The fear most owners have about an AI answering their phone isn’t that it’ll sound robotic. It’s that it’ll say something obviously wrong and make the business look stupid. That fear is reasonable. Here’s a real example of exactly that happening, what caused it, and how it was fixed.
What happened
A catering business had an AI receptionist handling inbound calls, including questions about same-day orders. The cutoff was mid-afternoon: orders placed before that time could go out the same day, anything after would be next day.
A caller rang at 11:20am. Plenty of time. The AI told them same-day orders had already closed “as it’s past 2pm.”
The caller almost certainly assumed the business had its act together. They probably went elsewhere.
The cause
The AI’s internal clock was reading UTC, Coordinated Universal Time, the global baseline that runs zero to twenty-four hours with no adjustments for where you actually are. The business is in the UK. In winter, UK local time and UTC are the same. But at the time of this call, the clocks had moved forward, so UK local time was UTC plus one hour.
The model saw a timestamp that read roughly 10:20 UTC. It converted that to local time itself, got the arithmetic wrong, and concluded it was past the cutoff. It wasn’t. It was 11:20am. The order window was open. The AI closed it anyway.
The root problem: the system was leaving timezone maths to the AI model, rather than computing the correct local time first and handing it over as a plain fact.
Three ways to respond to this kind of risk
Option 1: Don’t use AI voice at all
Pull the AI receptionist and go back to a human answering the phone, a voicemail, or a missed-call text service. No AI means no AI errors. The cost is real though: someone has to be available, and when they’re not, calls go unanswered. If you’re already losing calls because you can’t always pick up, removing the AI doesn’t solve the underlying problem, it just trades one failure mode for another.
Option 2: Accept that occasional errors are the cost of automation
Every system has failure modes. A human receptionist mishears things, gives outdated information, forgets to check the calendar. You could decide that a rare wrong answer is an acceptable trade-off for the calls that get handled correctly at 7am on a Saturday when nobody’s in. That’s a legitimate position, and for some businesses it’s the right one. The problem is that “occasional” is hard to measure if you’re not monitoring calls, and a wrong answer on a time-sensitive question, like an order cutoff, doesn’t just inconvenience a caller. It loses a sale that the caller thought they were entitled to place.
Option 3: Find the specific cause and fix it
This is the option we took. The fix is straightforward once you know what’s wrong.
Instead of letting the AI model figure out what time it is and whether a cutoff has passed, the system now computes the current local time explicitly, before the model does any reasoning, and injects it as a plain fact into the context the model reads. The model never has to do timezone arithmetic. It just sees: “Current time: 11:20am, Thursday, UK local time.” Then it reasons from that.
In practice, that looks something like this. Before the AI processes any call logic, the system runs:
from datetime import datetime
import pytz
business_tz = pytz.timezone("Europe/London")
now_local = datetime.now(business_tz)
current_time_string = now_local.strftime("%I:%M%p, %A %d %B %Y")
# Produces e.g. "11:20AM, Thursday 14 October 2026"
# This string is injected into the system prompt before any call reasoning
The model never touches a raw UTC timestamp. It gets handed a human-readable local time, computed by code that has no ambiguity about daylight saving or timezone offsets, because the timezone is declared explicitly, not guessed.
The rule of thumb: any rule in your AI that involves a time, a cutoff, an opening hour, or a deadline needs the clock pinned to your actual timezone in code, not assumed by the model.
How to check if your setup has this problem
You don’t need to read the code yourself. Ask whoever built or manages your AI receptionist one question: “When the AI decides whether we’re open or a cutoff has passed, where does it get the current time from, and is that time computed in our local timezone before the model sees it?”
A confident, specific answer is a good sign. A vague one, or silence, is worth investigating.
You can also test it yourself. Ring your own AI line at a time when the answer matters, say, ten minutes before a cutoff or just after you open. If the AI’s answer doesn’t match reality, you have the bug.
The recommendation
Option 3. Fix the specific cause. The timezone bug is not a reason to distrust AI receptionists in general. It’s a reason to make sure the one you’re running has its clock set correctly. The fix is a small piece of work with a clear, testable outcome: call before the cutoff, get the right answer. Call after it, get the right answer. Every time, regardless of daylight saving.
The businesses that get the most out of AI answering are the ones that treat it like any other system: commission it properly, test the edge cases, and fix what’s broken. Not the ones that either avoid it entirely or cross their fingers.
Want this checked on your line?
If your AI receptionist handles anything time-sensitive, opening hours, same-day cutoffs, booking windows, it’s worth knowing whether the clock is pinned correctly. We can look at how your system handles time and tell you whether there’s a problem before a caller finds it first. Find out how we work, and start a conversation there.




