field note
Chat or Decide: Which Kind of AI Your Office Actually Needs
By Shafen Khan//Updated /5 min read
short answer
Most small businesses need AI that decides, not AI that chats. A chatbot writes you a paragraph. Most of the work inside a small business is a small call made dozens of times a week: is this junk or real, does it go straight through or to a person first, does it wait until Tuesday. A decision model picks one answer from a fixed list and says how confident it is, but it cannot tell you why.
Bring the calls your team makes all week. Get a straight answer on which to hand to software.
tl;dr
- A chatbot writes you a paragraph. Most of the work inside a small business is a small decision made dozens of times a week: junk or real, straight through or to a person first, now or Tuesday.
- TypeSafe released Jev in early access on September 15, 2026. It only decides: one answer from a fixed list plus a confidence number, fast and cheap. It cannot tell you why.
- Do not buy anything yet. Count for one week how often your team sorts, routes, flags or prioritizes something. That list tells you whether any of this is for you.
A new inquiry lands in the inbox overnight. Somebody has to decide whether it is a real customer or junk. An order comes in through the website. Somebody has to decide whether it goes straight through or needs a person to look first. A maintenance request arrives at eleven at night. Somebody has to decide whether it waits until Tuesday.
None of those is a conversation. All of them happen dozens of times a week, and in a small business they are made by a person who is also doing four other things. For three years the industry has been selling small offices a way to have a conversation.
The shape of the work
A language model generates text. That is genuinely valuable when a paragraph is what you want: a first draft of an email, a summary of a long document.
It is the wrong shape when you need a thing to end up in the right place. Here is the tell. If anyone has built you a system that asks an AI a question and then digs the answer back out of the sentence it wrote, that digging is where it quietly breaks. The model phrases its answer slightly differently one Tuesday, and something lands in the wrong pile without anyone being told.
What changed on September 15
A company called TypeSafe released a model called Jev, in early access behind a waitlist. Jev does not write. You give it a situation and a fixed list of possible answers, and it hands back one of them plus a number for how confident it is.
The plainest way to put it: the difference between asking a colleague to write you a memo about which pile a document belongs in, and having a colleague who simply puts it in the pile.
What it is genuinely good at
- It is fast. Answers come back in 70 to 500 milliseconds, so the decision can happen inside a process while somebody waits, not on a schedule overnight.
- It is cheap enough to stop counting. For a small repeated sorting call, the vendor's published price works out to fractions of a cent each.
- It cannot hand back something unreadable. The answers are fixed before the question is asked, so it cannot invent a fifth category when you gave it four.
What it is bad at, which no vendor page will lead with
- It cannot tell you why. A confidence number is not a reason. If a decision might one day have to be explained to a regulator, an auditor, an insurer, or an attorney, this is the wrong tool. That rules it out of a great deal of medical, legal, and safety work immediately.
- It can still be wrong. It cannot invent an option outside the list you gave it. It can choose the wrong option from inside that list.
- The numbers are not settled. Accuracy figures quoted by different write-ups differ by roughly eight points on the vendor's own benchmark, and TypeSafe has said the pricing may be subsidized.
Where this actually fits a small business
Not at the front door, where everyone keeps putting AI. In the middle, where the work already is: separating overnight inquiries from spam, routing standard orders straight through and special requests to a person, sorting maintenance requests by urgency before anyone gets dispatched.
The design that matters is the threshold. Anything the model is sure about flows through. Anything it is uncertain about goes to a person, who now spends the morning on the twelve genuinely ambiguous cases instead of the four hundred obvious ones.
Where to start
Not with Jev. It was four days old when this was written, behind a waitlist, with a price the vendor says may not hold. What has changed is the question for the next sales call: am I being sold a conversation or a decision, and which one does my bottleneck need.
Then count. For one week, have everyone write down each time they sort, route, flag, or prioritize something: the small repeated calls an experienced person gets right in about two seconds. If the list is short by Friday, you do not have this problem. If it runs to thirty lines, you have found the part of your business this class of tool is for.
sources
- TypeSafe: Introducing System One Models and Jev the vendor's own launch announcement
- TypeSafe developer documentation API surface, model behavior and limits
- Developers Digest: TypeSafe Jev, System One models release guide independent write-up, pricing and latency figures
Read and checked on September 19, 2026. Vendor claims are marked as such. Where published figures disagreed, that is said in the text rather than resolved silently.
common questions
Questions I get asked on this
Can I use this in my business right now?
Not yet. Jev went into early access on September 15, 2026, behind a waitlist, and it is the first release of its kind. Nobody should rebuild a working process around a model that new, at a price the vendor has said may be subsidized. What is useful today is the distinction it makes obvious, not the product.
What is the difference between a chatbot and a decision model?
A chatbot produces language. You ask it something and it writes back, and then a person reads it or your software has to dig the answer back out. A decision model skips the writing. You hand it a situation and a fixed set of possible answers, and it returns one of them plus a number for how confident it is. A chatbot is for when you want a paragraph. A decision model is for when you want the thing to go in the right pile.
Does it really never make things up?
That claim is worth splitting in half. Because the possible answers are fixed before you ask, the model cannot invent a category that does not exist or hand back something your software cannot read. That part is true and genuinely useful. It can still pick the wrong answer from the list. Only the first claim holds.
Is this going to replace my staff?
It replaces a step, not a person. The steps worth handing over are the small repeated sorting calls an experienced member of staff makes correctly in two seconds and resents making forty times a day. The judgment calls that need a human are the ones this tool is worst at, because it cannot explain its reasoning. Uncertain cases go to a person, which is why the confidence number matters more than the speed.
pass it on
Send it to a coworker, your boss or a friend who has this problem.
Handing it to someone who prefers paper? Get the one-pager to pass on.
Run the count.
Spend a week writing down the small repeated calls your team makes, then bring the list. You will hear which of them are worth handing to software and which should stay with a person. If none of them are ready, that is what you will hear.