jonas.splinter
Search
Blog

Let an AI sort the inbox, then snooze what can wait

How a small AI model labels each new mail as high or low value, why low-value mail is snoozed to two fixed times a day, and why I let it err towards important.

Not every mail deserves my attention the moment it arrives. A buyer asking for a price does. A newsletter, a receipt or a polite "no thank you" can wait a few hours without anyone noticing. For a while I sorted that by hand, which meant reading everything twice. Now a small AI model reads each new mail in my shared inbox first, labels it as worth my time or not, and the inbox puts the second group to sleep until a set time of day. This post describes how that works in practice, the choices that made it reliable, and the one mistake I made early on. It runs on my real inbox every few minutes, so everything below comes from daily use.

How the sorting works

Every new conversation in my FreeScout help desk gets one question: is this mail high value or low value. High means a person who might buy, a client who needs an answer, or something I asked for myself. Low means everything that can wait.

The model that answers is Clef, a decision model rather than a writing model. It does not produce a paragraph of reasoning. It picks one of the options I give it and returns how likely each option is. That makes the answer easy to act on and easy to check: the same mail gives the same answer every time.

The answer becomes a tag on the conversation, such as conv-high or conv-low. The tag is the durable record of the decision. Everything else hangs off it.

I started with three values: high, medium and low. Medium turned out to be a place where hard cases went to hide, and I could not say what I wanted done with it. I folded it into low. Two values, each with a clear action, work better than three where one means "not sure".

Snooze, do not sort into folders

The obvious next step would be to move low-value mail into a separate folder. I chose not to. A folder is where mail goes to be forgotten.

Instead, each tag can carry a snooze schedule. Low-value mail disappears from the active list and comes back at 13:00 or 21:00, whichever is next. High-value mail is never snoozed and stays in front of me. The result is an inbox that, most of the day, only holds conversations that need a reply now, and two short moments where I clear the rest in one go.

A few details turned out to matter:

  • Times are sorted when saved: the next wake-up is the first time still ahead, so an unsorted list would quietly skip the sooner slot.
  • Turning a schedule off keeps its times: disabling and erasing are different decisions.
  • The wake-up time is worked out when the snooze is applied, not when the preview was shown, so a mail never gets snoozed into the past.
  • If the tag fails to save, nothing is snoozed: hiding a mail whose decision never landed would lose it.

When in doubt, call it important

The two kinds of mistake do not cost the same. A low-value mail marked high costs me a few seconds of reading. A buyer marked low waits until 13:00 and may have written to someone else by then.

So I do not simply take the model's top answer. If the chance of high is at least 30 percent, the mail counts as high, even when low is the more likely answer. That lets more ordinary mail through to the active list, and that is the point. I chose the direction of the error on purpose.

One question cannot hold two truths

The early mistake was a login code. A mail with a login code is an automated notice, which I had described as low. It is also something I asked for a minute ago, which I had described as high. With both descriptions in one question, the model flipped between the two from one run to the next.

The fix was to stop asking one question to hold both ideas. The model now also answers a separate yes or no question in the same call: did I ask for this mail myself. When that answer is a clear yes, the code overrides the main choice and the mail is high. The separate question gave a steady answer where the combined one never did. When I need a rule like that, I write it as its own small question and let code combine the answers.

Read before you let it act

Automatic sorting has two switches, and both started off. The first lets the model classify new mail and log its answer. The second lets those answers actually tag and snooze real conversations.

I ran with only the first switch on and read the log before turning on the second. That order matters: a wrong label in a log costs nothing, while a wrong snooze hides a real person. The log has stayed on since, so every decision can still be traced back to the answer and the probabilities behind it.

Two smaller rules keep the automation honest. The list of mail still to sort is defined by the missing tag, so a mail whose tag failed to save comes round again on the next run. And with no tags configured, the sorter does nothing at all, rather than trying to classify the whole help desk.

What I do not know yet

Most of my mail is in English. I have not measured how well the sorting works for Dutch or German mail, and I would not trust it there without a test. I also do not know how stable the model's answers stay when its provider updates it, because the model name carries no version. I re-run my comparison on older mail when I hear of a change, but a quiet update would not announce itself.

Frequently asked questions

Does the model ever reply to customers itself?

No. For this task it only chooses a label. Replies are written and sent by me, and anything that would send mail on its own sits behind separate, stricter checks.

What happens to a low-value mail if I am away?

It keeps coming back at the next slot, in the active list, until someone handles it. Snoozing delays a mail; it never closes or deletes it.

Why not let the model pick the snooze time as well?

Because the time is a policy, not a judgement about one mail. I want low-value mail to come back at the same moments every day, whichever mail it is. Keeping the schedule on the tag makes that rule visible in one place.

What to do first

  1. Write down the two or three kinds of mail that truly need a fast answer, and call everything else low.
  2. Give each label one clear action, and drop any label you cannot give an action to.
  3. Decide which mistake is more expensive and set your threshold to lean towards the cheaper one.
  4. Turn on classification with logging only, and read the log for a few days before letting it tag or snooze.
  5. Pick two fixed times a day for low-value mail to return, and clear it in one sitting.

The mail this inbox handles is mostly about Blockspot.io (opens in a new tab). For the safety side of the same inbox, read A warning bar in my inbox.

Related updates

Short notes on the same topics