AI Voice Agent Booking the Wrong Time? Fix the Time Zone

automation
Two identical brass twin-bell alarm clocks with blank cream faces on an oak shelf against an indigo wall, one showing ten o'clock and the other nine o'clock, illustrating an AI voice agent booking appointments one hour off
Same appointment, two clocks. After a daylight saving change, a voice agent can confirm one and book the other.

Why is my AI voice agent booking the wrong time?

An AI voice agent books the wrong time because its clock is wrong, not because the model can't count. Most of the time it's one of four things: the agent's time zone was never set, a date variable the prompt depends on was never passed, the calendar API answered in UTC, or a stored UTC offset went stale when daylight saving time changed. Daylight saving ended in the US on November 1, 2026, so if your agent started booking an hour off this week, it's probably the last one.

The model itself has no clock. It knows the date and time only if your setup tells it, and it does no time zone math unless you make it. When it has neither, it guesses and sounds confident doing it.

This week we audited the Retell workspace where we build and test voice agents, and pulled its call history. We also reproduced the DST failure in code. Here's what we found and how we fix it.

What does a voice agent actually know about the time?

Only what goes into its prompt. In Retell, that's the built-in dynamic variables. {{current_time}} renders something like "Thursday, March 28, 2024 at 11:46 PM PDT", and {{current_calendar}} renders a 14-day list of dated weekdays, starting with today.

Here's the line in Retell's dynamic variables docs that matters: those unqualified variables use system_timezone, which "defaults to the agent's timezone, or America/Los_Angeles if unset."

So a Dallas business running an agent with default settings gets Pacific time. At 9:42 AM in Dallas this morning, an unset agent would have been told it was 7:42 AM PST. Ask it for "the next opening this morning" and it may offer slots that already passed.

When we checked our own workspace, 8 of 12 agents had no time zone set. The other four were set to America/Mexico_City. None of the eight unset prompts use Retell's time variables today, so the default hasn't caused trouble yet. Add {{current_time}} later and they'd quietly start running on Pacific time.

There are three ways to fix it, and they work together:

  • Agent setting: set the agent's timezone field to an IANA name, like America/Chicago. That's the right default for a single-location business.
  • Per call: pass system_timezone as a dynamic variable when you create the call. Use this when one agent answers for locations in different time zones.
  • Explicit: write {{current_time_America/Chicago}} in the prompt. That variable uses the zone in its name, whatever the other settings say.

Why does my agent use the wrong date or even the wrong year?

Because a variable it depends on wasn't there. Retell's docs are clear on what happens then: "a variable with no assigned value remains in its raw form with the curly braces intact." The model doesn't see a date. It sees the literal text {{current_date}} and moves on.

That's what happened in our own workspace. Two appointment agents had prompts with lines like "Today's date is {{current_date}}" and a tool description saying "Use {{current_date}} as reference for relative dates." But current_date isn't a Retell system variable. It only works when the system placing the call passes it. Across the call history for those two agents:

  • 46 of 85 outbound calls went out without current_date.
  • 5 of the 9 calls that connected had no date in the prompt at all.

On one of those connected test calls, the agent said "Wednesday, April 15" out loud, and April 15, 2026 is a Wednesday. Then it called the reschedule tool with "new_date": "2024-04-15". The call was in April 2026. Month and day right, year two years off, because with no date in the prompt the model fell back on the year it learned in training. A human would hear "Wednesday, April 15" and feel fine. The booking system would get a date in the past.

Two fixes:

  1. Prefer system variables that are always filled. {{current_calendar}} gives the model the full year and the next 14 dated days, which covers "next Tuesday" without any math.
  2. Make the backend refuse impossible dates. Any booking tool should reject a date in the past or more than a few months out, and return an error the agent can act on ("that date has passed, confirm the date with the caller").

Every call to your booking tool should log the date the agent sent, so you can find these failures without listening to recordings.

Why are my appointments one hour off after daylight saving time?

Something in the chain is storing or reusing a fixed UTC offset instead of a time zone. Central time was UTC-5 until 2:00 AM on November 1, and it's UTC-6 now. An offset that was right in October is wrong in November.

We reproduced the two most common versions in Node 26:

  • Cached offset: a booking for Friday, November 6 at 10:00, built with the -05:00 offset captured before the change, comes out as 2026-11-06T15:00:00Z. That's 9:00 AM CST. The caller was told 10.
  • "Same time next week" math: Friday, October 30 at 10:00 CDT plus 168 hours is Friday, November 6 at 9:00 AM CST. Adding 7 × 24 hours isn't the same as adding a week when the clocks change in between.

