Not subscribed? Sign up to get it in your inbox every week.
PRESENTED BY OMNISEND.COM
Leave Klaviyo in 5 days
Most ecommerce teams treat an ESP migration like a quarter-long project. With Omnisend, you can move off Klaviyo in five days: lists, flows, and history included, email and SMS in one place when you're done.
Customers get $79 back for every $1 spent, and 150,000+ brands already run on Omnisend.
Once you're over, it plugs into ChatGPT and Claude too — ask for a revenue breakdown or a new segment without opening another dashboard.
Use code THEBOTTLENECK for 30% off your first 3 months.
Decisions got cheap
Most of the AI you buy today is built to talk.
It writes the email. It drafts the brief. It explains the spreadsheet. That is useful. It is also the expensive, slow path for a huge share of the work operators actually need: is this lead a fit, which GL account does this receipt belong in, should this ticket escalate, how strong is this account against our ICP.
Those are not writing jobs. They are decisions. And until last week, the default way to automate them was to rent a frontier chat model, beg it for JSON, parse the answer, and hope it did not invent a fourth option you never offered.
TypeSafe.ai just shipped a different bet.
On September 15, the company launched Jev (https://typesafe.ai/blog/introducing-system-one-models-and-jev), the first public "System One" model: a model that does not generate prose at all. You send it state and typed questions. It returns structured answers with probabilities and a confidence score. Input tokens are $0.042 per million. Output tokens are free. End-to-end latency is in the 70–500ms range. Their headline comparison on System One workflows is roughly two orders of magnitude faster and cheaper than asking frontier LLMs to do the same decision work.
That is programmatic decision-making at a price where you stop rationing intelligence.
What Jev actually is
TypeSafe's founder, Diogo Almeida, helped build the instruction-following research that became ChatGPT. His public argument now is blunt: RLHF optimized models to please humans in conversation, which is why they still need a human in the loop when you try to bury them inside software.
Jev is the opposite research direction. TypeSafe calls the class System One Models, after Kahneman's fast System 1 thinking. The model is named after William Stanley Jevons, the economist behind Jevons paradox: when coal engines got more efficient, coal use went up, not down.
TypeSafe's claim is that the same will happen to machine intelligence. Make the decision unit cheap enough and demand explodes.
The product shape is simple:
Choice: pick one of up to 255 labeled options. You get the winner, a probability for every option, and a confidence.
Score: place the input on an ordered scale you define. You get a (possibly fractional) score, probabilities across levels, and confidence.
Noul (yes/no): the probability that a condition is true. No separate confidence field, because the probability is already the distribution.
You can ask many of these questions in one call. They run in parallel against the same state. You pay for the input once. That is the engineering trick that matters for ops: fan out a dozen classifiers over one record without lighting twelve frontier LLM invoices.
Important constraint, said plainly: Jev does not write. It will not draft your outreach or summarize your board deck. It decides. Your code does the rest.
Why this matters for operators
The operator jobs that eat headcount are rarely "write me a paragraph." They are high-volume, fuzzy judgments that used to need either a person or a brittle rules engine.
Three examples that map cleanly onto Jev's primitives:
Accounting categorization. Receipt text and vendor name in. Choice out: which GL account, which cost center, is this a duplicate. Confidence gates whether finance auto-posts or queues for review.
Lead and ticket routing. CRM row or support thread in. Choice for owner team, Score for urgency, Noul for "is this even a real opportunity." High-confidence paths auto-route. Low-confidence paths hit a human queue.
GTM TAM scoring. This is the one that should make revenue operators sit up. Pull a list of every company in a segment. For each row, ask Jev a small battery of questions: ICP fit, buying-committee clarity, timing signal, competitive displacement likelihood. Rank the whole universe. TypeSafe's pricing math is the point. At $0.042 per million input tokens, scoring a large TAM with short state per company lands in "maybe a dollar" territory for many realistic list sizes, not "open a procurement ticket" territory. Exact cost still depends on how fat each record is, so measure it on a sample. The shape of the cost curve is what changed.
The unlock is not that AI can classify. You could already coerce GPT into a label. The unlock is that classification is now cheap enough, typed enough, and confidence-aware enough to put inside the workflows where the volume lives.
The Monday test
Pick one categorization job your team already does by hand or with rules: expense coding, lead tiering, churn-risk flags, inbound ticket routing. Export 200 real examples. Write the Choice / Score / Noul questions the way you would brief a sharp intern. Run them through Jev. Then look at three numbers only:
How many rows cleared a confidence bar you would trust to auto-act.
What the batch cost.
Where the model was confidently wrong.
If the auto-act share is high, the dollar cost is trivial, and the confident errors are rare and patterned, you just found a workflow that should not have a human in the loop for the easy majority. Wire the threshold. Keep the tail.
If the model is cheap but confused, you learned something useful faster than a six-week "AI classification" project: the question definitions need work, or the job still needs richer state.
—
Decisions were the part of ops AI that never quite paid off at scale, because we kept buying writing machines to do judgment work. Jev is a judgment machine. The products that win next will bury it in the workflows where millions of tiny true/false and ranking calls actually move the P&L.
Till next time,
Chris
