If your prompt tells an AI voice agent to "book the appointment using the booking tool" and that tool isn't connected, the agent will often still tell the customer they are booked. The call sounds perfect, the customer is happy, and no booking exists. This is one of the most common and costly mistakes in new deployments. Here is how to spot it and the pattern that prevents it.
Build an agent that books reliably: Start free on Edesy — Rs 50 free credit, no demo needed.
How it happens
In one deployment, the agent's instructions referred to a booking step by name, copied from an earlier version of the setup. The step wasn't configured in the live agent. The agent did what the script said conversationally: it agreed a time, said "your visit is booked", and ended the call. Nothing was created in the CRM, so no one showed up for some of those visits.
The same thing happens when:
- a tool is renamed but the prompt still uses the old name;
- a tool's API key or URL expires and calls start failing;
- a prompt is copied from another agent that had different tools.
How to find it
- Pull a day of calls where the outcome or transcript says a booking was made.
- Count the bookings that actually exist in your calendar or CRM for the same day.
- If the second number is lower, open a few of the missing ones and read the transcripts.
Do this in the first week of any new agent. See the first-week monitoring routine.
The fix: confirm on the call, create after the call
For most bookings, the simplest reliable pattern is:
- On the call: the agent offers two slots, agrees one, and repeats the date and time clearly. It says the team will confirm before the visit, not "it is booked in our system".
- After the call: the agreed date and time are extracted as structured fields.
- Automatically: those fields create the meeting or visit in your CRM or calendar and assign it.
This removes the live tool as a point of failure. If the post-call step ever fails, you still have the agreed slot as data and can recover it. More on this: post-call CRM automation.
When you do use live tools
Live tools are still the right choice for checking real-time availability or looking up an order. If you use them:
- Name only tools that exist. Every tool in the prompt must be configured on that agent.
- Tell the agent what to do if a tool fails: "If you cannot check availability, take the customer's preferred time and say the team will confirm."
- Never let it claim success it can't see. "Only say it is booked if the booking tool returns a confirmation."
- Test after every change with a few calls to your own phone.
See connecting custom API tools to your agent and prompt rules from real deployments.
A deployment that uses this pattern
A rooftop solar installer in West Bengal books free site surveys with the confirm-then-extract pattern: the agent agrees the slot on a Bengali call, and the meeting is created in the CRM after the call. It has booked 300+ surveys so far. Read the solar installer case study.
Start
Create a free account,
add meeting_date and meeting_time fields to your agent, and call your own phone to book a test
visit. Want us to review your booking flow? Talk to us.