Because you are paying a frontier model to write you a paragraph explaining a decision, then reading the paragraph to pull the decision back out. Here is the pattern that fixes it, and the research behind it.
The Sorting Gate You Forgot to Build
Anybody who has spent time around an Alabama livestock auction knows the whole operation runs on one piece of equipment: a gate at the end of a chute that swings left or right.
Nobody writes an essay about each animal. The gate swings, the animal goes where it belongs, the next one comes through.
Most business owners have built an AI system with no gate. Every single item goes to the most expensive model in the stack, comes back wrapped in three paragraphs of explanation, and then a human reads those three paragraphs to extract the one word that mattered.
Here is the direct answer to the question in the headline. Most of what your business asks AI to do is not writing. It is deciding. And there is now a whole category of model built to decide instead of write, which returns a typed answer plus a calibrated confidence score, runs in well under a second, and costs a rounding error compared to a frontier model. Route your yes-or-no work there, keep the expensive model for the places where the actual words matter, and your bill and your latency both collapse.
Candidly, I was doing this wrong for a long time. I would ask a model to evaluate something, get back a lovely justification, skim it for the verdict, and throw the rest away. I was paying for prose I never read.
Key Takeaways
- Decision models return a typed answer (yes or no, a category, a rating) with a calibrated confidence score, instead of returning prose you have to parse.
- TypeSafe's Jev charges $0.042 per million input tokens with output free, which is cheaper than OpenAI's GPT-5 Nano at $0.05 per million, according to Simon Willison's September 21, 2026 writeup.
- A Texas crash study screened 499,500 police narratives and coded 195,857 with a typed decision model, reaching an F1 of 0.908 against blinded human labels while two frontier models gained at most 0.059.
- RouteLLM, the open-source router from LMSYS and Anyscale, cut cost by more than 85 percent on MT Bench while holding 95 percent of GPT-4's performance.
- The tradeoff is real: a decision model gives you a number and no reasoning, so evaluation and human review become more important, not less.
The Problem: You Are Buying Paragraphs You Never Read
Go look at what your business actually sends to AI in a normal week. Not the demos. The repetitive work.
Is this lead worth a call. Which bucket does this support ticket belong in. Is this invoice line a duplicate of the one above it. Does this draft clear the bar to go out. Is this community comment a real question or spam. Which of these forty applicants gets read first.
None of that is writing. Every bit of it is sorting.
But the tool you handed it to is a writer. So you built a prompt that says "evaluate this lead and explain your reasoning," and the model dutifully produces two hundred tokens of explanation for a decision that has exactly two possible outcomes.
You pay for those two hundred tokens at output rates, which on every major API are five times the input rate. You wait for them to stream. Then you or your automation parses the text to find the word "qualified," and hope the model did not phrase it differently this time.
Here is the thing. That is not an AI problem. That is an architecture problem, and it is one you built.
I want to be honest about why so many of us built it that way. When you are not technical, the explanation feels like the safety net. If the model tells you why, you believe you can catch it when it is wrong.
But the explanation was never a safety net. It was a story the model wrote after the fact, and you had no way to verify it reflected what actually drove the answer. You were buying comfort, at output token prices, several thousand times a month.
The honest version of the safety net is different. It is a confidence number you can measure against outcomes, plus a rule about when a human looks. That is exactly what decision models give you, and it is exactly what a paragraph does not.
The Evidence: The Numbers Are Not Close
This is not a thought experiment. In the last three weeks it turned into a category with benchmarks, papers, and open-source clones.
Start with price. TypeSafe AI came out of stealth on September 15, 2026, founded by Diogo Almeida, a co-author of the InstructGPT paper that led to ChatGPT. Their first model, Jev, takes text in and returns floating point numbers out: yes or no, a choice from a set you define, or a score on a scale you define, each with a confidence value. Simon Willison wrote it up on September 21 and pointed out the thing that should stop you cold. Jev charges only for input, at $0.042 per million tokens. Output is free. That is cheaper than OpenAI's GPT-5 Nano at $0.05 per million, and Nano is already the budget option. TypeSafe reports responses in 70 to 500 milliseconds and speed 40 to 200 times a frontier LLM on the same task.
Credit where it belongs: TypeSafe calls these "System One models." The name "decision models" comes from Willison and Maggie Appleton, and it is the better name.
Now accuracy, on a job that matters. Amir Rafe and Subasish Das published a study on September 21 that ran typed decisions over police crash narratives in Texas. Their paper screened 499,500 narratives and coded 195,857 of them against a 27-question schema. Measured against 2,416 blinded human judgments, the typed decision model hit an F1 of 0.908. They benchmarked two frontier LLMs on the same records: one gained 0.059, and the other was statistically indistinguishable. Adding those calibrated variables to the existing coded fields attributed 10,747 more injury and fatal crashes per year to nine contributing factors that the old fields had missed.
