Two weeks ago, if you wanted AI that answers with a decision instead of an essay, there was exactly one vendor to buy it from. As of this month, there are three.
The category is decision models: AI that reads whatever you hand it — an email, a support ticket, an invoice — and returns one typed answer with a confidence score, in a fraction of a second, for close to nothing. Not "here's my analysis of this lead." Just: qualified, 94% confident.
Why that matters for a small business: most of the AI work you'd actually automate is deciding, not writing. Sorting, routing, flagging, approving. If a chatbot writes prose around every yes/no question, you're paying essay prices for one-word answers. Decision models are the cheap, instant tier built for exactly that job — and this week the category went from experiment to a real market.
The three vendors
TypeSafe's Jev — the first mover. The original commercial decision model. You define the input and a fixed list of possible answers; it returns one with a confidence score. Two limits worth knowing: it never generates prose (that's the point, not a defect), and the answer list caps at 255 choices. Before that sounds cramped — most business decisions have fewer than ten possible answers. If yours has more than 255, you don't have a decision to automate; you have a catalog lookup.
OpenAI's Decisions API — the zero-new-vendor path. Shipped at DevDay, built on GPT-6 Luna — the smallest of OpenAI's new models, priced at $0.10 per million input tokens on the standard API — it focuses the model on user-defined questions with finite pre-defined answers and returns one with a confidence score in about 150 milliseconds. The significance for a small business: it runs inside the ecosystem you may already pay for. Same account, same bill, no new vendor to evaluate or security-review. One honest note: the API's own pricing hadn't been published at announcement, so confirm costs before building on it.
AWS's Strands Decider 2B — the free, private one. Open-sourced, 1.9 billion parameters, small enough to run on your own machine — a median of about 115 milliseconds per answer on a consumer desktop GPU, in AWS's published numbers — with the full training recipe public. The significance: it's the option where nothing ever leaves your machine, and it costs nothing beyond hardware you already own.
Pick by constraint, not by benchmark
Vendor coverage of this category compares benchmarks no small business will ever run. Pick by your constraint instead:
| Your constraint | Pick | Why |
|---|---|---|
| No engineering team; want to start this week | OpenAI's Decisions API | Runs where you already work — no model files, no infrastructure, no ops |
| Data that shouldn't leave the building (health, legal, financial records) | Strands Decider 2B, run locally | Open weights on your own machine; nothing is sent anywhere |
| Want to test the category before spending a dollar | Strands Decider 2B | Free and open — run it against last month's real data at zero cost |
| Already built on Jev | Stay put | The category is two weeks old; any feature gap costs less than a rebuild |
| Want decisions and chat on one bill | OpenAI's Decisions API | Same account and billing as the chat models you may already use |
What a decision looks like in practice
Four decisions almost every small business makes daily, as decision-model calls:
-
Lead qualification. Input: the lead form submission. Answers:
qualified/needs-more-info/not-a-fit. Not-a-fit gets the polite decline template; qualified pings the owner within a minute. -
Support triage. Input: the ticket. Answers:
billing/scheduling/complaint/urgent. Urgent jumps the queue and pages a human. -
Invoice exceptions. Input: the invoice. Answers:
match/duplicate/exception. Clean ones post automatically; exceptions land in the bookkeeper's folder. -
Review sentiment. Input: the new review. Answers:
happy/neutral/angry, crossed withservice/quality/price/wait-time. Angry-plus-anything triggers a same-day draft reply for approval.
Notice what's missing from all four: paragraphs. The value isn't the words — it's that the sorting happens instantly and only the exceptions reach a human.
Three catches before you buy
Put an escape hatch in every answer list. Add a none-of-these option and route it to a human. Force every input into your categories and the weird stuff gets mislabeled confidently — which is worse than an empty bucket.
Test on last month before you switch. Decision models are cheap, which tempts you to flip the switch on day one. Cheap is only cheap if it's right. Two weeks of parallel answers on real data settles it.
The vendor matters less than your answer lists. Before picking anyone, write down the five decisions you'd automate, the possible answers for each, and where the exceptions go. That document is 90% of the project, it's vendor-neutral, and it ports to all three vendors — so when the fourth and fifth ship (they will), you move without rebuilding.
Bottom line
The "just decide" tier went from one vendor to three in two weeks. That's the signal the category is permanent — and your cue to decide which constraint you're solving first: privacy, budget, or no engineering team. Pick the constraint and the vendor picks itself.
Tonight: list the five decisions your business repeats most, with the possible answers for each. That list works on all three vendors — and it's the part that actually takes work.
Want the model-selection guide and cost frameworks in one place? The AI Operations Playbook covers 50 business tasks AI can handle today, with cost frameworks, a model-selection guide, and an ROI calculator built for small businesses.
Top comments (0)