Your AI Agent Is Reading Spam (And You're Paying For It)

If you use an AI agent to handle your email, here’s what’s quietly happening: every single message goes to a big, expensive model first. Newsletters. Promo blasts. Spam. Even scams. Your agent reads all of it, and you pay for every word.

The fix is simple. Put a bouncer at the door.

That bouncer is Jev, a new kind of AI model from TypeSafe AI.

Here’s what makes Jev different. It can’t write a sentence. It doesn’t chat. It only makes decisions. You give it a short list of choices, and it picks one in a fraction of a second, for almost nothing. Think of it like a mail room clerk who never opens a letter fully, just glances at it and drops it in the right tray.

For an inbox, the trays look like this:

1. Needs a reply
2. Brand deal
3. Spam
4. Scam

Only the emails that actually need a thoughtful answer get passed through to Claude to write. Everything else gets sorted and set aside. Your agent reads less, and your bill drops.

The proof: in one test shown in the video, Jev sorted 1,700 emails for 18 cents.

Because Jev can only pick from your list, it never invents a new category you didn’t ask for. It also tells you how sure it is. So you can set a simple rule: if Jev is less than, say, 60 percent sure, send that email to a real person instead.

Here’s the fine print though. That 1,700 email number comes from one creator’s test, not an independent benchmark. Before you trust it with your inbox, try this:

1. Take 50 emails you’ve already sorted by hand
2. Run Jev on them without showing it your answers
3. Count how often it agrees with you
4. Look closely at the ones it got wrong

That tells you whether it’s ready for your inbox, and where to adjust.

The same bouncer idea works anywhere an AI is reading a pile of stuff: support tickets, DMs, social comments, form submissions. Let the cheap, fast model decide. Let the smart model do the writing. And keep a person in the loop for anything it isn’t sure about.