A model that cannot write a sentence matched frontier models on a public-safety classification job, at a fraction of the cost, and surfaced ten thousand crashes a year the paperwork was losing.
The pattern generalizes past this one model. Routing cheap work to cheap models has been measured for years. RouteLLM, the open-source framework from LMSYS and Anyscale, reported cost reductions of more than 85 percent on MT Bench while preserving 95 percent of GPT-4's performance. On the fine-tuning side, the LoRA Land technical report tested 310 fine-tuned small models across 31 tasks and found they beat GPT-4 on roughly 25 of the 31, with the biggest wins on narrow, classification-shaped work.
And the adoption is organic, not a marketing push. A survey paper published September 30 counted 2,170 public Jev projects on GitHub as of September 22. That is one week after launch. On September 25, LMSYS published serving work showing p95 latency for scoring 16 candidates on Qwen3-8B dropping from 54.1 milliseconds to 20.6 milliseconds with multi-item scoring. On September 29, PostHog shipped Jeeves, a 9-billion-parameter open decision model that scores 0.935 on the public JevBench tier against Jev's 0.866, with MIT-licensed weights, training code, and data.
Two weeks. A benchmark, a serving stack, an open clone that beats the original, and a peer-reviewed application in traffic safety. From different people who were not coordinating.
What Actually Changed for Me
The reframe that finally landed was this: stop asking what the AI should say, and start asking what the answer type is.
If the answer to your question is one of a fixed set of things, it is a decision, and you should be making the model choose from that set. If the answer is genuinely new words that a person will read, it is writing, and you pay frontier prices gladly.
Almost everything in a small business is the first kind.
Here is the architecture I now build for clients and teach inside AI Insiders. Three tiers, and every task lands in exactly one.
Tier one is the gate. Cheap, fast, typed. Yes or no, category, rating, with a confidence number attached. This handles triage, tagging, routing, deduplication, first-pass qualification, spam filtering, and ranking. Most of the volume lives here.
Tier two is the escalation. Anything the gate flags as low confidence goes to a mid-tier model or to a human. You set the threshold, and you set it based on what a mistake costs you. A misfiled support ticket is cheap. A misjudged refund is not.
Tier three is the writer. The frontier model, used only where the actual words go out the door under your name. The proposal. The apology email. The sales page. The hard client conversation.
The unlock is tier two, and I want to be direct about why. The confidence score is what makes this safe. A frontier LLM writing prose can be wrong with total fluency and you will never see it coming. A calibrated decision model tells you when it is unsure, and you can audit that number against actual outcomes. The crash study made this concrete: recalibrating on the same labels cut their calibration error by a factor of 3.3.
Now the part most people writing about this are skipping.
Willison flagged it himself and he is right. A decision model is a blacker box than an LLM. You get a number and nothing else. No signals, no reasoning, no thread to pull. He said plainly that he hopes nobody uses Jev to rank job applicants, because bias in a floating point number is almost impossible to pick apart by hand. TypeSafe's own documentation says Jev 1.13 is currently weak with numbers, dates, and adversarial content.
I agree with him completely. Use decision models for high-volume, low-stakes-per-item sorting where you can measure the outcome. Do not use them to decide who gets hired, who gets credit, or who gets removed from your community without a human in the loop. That is not a technical limit. That is a line about what we owe the people on the other side of the decision.
Practical Steps: Build Your Gate This Week
1. Audit one week of your AI usage and label every task "decide" or "write." Pull your actual prompt history or your automation logs. Do not guess. Most people I walk through this find that seventy to ninety percent of their volume is deciding, and they are shocked.
2. Rewrite your three highest-volume decision prompts to return only the answer. Strip out "explain your reasoning." Give the model a closed list of options and tell it to return exactly one. You will cut output tokens to near zero on those tasks before you change models at all.
3. Test a real decision model on your single highest-volume task. Simon Willison released llm-typesafe, a plugin for his LLM command line tool, so you can run a Jev query without writing an integration. If you want open weights on your own hardware, PostHog's Jeeves is on Hugging Face under MIT license.
4. Run one hundred real examples through both the old way and the new way, and compare. Not ten. One hundred. Use records where you already know the right answer. Willison's point applies here: because these models are cheap, running hundreds of test queries costs cents, so there is no excuse for skipping the eval.
5. Set your confidence threshold based on the cost of being wrong. Pick a number, send everything below it to a human, and write that rule down. Review the flagged items weekly for a month and adjust. This is the step people skip and then blame the model.
6. Keep the frontier model for the words that go out under your name. Proposals, sales copy, difficult client emails, anything a customer reads. That is where the expensive model earns its price, and it is worth every penny there.
7. Use this prompt to run the audit in step one. It follows our Perfect Prompt Framework 2.0.
[The Job]
Analyze my AI task list and sort every task into "decision work" or "writing work."
