My AI doesn't know what time it is

On a show day my AI assistant told me we had 25 minutes to air. We had about 85. Every clock in the building was right. The fix was not a better rule. It was a fact on the screen.

Share
A laptop on a desk at sunset with a wall clock in the background

The 60-minute mistake

It was a show day. The countdown matters on show days: promo posts go out on a schedule, sound check is at a fixed time, and the stream starts when it starts.

My AI assistant told me we had 25 minutes to air. We had about 85.

My first guess was a server with the wrong time zone. I checked. The workstation, the Mac running the stream, and the web server all agreed to within two seconds. Nothing was wrong with any machine.

The error was in how the AI was keeping time. It had checked the clock a few times early in the session. After that, it was estimating "now" by adding up how much work it felt it had done since. Work does not map to minutes. The estimate drifted across a long session until it was an hour off, and a wrong countdown does real damage: it rushes you through steps in the wrong order and makes you put off work you had plenty of time for.

Why a rule didn't fix it

The first fix was a written rule: before you say what time it is, check the clock, in that same turn. It helped, and it still failed sometimes, because a rule only works if the AI remembers to follow it every single time. That is the weakest kind of fix there is.

An AI assistant's sense of the date is also thinner than you'd think. It gets told the date once, when the session starts. No time, no day of the week. Work past midnight and even the date is stale.

The fix: put the clock in front of it

Now a small script runs on every message I send and puts one line in front of the AI before it reads anything else: the local date, the time, the day of the week, whether daylight saving is on, and the week number. The AI does not have to remember to check. The answer is simply there.

The same script became the one time source for my automation. Before, about twenty scripts each worked out the time their own way, some in the local zone and some in UTC. Now 31 scripts read the same clock.

What testing found

  • I checked the new clock against the old method at 53 moments across the year, including both daylight saving changes and the repeated hour in November. Zero mismatches.
  • Moving it to the second machine turned up two bugs. One was a hard-coded folder path, and the other was a command that does not exist on macOS. Both failed silently: the script fell back to the old method and still printed a plausible time. If I had not been testing for that exact case, I would never have known.
  • It took about 35 minutes to build, against a two-hour budget.

What this means for your business

If you use AI for anything where time matters, like scheduling, deadlines, dated documents, or "what's due this week", do not assume it knows what time it is. Give it the time, every time, from one source you trust.

There is a bigger version of this lesson too. When an AI keeps making the same mistake, the tempting fix is a stricter instruction. The fix that lasts is to change what the AI can see, so the right answer is already in front of it. If you want help finding those fixes in your own workflow, get in touch.