Playbook25 September 2026

How to tell if a decision you repeat at work is worth handing to cheap AI judgment

Four questions tell you whether a repeated sorting decision, like leads, tickets or email, is a fit for a cheap automated check, and when the call should stay with you.

Every morning you go through the same pile. New leads, support tickets, emails that piled up overnight. For each one you ask roughly the same small question: is this urgent, is this relevant, does this deserve attention. The answer takes a second, but when there are hundreds of items, those seconds are the biggest part of your day.

Here's a four question test for whether that decision is a fit for a cheap automated check, a prompt that turns a vague call into clear questions, and when the decision should stay with you anyway.

Start with the basics

There's a category of AI tool now whose only job is answering one small question, over and over, fast and cheap: which category this belongs to, how urgent it is, whether a statement is true. It doesn't write anything or make a complex call, it just judges. TypeSafe, which built one such tool called Jev, is reportedly raising money at a valuation over $10 billion, up from a $200 million seed. The reason isn't that it's smarter than a regular language model, it's that it's 40 to 400 times cheaper and far faster, which makes it something you can run on every single item instead of a sample.

Four questions, before you build anything

Answer plainly. You'll know the answer to most of these as you read them.

  1. Do you have a pile, not a single item?Hundreds of tickets, leads, emails, comments. A one off case is faster to just handle yourself.
  2. Can you write out every possible answer in advance?A category from a closed list, a yes or no, or a scale you can describe in words. If you can't say what a good answer looks like before you start, the task isn't ready yet.
  3. Is a wrong answer cheap to catch?Something you'd spot on the next pass and fix easily. If a mistake touches money, hiring, or security, don't leave the call to automated judgment alone.
  4. Can you hand over the evidence as plain text?An email, a transcript, a product description. If checking it needs a full account or a complex image, this is a weaker fit.
Hand it overAll four are yes, and it doesn't touch money, hiring, or security. Run it on the whole pile, not a sample.
Use it to shortlist onlyAll four are yes, but it touches money, hiring, or security. Let it rank or flag everything, then send the borderline cases to a full human review or a slower model.
Break it apart firstYou've got one vague question like 'is this a good lead'. That's several small questions stacked together, write each one out separately before you start.
Keep it with youThere's no real pile, just a one off case, or you can't say in advance what a good answer looks like.

Wait, before you trust the score

A ranking has no reasoning attached to it, just a number. That's fine for sorting a queue, it's not fine for deciding who gets hired, whose payment goes through, or what counts as a security risk on its own. The safer pattern is hybrid: let the cheap check handle volume and rank everything, then route anything borderline to a fuller review, a person or a slower model, before it becomes a real decision.

One vague question is usually five small ones

'Is this a good lead' isn't one judgment, it's industry fit, company size, buying intent and a few more, all folded into a single call. Write each one out on its own, describe every level of any scale in words instead of leaving it to guesswork, and ask more questions than feels necessary. Before you let a check like this run your inbox or your pipeline, label 50 items yourself first and compare.

Prompt to copyI have a decision I make over and over about [describe the pile, e.g. new leads, support tickets, incoming email]. Instead of one vague question, help me break it into three to five small, clear questions, each answerable as yes or no, a choice from a closed list, or a score on a scale described in words. For each question, give me an example of a good answer and a bad one, so I know if I've worded it right.

This wording matters because it forces the vague call into pieces you can actually test, one at a time, instead of hoping the AI reads your mind on the first try.

What happened to people who tried this

Jonathan Yunikowski asked what his inbox would look like ranked by importance instead of by time. A cheap judgment check rated 100 emails in 453 milliseconds, for about a tenth of a cent, and matched his own ratings on every single one. "This is a use case that I want right now, not tomorrow," he wrote. "In fact, I wanted it yesterday." Stephen Tey, who runs the link company Dub, fed a cheap classifier 10,000 previously caught malicious domains to flag bad links going forward, something he said his team had been wrestling with since day one. "With Jev, we solved it in two hours," he wrote. And Ian Nuttall ran eight questions, topic, hook, tone, over nearly 3,300 of his own old posts for about 13 cents, then compared the results against what actually got engagement.

Do this next

Pick one decision you make more than a few times a week, sorting leads, flagging emails, tagging tickets. Run it through the four questions above. If it passes and doesn't touch money, hiring, or security, use the prompt to break it into small questions, then test it against 20 or 30 items you judge yourself before you let it run on anything larger.

Sources
All guides