Garry Jakubiak

Writing · 11 October 2026 · 2 min read

Jev: easing the load on the human in the loop

A first-hand account from Private Twin: an agent that gives its opinion on every proposal before I do, and three decisions out of four taken off my plate over three working days.

The Jev logo on a glossy pink tile surrounded by glass, with the words “3/4 of decisions automated”.

Private Twin keeps track of the context of my work: what was decided, who owns what, what is still open. To keep that context up to date, it reads the transcripts of my meetings, my emails and my activity, and proposes updates: a commitment made in a meeting, a task moving forward, a piece of information about a case. Nothing goes in without my approval. This is what is called a human in the loop: a step where a person validates what the machine proposes.

I soon felt the downside: every proposal is a decision to make, and a busy day produces dozens of them. So I added Jev.

What Jev does

Jev reads each proposal alongside the excerpt from its source, the passage of the transcript or the email, and asks itself a few simple questions:

  • does the source actually say this?
  • does the proposal concern the right person?
  • is the proposal useful?
  • is it an advert, a sales pitch or an automated message?

From its answers comes an opinion, keep or dismiss, with its level of confidence. When it cannot decide, it says so: “needs checking”.

Three outcomes instead of one

What Jev saysWhat happensWho decides
Keep, at least 80% surethe proposal is accepted automaticallyJev
Dismiss, at least 80% sureit is rejected automaticallyJev
Everything else: less sure, or “needs checking”it is left for me to review manuallyme

The results over three working days

Three days analysedDecisions
Proposals requiring a decision55
That Jev would have decided on its own42, three out of four
Of those, where I disagreed with it2
Left to review manually13

Instead of 55 decisions, I would have had only 13 to make.

A balance that suits me

Of the 42 decisions Jev would have taken on its own, I disagreed with it on 2, and they were not the ones that mattered most. Jev is not infallible, and it does not need to be. The 80% threshold is a slider between time saved and errors accepted; over these three days, it sits in the right place.

Keeping a human in the loop does not mean making them reread everything. Three quarters of the decisions gone, for two minor decisions slipping through: I am happy with the result, and it makes me want to keep going.