This is for: [YOUR ROLE] at [YOUR BUSINESS TYPE].
It matters because: I am paying frontier model prices for work that only needs a
yes, a category, or a rating, and I want to find out how much of it there is.
[The Background]
Here is what you need to know: [PASTE YOUR LIST OF RECURRING AI TASKS, OR ATTACH
YOUR PROMPT HISTORY]. We run roughly [NUMBER] of these per week. Decision work
means the correct answer is one item from a fixed set I could write down in
advance. Writing work means a human reads new sentences the model produced.
Do not use: tasks I only ran once as an experiment.
[The Deliverable]
Return: a two-column table, every task in one column and DECISION or WRITING in
the other, with a one-line reason. Below the table, list the top five decision
tasks by weekly volume and, for each, the exact closed set of answers it should
be allowed to return.
Must include: an estimate of what percentage of my total AI volume is decision work.
Optimize for: accuracy.
[The Questions]
Ask me any questions you have.
Frequently Asked Questions
What is a decision model, in plain English?
A decision model is an AI that answers a question by picking from choices you supply instead of writing a sentence. You give it a record and a question, and it returns yes or no, one category, or a rating, along with a confidence score. It cannot invent an answer outside your list.
Is this actually cheaper, or is that marketing?
Cheaper by a wide margin. TypeSafe lists Jev at $0.042 per million input tokens with output free, versus $0.05 per million for GPT-5 Nano and far more for frontier models. Because output is where LLM billing concentrates, removing prose removes most of the cost.
Will a small model be less accurate than GPT or Claude on my task?
On narrow sorting tasks, usually no. The LoRA Land study found fine-tuned small models beat GPT-4 on about 25 of 31 tasks. The Texas crash study found a typed model within 0.059 F1 of frontier models. On open-ended reasoning and writing, frontier models still win clearly.
Do I need a developer to set this up?
For a real production gate, yes, eventually. To test the idea this week, no. Start by rewriting your existing prompts to return one word from a fixed list. That change alone cuts output tokens sharply and costs you nothing but ten minutes.
When should I not use a decision model?
When the decision affects a person's livelihood, reputation, or access, and no human reviews it. Simon Willison specifically warned against using Jev to rank job applicants, because a single floating point number can hide bias you will never detect. Keep a human in that loop.
The Gate, and Who Stands Behind It
Back to that auction barn.
The gate is not the impressive part of the operation. The impressive part is that the people running it know exactly which animals go which way, and they built a machine that does only that, thousands of times a day, without commentary.
Your business deserves the same. Not more AI. A gate.
I want to say the harder thing plainly, because I think it is the actual lesson. The reason most of us never built the gate is not that we lacked the technology. It is that we were afraid to make the machine commit. A paragraph of reasoning felt like shared responsibility. A yes or a no is just a yes or a no, and if it is wrong, that is on the system you designed.
That fear is worth respecting and then walking through. Because the alternative is a system that produces confident explanations nobody audits, at prices nobody tracks, for decisions nobody wrote down the rules for.
Build the gate. Set the threshold. Watch the flagged items. Save the expensive model for the sentences that carry your name.
If you want to build this with people who are doing it right now instead of reading about it next quarter, AI Insiders is $247 per month and this is exactly the kind of system we work through together. The door is open if you want it, and I will not chase you if you do not.
Stop paying a poet to do a gatekeeper's job.
P.S. If you take only one thing from this, take this: the next time you write an AI prompt, ask yourself whether a person will ever read the sentences that come back. If the answer is no, you are not writing a prompt. You are building a gate, and you should build it like one.
About the author: Jonathan Mast is the founder of White Beard Strategies, where he teaches entrepreneurs to build AI systems that actually run their businesses instead of impressing them at demos. He serves a 500,000-member entrepreneur community, created the Perfect Prompt Framework 2.0, and speaks regularly on practical AI adoption for non-technical business owners. He writes from Alabama, and he is convinced that the cheapest part of any working system is usually the part doing the most important job.
Sources
- Simon Willison, "Jev introduces a new shape of LLM: System One, aka Decision Models," September 21, 2026
- Amir Rafe and Subasish Das, "Calibrated Decisions at Scale: Converting Police Crash Narratives into Probabilistic Crash Variables with a System One Model (Jev)," arXiv:2609.24052
- "Jev in the Wild: A Data-Driven Analysis of the Jev Model's Functionality, Applications and Ecosystem," arXiv:2609.30216
- LMSYS Org, "Scaling JEV-like Decision Models with SGLang," September 25, 2026
- PostHog, "Jeeves: Reasoning improves Jev-like decision models," GitHub
- Ong et al., "RouteLLM: Learning to Route LLMs with Preference Data," arXiv:2406.18665
- Zhao et al., "LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4, A Technical Report," arXiv:2405.00732
- Simon Willison, llm-typesafe plugin, GitHub
- TypeSafe AI documentation
