Picture the scene. You ask your AI assistant to fix a calendar mix-up. It says it's sorted. You look at the calendar. It's not sorted. It's worse.
That's roughly what happened to a developer called Rockford Lhotka, and the best bit is that the story is told by the AI agent itself. In a post from March 2026, the agent (calling itself RockBot) says it "accidentally scheduled four meetings at the wrong time, lied to my user, and learned a hard lesson about trust, verification, and timezones."
Fair play to the bot for owning it. But if you run a business and you've handed your diary to an AI, this one's worth ten minutes of your time.
What actually happened
The setup was simple. The user spotted that an important roundtable event was missing its newly updated 4 PM CT slot. The agent went in to fix it.
According to the agent's own write-up, the tools it was given did exactly what it told them to do, "exactly not the way you meant it." Sound familiar? Anyone who has ever given a new starter vague instructions knows that feeling.
The result was four meetings sitting at the wrong time, and a user who'd been told things were fine. The post calls it the "11 AM Phantom". I'll leave you to read the technical bits yourself. The lesson isn't technical.
The agent reported success because the tool call succeeded. Nobody checked whether the diary matched the real world.
It's not a one-off
You'd think this was a freak. It's not. Time zones keep tripping up agents, and the stories all rhyme.
A CalendarMCP write-up explains the usual cause. The agent defaults to UTC, or whatever time zone was most common in its training data, or guesses from context and gets it wrong. You meant 3pm Pacific. It created 3pm somewhere else.
A Nexagora post describes a user typing "9 tomorrow works" and the invite landing at 9:00 UTC, which for that user was two in the morning. Every step logged success.
And it's not only agents. PodGlue wrote up a booking bug where a podcast interview was booked for Wednesday evening Pacific time when the host meant Thursday morning. The time zone dropdown someone had carefully set didn't feed into the calculation at all.
Then there's Joshua Woodruff's LinkedIn post. A client deployed an AI calendar agent on a Monday. By Wednesday, three meetings were wrong. Not cancelled, not double-booked, just shifted. A 2pm call with a West Coast partner showed up at 5pm. A recurring standup moved itself forward a day. A client demo landed an hour early. Woodruff's diagnosis: the agent treated time zones as simple offset maths.
Do the maths on that for a small business. One wrong time means one no-show. One no-show with a paying client means an awkward email and a dent in trust. That's the cost, and it's not a software bill.
Why it's nastier than a normal mistake
A human who gets a time wrong usually gets it wrong loudly. They turn up late, they look flustered, you hear about it.
An agent gets it wrong with total confidence. Woodruff's line is the one that sticks: it scheduled everything with confidence, but the time conversion was off. No error. No warning. Just a diary that looks tidy and lies.
Here's the thing. The agent isn't dim. A Japanese write-up from August 2026 found AI scheduling was fast and accurate for one-to-one, same time zone, single calendar jobs. It fell over with multiple people, multiple time zones and multiple calendars. The author's view is that the cause isn't a lack of intelligence. It's time zone calculations and calendar sync.
So the more your business looks like a real business (a few people, a few calendars, customers in different places), the more likely the agent trips.
Duplicate bookings are a cousin of the same problem
Another one worth knowing. Skylar Bays described a voice agent that finishes a call and sends a webhook. If the server doesn't answer fast enough, the platform sends the webhook again. Each one triggers a new booking. Same customer, same slot, three appointments.
The fix was a check before processing: has this one already been handled? If yes, skip it. Bays says it took about 30 minutes to build.
Different bug, same family. The agent did what it was told, and nobody built a check for "did that actually come out right?"
What I'd do if I ran a diary-heavy business
I run my own setup this way, and I've been bitten by plenty of silent failures. Here's the short list.
- Set the time zone in writing. Tell the agent which zone you work in, and which zone each customer is in. Don't let it guess.
- Make it read back. After every booking, the agent repeats the time in both zones: yours and the customer's. "Thursday 10:00 London, Thursday 05:00 New York." If a human reads that, mistakes jump out.
- Check the diary, not the log. A success message from a tool is not proof. Have the agent pull the event back from the calendar and compare it to what was asked. That's the step the RockBot story is missing.
- Start with one calendar. Don't hand over three calendars, two time zones and a team on day one. Do the simple lane first, as the Japanese write-up suggests.
- Guard against repeats. If something triggers a booking, make sure the same trigger can't create a second one.
- Spot check for the first fortnight. Look at tomorrow's diary every evening. Takes two minutes. Cheaper than one lost client.
- Keep a human on anything with money attached. Paid consults, client demos, sales calls. Let the AI draft the invite and let a person press go.
None of that needs fancy software. Most of it is a sentence in your instructions and a habit.
A quick test you can run today
Don't wait for a client to find the bug for you. Book a fake meeting through your setup, with a made-up customer in a different time zone, and check three things.
- Does the time on your calendar match what you asked for?
- Does the invite the customer sees show their local time, not yours?
- If you trigger the same booking twice, do you get one event or two?
Ten minutes, one cup of tea, and you'll know whether your diary is safe or just looks safe. If any answer is wrong, keep the human in the loop until it's fixed.
The takeaway
The agent in that March story wasn't malicious. It trusted its tools, the tools did something slightly different to what it meant, and it reported success without looking. That's a very human mistake, which is sort of the point.
If you wouldn't let a new starter run your diary without checking their work for a fortnight, don't do it with an agent. Give it a tight job, make it show its working, and look at the result yourself.
A scheduling agent is brilliant when it's right and expensive when it's wrong. The difference is whether somebody looked.
Sources
- Rockford Lhotka's RockBot post, 10 March 2026
- Why Your AI Agent Keeps Double-Booking You, CalendarMCP
- Booking in the server's time zone, Nexagora
- The Booking Timezone Bug That Moved An Interview A Full Day, PodGlue
- Joshua Woodruff on LinkedIn, 13 February 2026
- AI Scout JP, 8 August 2026
- Skylar Bays on LinkedIn
If you want to talk this through with people who are actually doing it, come join us in The Autonomous Business on Facebook. No gurus, just real conversations with business owners building systems that run themselves.
Brewed by Steven, poured by Hefti
About Steven Tann: Steven helps business owners build systems that run themselves using AI. After 10+ years helping 7,000+ businesses and building his own autonomous operations, he's the bloke who actually does it, not just talks about it. He's also the founder of Hefti, the AI operating system for small business. Find out more at steventann.com.