> ## Content Index > Fetch the complete content index at: https://www.darlison.com/llms.txt > Use this file to discover other available public pages before exploring further. # Your AI Is Overpaid For Most Of What It Does - URL: https://www.darlison.com/your-ai-is-overpaid/ - Published: 2026-10-25T09:37:07.000Z - Updated: 2026-10-25T09:37:07.000Z - Description: Most of the AI steps in an automation are not writing anything. They are judging. I tested a model built only for yes-or-no answers against three general AI models. It was at least 17 times faster and 4 times cheaper. - Author: Byron Darlison - Tags: AI, open source, Entrepreneurship *Jev is a new kind of AI model that only answers yes-or-no and multiple-choice questions. I gave it a task. It typically answered in about a tenth of a second, and the whole job cost US$0.07\. I gave the same task to three general AI models. Claude Haiku 5.5 took about 2 seconds and cost US$0.33\. GPT-6 Astra took about 4 seconds and cost US$16.52\. Claude Fable 5.1 took about 5 seconds and cost US$21.51\. Jev is at least 4 times cheaper and at least 17 times faster. It was slightly less accurate than the others. *I was stunned by what this makes possible.** When [TypeSafe AI](https://typesafe.ai/?ref=darlison.com) released [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev?ref=darlison.com), I started experimenting with it. Jev is an AI model, but a very different kind from the chat assistants most of us use. It does not write anything. You give it some text and a question with fixed answers, either yes or no or one choice from a short list. It returns how likely each answer is. TypeSafe built it for automations that run on their own, not for chat. I wanted to find out what a model that narrow is good for. ![Jev by TypeSafe AI logo](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/jev-by-typesafe-ai-logo.png) The experiments changed how I look at my own automations. Most of the AI steps in them are not writing anything. They are making a judgment. Is this invoice paid? Does this email need a reply? Is this message from a client? Did the person answer the question? Those judgments usually go to the same large model that writes. That works, but it is the wrong tool. It is slower than it needs to be, it costs more than it needs to, and the answer has to be dug out of a paragraph. A model built to answer yes or no can handle many of these checks faster and for less, and hands back a number you can act on. Test it against the model you use now. ## Two kinds of AI work AI does two different jobs inside an automation. The first is judging. You give it some text and a question with a fixed set of answers: yes or no, or one choice from a short list. Is this a complaint? Which client is this about? Has this been answered? The second is writing. You want words back: a reply, a summary, a follow-up question. Large models do both well. Judging does not need a writer, though. A model built for judging reads the text and returns a probability, for example 0.92 that the answer is yes. There is no prose to interpret and no format to drift. You set a bar, and anything over the bar counts as yes. ## Let the judge run, wake the writer Once the two jobs are separate, the design follows. The cheap judge runs on everything, all the time. The expensive writer runs only when the judge flags something worth writing about. This changes what you can afford to automate. When a check costs a small fraction of a cent and comes back in a fraction of a second, you can check every email, every invoice and every line of a transcript instead of a sample. The writer's cost then grows with the number of things worth writing about, not with everything coming in. ![A stream of emails, invoices and transcript lines passes a blue Judge gate; one green check goes to the Writer](https://storage.ghost.io/c/42/33/4233497c-fcb6-4830-b40a-012948df926f/content/images/2026/10/judge-runs-writer-wakes-ai-automation-diagram.jpg) ## A live example: Call Coach The first real experiment came from an episode of How I AI, where Claire Vo talked with John Lindquist about Jev: ["Jev: 8 real use cases"](https://youtu.be/dAIIaepNhQM?ref=darlison.com). One of John's examples was a coach that listens while you give a presentation. I wanted the same thing for interviews. When I interview someone, I face a trade-off. If I track my question list, I stop listening. If I give the person my full attention, I lose my place in the list. So I built Call Coach with Claude. Call Coach listens to both sides of a call and transcribes it on my Mac. After each turn, Jev answers a set of narrow questions. Which question on my list did I just ask? Has the other person answered it? Is there something in that answer worth following up: a number, a turning point, a story? Questions are ticked off as they are answered, and the next question to ask appears on screen. A writing AI wakes up only when Jev flags an answer worth digging into. It then suggests one short follow-up question. The follow-ups turned out to be the most valuable part. In a long interview, the answer worth chasing often goes by in a sentence or two. A number mentioned in passing. A decision described as if it were obvious. A story that stops just before the interesting part. When I am also keeping track of a question list, I can easily miss