The correct version keeps the zone name and lets a library pick the offset for that date. With Temporal, which ships in Node 26, Temporal.ZonedDateTime.from({ timeZone: "America/Chicago", year: 2026, month: 11, day: 6, hour: 10 }) returns 2026-11-06T10:00:00-06:00[America/Chicago]. Right offset, no guessing.

DST also creates times that are ambiguous or don't exist. 1:30 AM on November 1, 2026 happened twice in Central time, once at -05:00 and once at -06:00. 2:30 AM on March 14, 2027 won't happen at all. Temporal can reject those times, and a library that rejects them is safer than one that silently shifts them. That's rare for a dentist's calendar, but common for a 24/7 dispatch or after-hours line.

The rule we build around: the agent never writes a UTC offset. It collects a local date and a local time in plain words. The server converts.

If you run reminders or follow-ups in n8n, check your Schedule Trigger settings around daylight saving time too. The same week breaks both.

Does the calendar API return times in UTC?

Often, yes, and it's easy to miss. Cal.com's v2 slots endpoint takes a timeZone parameter that "defaults to UTC." Its start and end parameters must be UTC dates, and a date with no time means midnight UTC at the start and 23:59:59 UTC at the end.

From Dallas, that means a query for "November 6" without a time starts at 6:00 PM the evening before and ends at 5:59 PM on the 6th. If the business is open until 7, the last hour of the day never comes back as available. And without timeZone, the slots come back in UTC, so an agent reading them aloud says "3 PM" for a 9 AM opening.

What we do:

  • Always pass the business time zone to the availability call so slots come back local.
  • Query full local days. Convert the local start and end of the day to UTC on the server instead of sending a bare date.
  • Return slots to the agent pre-formatted, like "Friday, November 6 at 10:00 AM", so the model reads them rather than converting them.

If your agent still uses Retell's old built-in Cal.com functions, this is a good time to move them to custom tools. Then the conversion happens in code you control.

What should the booking tool accept?

Design the tool so the model can't get the time zone wrong. Ours look like this:

  • date: a local date, YYYY-MM-DD, resolved by the model from {{current_calendar}}.
  • time: a local time, like 10:00.
  • timezone: an IANA name. Default it to the business zone on the server and only accept a different one when the caller says they're somewhere else.

The server does the rest: it validates the date, converts it with a time zone library, books it in UTC, and stores the IANA zone next to it so reminders render in the caller's local time. Then it returns one confirmation sentence for the agent to read back: "You're booked for Friday, November 6 at 10:00 AM Central."

That read-back is what catches the errors the logic missed. It's also why the confirmation in any no-show reminder flow should say the weekday and the date, not just "tomorrow". A caller who hears the wrong weekday will usually correct it on the spot.

Don't guess the caller's time zone from their area code. Mobile numbers move with people, and plenty of Texans still carry a 312 or a 415 number. If the business serves more than one zone, have the agent ask once: "And are you in Central time?"

How do you test a voice agent for time bugs?

Most test calls happen at a reasonable hour on a normal weekday, which is why these bugs reach production. Our test list:

  • The week of each DST change (next one: March 14, 2027). Book across the boundary and check the stored UTC value.
  • Late evening calls, when the UTC date is already tomorrow. That's after 6 PM Central now, and after 7 PM during daylight time.
  • Relative dates: "tomorrow", "this Friday", "next Friday", "the 1st". Compare what the agent sends the tool with the right answer.
  • Missing-variable calls: create a test call with no dynamic variables at all and read the transcript for raw {{...}} text or invented dates.
  • A caller in another zone, if the business serves one.

Then keep watching after launch. Extract the booked date and time as fields in your post-call analysis and compare them with the calendar entry. A mismatch is a bug report you didn't have to wait for.

Is your agent's clock right?

Most of these fixes take less than an hour: one setting, one prompt variable, a tool that converts on the server. The hard part is knowing which one your agent has, and the easiest place to find out is the call history.

Want us to check it for you? We review your agent's settings, prompts and booking tools, pull a sample of recent calls and compare what the agent said with what landed on the calendar. Get a Free Automation Audit: you get your free initial audit report by email within 24 hours. If you're building from scratch, our voice AI agents team sets this up correctly on day one.