AI Used to Write. Now It Decides.

In plain English

The AI most of us use writes things. A new kind of AI makes small decisions instead. You give it some text and ask "which one", "how good" or "yes or no", and it answers with a number that says how sure it is. The one I use is called Jev, made by a company called TypeSafe AI. The idea behind it is the reason I think now is the right time to start a One Person Company. The work that used to need a second person was never the writing. It was the deciding.

我们平时用的 AI，是帮你写东西的。现在有一种新的 AI，是帮你做小决定的。你给它一段文字，问它「选哪个」「几分」「是不是」，它给你答案，还附一个数字，告诉你它有多确定。我在用的这个叫 Jev，是 TypeSafe AI 做的。它背后的想法，就是我觉得现在最适合开始一人公司的原因。以前要多请一个人，从来不是为了写，是为了判断。

TL;DR

Jev answers three kinds of questions: pick one, give a score, yes or no. Every answer comes with a confidence number. That number is the useful part, because it tells you when to trust the machine and when to look yourself. For a one-person business, that covers the small judgment calls that used to need staff. I use it for sorting, scoring, checking and searching. The price is not published yet.

Jev 只答三种问题：选一个、打分数、是或不是。每个答案都附一个「有多确定」的数字。这个数字才是重点，它告诉你什么时候可以放手，什么时候要自己看。对一人公司来说，以前要请人做的那些小判断，现在有东西可以接了。我用它来分类、打分、查对错、找资料。价钱还没公开。

The day I read their docs

On 22 September I opened the documentation for a company called TypeSafe AI. I expected another chatbot.

It was not one. Their model, Jev, does not write a single paragraph. Their own description calls it a model "built to make fast, structured decisions that software can use directly."

Every AI tool I use every day writes something for me. This one gives a verdict and stops.

The AI we already use

"Thanks for your message! Based on what you've shared, it sounds like you may be interested in our services. Here are a few things to consider..."

A model that decides

Wants a quote? yes · 0.91 Urgency (1 to 5) 4 · conf 0.78 Which service? training · 0.83

Illustration only. The numbers on the right are made up to show the shape of an answer, not results from a real run.

What Jev actually does

You send it a piece of text. That could be an email, a form, a contract, a list of products. Then you ask it questions about that text. There are only three kinds.

Choice

Pick one option from a list you give it.

Score

Rate the text against a rubric you write.

Yes or no

How likely is this statement true, from 0 to 1.

You can ask many questions at once. Their docs say they run in parallel and "adding questions barely changes the response time."

They tested it. One long document, 13 questions in one request, against the same 13 questions sent one by one. Asking together came out 12.2 times cheaper and 10 times faster. That is their number, from their own test.

The number that tells you when to step in

Here is the part I keep thinking about.

Every answer comes with a confidence number. TypeSafe has a pattern for it called confidence routing, and the credit for this idea is theirs. Their example is a banking assistant. Below 0.6 confidence on anything, a human takes over. Checking a balance is fine at 0.6. Approving a money transfer needs more than 0.85.

0.6

0.85

Below 0.6A person looks at it.

0.6 to 0.85Low-risk jobs go ahead.

Above 0.85Even the risky jobs go ahead.

The answer tells you what; confidence tells you whether to act.
TypeSafe AI documentation

I run my own business on a lot of AI, and I have approval gates everywhere. Nothing gets paid, posted or sent without me. What I never had was a number that says when I actually need to look. So I look at everything. That is the part that eats the day.

Why this makes now the time to start

When people ask me if one person can really run a company, they usually worry about the writing. Posts, emails, proposals.

That was solved a while ago. The writing was never the real reason you needed a second person.

The real reason was the hundred small calls a day. Is this enquiry serious or just asking? Which job goes first? Is this draft good enough to send? Does this contract have the clause we always need?

You cannot write a prompt for "use your judgment." But you can write a rubric. You can write a list of options. You can write a yes or no question.

You write the rule. The machine sorts. The number tells you which ones still need you.

That is the shift. For the first time, the small judgment calls have somewhere to go that is not a hire. That is why I would start a One Person Company now, and not wait for next year.

It was never the writing that needed a team. It was the deciding.
Weiss Ang

What I am honest about

The numbers are theirs. Every figure in this post comes from TypeSafe's own tests.
I could not find a price. So I cannot tell you it will save you money.
It reads text only. Their docs say images, audio and video are not supported yet.
It does not write. Keep the tools you already use for writing. This is a different job.

TypeSafe publishes its own examples, and two of them show the idea well:

Catching made-up quotes. It checks whether a quote really appears in the source document. If it doesn't, it marks the quote as made up. Answers it is unsure about go to a person to check.
Finding the right document. In their legal search test, 1,200 checks cost about US$0.06 in total.

How I use it myself, without the details:

Sorting. Putting each item into the right category.
Scoring. Rating work against a rule I wrote once.
Checking. Asking whether something AI wrote is really backed by its source.
Searching. Scoring a big pile of documents, then reading only the top few.

The machine does the sorting. I only look at the ones it is unsure about.

Start this week, no tool needed

You do not need Jev to use the idea.

List the small calls you make every day. Just the ones that repeat. Enquiries, orders, approvals.
Turn one of them into a question with a fixed answer. "Pick one of these four." "Score 1 to 5 using this rule." "Yes or no."
Decide your line. Which answers can go ahead on their own, and which ones you always want to see.

That is the rubric. Once it is written down, any tool can use it, and you stop carrying it in your head.

If you want the bigger picture of running a business alone, start with what a One Person Company actually is. And the whole system I use is free in my community.

FAQ

What is Jev?

Jev is the model sold by TypeSafe AI. It does not write text. You give it some text and a set of questions, and it answers each one with a pick from a list, a score on a rubric, or a yes or no probability, plus a number that says how confident it is.

Have you used Jev yourself?

Yes. I use it for four jobs: sorting things into the right category, scoring work against a rule I wrote, checking whether AI writing is backed by its source, and picking the right few documents out of a big pile. The numbers in this post are from TypeSafe's own tests, not mine.

How much does it cost?

I could not find a price list in their documentation. Their own tests suggest it is cheap to run: in one, 1,200 checks cost about US$0.06 in total.

Can it read my videos, photos or voice notes?

No. Their documentation says Jev accepts text only, and that images, audio and video are not supported yet. A voice note or a video would need to be turned into text first.

Does this replace ChatGPT or Claude?

No. It does a different job. The tools you already use write, plan and explain. A model like Jev makes narrow calls: which one, how good, yes or no. You would use both, each for its own job.

Why does this make now a good time to start a One Person Company?

The part a solo owner could never hand off was the small, repeated judgment calls: which enquiry is serious, which job goes first, which draft is good enough. Those calls can now be sorted by a machine that also says how sure it is, so you only step in where it is unsure. That takes away one of the main reasons one person needed a team.

Sources

TypeSafe AI, Introduction. docs.typesafe.ai/introduction
TypeSafe AI, Parallel questions cookbook (the 12.2x / 10x test). docs.typesafe.ai/cookbooks/parallel_questions
TypeSafe AI, Confidence routing pattern. docs.typesafe.ai/patterns/confidence-routing
TypeSafe AI, Citation check cookbook. docs.typesafe.ai/cookbooks/citation_check
TypeSafe AI, Re-ranking cookbook (the US$0.06 test). docs.typesafe.ai/cookbooks/rerank_typesafe
TypeSafe AI, State (text-only input). docs.typesafe.ai/concepts/state

Build your One Person Company with me.

The whole system and the course are free inside my community. Put your name on the waiting list.

Join the waiting list

← All articles Home →