those moments and move on. Call Coach caught them. When Jev flagged one, a follow-up question appeared on screen, and I could ask it before the conversation moved on. It took me into specific areas I would otherwise have missed. A question list covers what you knew to ask before the call. The follow-ups cover what you only learn during it. I used Call Coach for the first time on 24-Oct-2026, in an 80-minute founder interview. By the end of the call, every question on my list and every follow-up was ticked. ## What the measurement showed I wanted to know whether the narrow model mattered, or whether general models would do as well. So we replayed the same 80-minute interview through Call Coach with four checking models. The transcript, questions and rules stayed the same. We changed only the checking model. The four models were Jev, Claude Haiku 5.5, Claude Fable 5.1 and GPT-6 Astra. Claude built the comparison and ran it. | | Jev | Claude Haiku 5.5 | GPT-6 Astra | Claude Fable 5.1 | | ---------------------------------------- | ------------ | ---------------- | ------------ | ---------------- | | Cost for the 80-minute call | US$0.07 | US$0.33 | US$16.52 | US$21.51 | | Typical response time | 0.12 seconds | 2.1 seconds | 3.7 seconds | 5.4 seconds | | Slowest response | 0.5 seconds | 7.4 seconds | 18.1 seconds | 18.7 seconds | | Answered questions ticked (of 25) | 20 | 22 | 22 | 24 | | Unasked questions wrongly ticked (of 3) | 0 | 0 | 0 | 0 | | Answers flagged for a follow-up question | 33 | 13 | 54 | 17 | Question tracking was close, with Jev slightly behind. Jev ticked 20 answered questions. The other models ticked between 22 and 24, with Fable the highest. None of the four ticked a question nobody had asked. Jev ticked 20 in this replay and 21 in an earlier replay of the same call. More tests would show how much its results vary. Speed was the real gap. Jev typically answered in 0.12 seconds. Haiku took 2.1 seconds, Astra 3.7 seconds and Fable 5.4 seconds. In a live conversation, that is the difference between a question being ticked as it is answered and the tool falling behind the person talking. Jev cost US$0.07 for the call. Haiku cost almost five times as much. Astra and Fable cost more than 200 times as much. For an automation that runs thousands of times a day, that cost gap decides whether you can check everything or only a sample. The limits are real. This is one call, replayed as text. Call Coach's settings were tuned on Jev, which favors Jev. Follow-up flags varied widely. Haiku flagged 13 answers, Fable 17, Jev 33 and Astra 54\. I did not grade how useful they were, so this test does not show which count was best. ## Have AI find the yes/no steps for you You do not have to work out which steps fit. Give this prompt to the AI assistant that runs or can read your automations. If you have none yet, ask it instead to find the yes or no checks your team does by hand each week. It does the review and tells you what switching would save. Review my automations and find the AI steps that a narrow judging model like Jev from TypeSafe AI could take over. Jev answers only yes/no or multiple-choice questions about text, and returns how likely each answer is. Check TypeSafe's current pricing at typesafe.ai. For each AI step: 1. Say what it does and which model it uses now. 2. Decide whether Jev could do it. It can when the step reads text and its answer comes from a fixed set. It cannot when the step must write words, calculate, look something up or take an action. If a writing step only needs to run sometimes, name the yes/no question that could decide when it runs. 3. For each step Jev could do, estimate how often it runs, what it costs and how long it takes today, and what it would cost and take with Jev. Use my usage logs or bills where you can see them, and mark which numbers are estimates. Then give me a short report: the steps worth switching, ranked by savings; the total change in monthly cost and response time; and what to test before switching. Ask me about anything you cannot see instead of guessing. Try the top recommendation on real work first. Compare its answers with your current model's before you switch the rest. The writing model still does the writing. It just stops doing the judging. ## Get Call Coach Call Coach is open source on GitHub at [ByronDarlison/call-coach](https://github.com/ByronDarlison/call-coach?ref=darlison.com). It runs on a Mac and needs a Jev key to tick questions automatically. I have released it as an open-source project. Anyone is welcome to take it, change it and build on it. Please treat it as an experiment, not a finished product. --- *Jev is made by* [*TypeSafe AI*](https://typesafe.ai/?ref=darlison.com)*. The starting idea came from How I AI with Claire Vo,* [*"Jev: 8 real use cases" with John Lindquist*](https://youtu.be/dAIIaepNhQM?ref=darlison.com) *(30-Sep-2026). The presentation coach example starts at 33:36\. Call Coach's design and code are my own work with Claude.* *Related articles:* - [*Robert Fritz’s Digital Decision Making*](https://www.darlison.com/robert-fritzs-digital-decision-making/) covers making decisions as a clear yes or no. - [*How to Discover Your Company’s Values Using AI*](https://www.darlison.com/how-to-discover-your-companys-values-using-ai/) puts AI to work on a real business job, with a prompt you can run.