A short report from your AI is only good news if you know what it checked. "Two clients are waiting on you, nothing else needs attention" makes for a lovely Monday, as long as the AI actually looked at everything else.

Most AI output gives you no way to tell whether it did. The fix is to make it say "nothing found" out loud, every time, for every item it was supposed to check.

Two kinds of nothing

Say you've asked an AI to go through your email every Monday and flag any client who's waiting on a reply from you. You have six active clients. This week it comes back with two names and a sentence about each. Both are right. You answer both and get on with your morning.

Then there are the other four. Maybe they were quiet. Maybe one wrote from a personal address the search didn't match. Maybe the search only went back five days and last Wednesday's email fell outside it. Maybe the AI read "no rush, whenever you get a chance" and decided that didn't count as waiting. Every one of those produces the same report: two names, and nothing about the rest.

An empty result and a skipped check look exactly the same on the page.

In a one-off chat you can live with this, because you're right there and you'd probably ask. A recurring report gets read differently. The whole point of having one is that you can skim it and move on, and skimming is exactly when an unexplained gap gets taken as good news.

Why the empty parts get cut

Nobody tells the model to hide things. It drops the empty parts because they look like filler. A line like "no email from this client this week" reads as clutter in a report that's meant to be useful, and a model writing toward a useful-looking answer cuts clutter first. If you've ever told ChatGPT to keep it short, you've asked for exactly this, and it did what you asked.

The second reason matters more. Unless the job spells out what to check, the AI has no list of its own to work through. It goes through what it finds and reports what stood out. Nothing in that process makes it notice that it never got to the client who emails from a personal Gmail, so a sentence like "I didn't check this one" never comes up.

Give every check its own line

The fix is boring, and it holds up. Write down what gets checked, as a list, and require a line for every item on every run, including the ones where nothing happened. Then add the line people forget: anything it tried and couldn't reach. That one goes first.

The Monday report turns into something like this:

- Couldn't check Fenwick. The search returned an error. - Acme: you replied Tuesday. Nothing waiting. - Birch: waiting on you since Thursday, a question about the second invoice. - Calder: no email this week. - Dunmore: waiting since Friday, wants to move the review call. - Ellis: last message was yours. Ball's in their court.

That's six lines where there used to be two, and it's faster to read. The clients come in the same order every week, so your eye goes straight to the two that need you and slides past the rest. You also aren't left with a low hum of doubt about four names nobody mentioned.

The line that says "couldn't check" is the most useful line in the report.

Without it, a broken search and a quiet client leave the same blank. With it, the failure shows up as a plain fact you can act on: open Fenwick's thread yourself this week, or sort out the search before next Monday. It sits above the good news because it's the only line that changes how much you should trust everything below it.

A missing line means something now

Once every check has a slot, the report has a shape, and the shape repeats. That changes what silence means. If Calder's line is simply gone one Monday, you're no longer looking at a quiet week. Something broke, and you can see it in two seconds because there's a gap where a line always sits.

This trade is worth making. An agent that reports its empty checks looks less impressive than one that only reports what it found, because most of what it says is "nothing here." But you can believe the quiet parts without checking them again yourself, and that checking was the time you were trying to save in the first place.

It does depend on the list staying put. If you paste a prompt into a fresh chat each Monday, the wording drifts, and sooner or later a client falls off the list without anyone noticing. A recurring job written once, run by an agent you own on your own machine, keeps its shape because the instructions don't change between runs. When a seventh client signs on, you add one line to the list, and the next report has seven slots.