a different situation, but take advice if you plan to call people back automatically. We are not lawyers, so treat this as a starting point, not legal advice.
How to launch a voice agent without annoying your customers
- Pull a week of call data. Count calls by type and time of day so you know what you are automating.
- Pick two or three call types to start with, usually new bookings, rescheduling and opening hours.
- Connect your real booking system so the agent sees live availability.
- Write the hand-off rules: who gets transferred calls, and what happens out of hours.
- Test with staff and friendly customers on real phones before going live.
- Start after hours or on overflow only, then review transcripts weekly and widen the scope as it proves itself.
FAQ
Is an AI receptionist better than hiring a part-time receptionist?
They do different jobs. A person handles judgement, warmth and complex calls; an AI agent covers evenings, weekends, busy periods and simple bookings without breaks. Many small businesses get the best result from both: the agent takes overflow and after-hours calls, and staff handle the rest.
Do I need a custom-built voice agent or an off-the-shelf tool?
If your calls are standard bookings in English and your booking platform is supported, an off-the-shelf tool is often enough. A custom build makes sense when you need another language or dialect, an unusual booking system, strict data handling, or complex rules about who can be booked when.
What if the agent books the wrong time?
Design for it. The agent should read every booking back before confirming, send a text or email confirmation, and log every call with a transcript so mistakes are easy to spot and fix. Review those logs weekly for the first month.
Talk to us about a voice booking agent
If missed calls are costing you bookings, we can help you work out whether a voice agent is worth it and what the first version should do. Book a free 30-minute scoping call, or see our fixed-price starting points on the packages page; a custom voice agent would typically start as an AI Agent Pilot.