What A Decision Model Is, Next To A Chatbot
A decision model is a small language model tuned to answer one kind of question with a choice from a fixed list, scoring each option so that software can act on the result directly.
A chatbot generates text and reasons through open problems. A decision model gives up that freedom in return for speed, low cost and constrained output, and Microsoft describes its model as built to deliver structured outputs that software can act on immediately. The idea of a model that returns probabilities for a fixed list is not new, since machine learning researchers have built probability-based classifiers for a long time, as The Register notes. What is new is that the options come with plain-language descriptions and a language model does the scoring.
Outputs come in three forms according to heise: a yes or no, one choice from a defined pool, or a ranking of options, always as numbers such as probabilities and never as an explanation.
Microsoft-Decision-1 As The Worked Example
Microsoft announced Microsoft-Decision-1 on 9 October as a model built for fast, structured decisions, and its figures show what the category promises.
| Property | Microsoft-Decision-1 |
|---|---|
| Base model | Qwen3.5-9B from Alibaba Cloud, post-trained for single-pass decision scoring |
| Accuracy | 83.5 percent across 36 benchmarks covering nearly 150,000 questions |
| Latency | 85 milliseconds; 2.5 times faster than H2O-Lightning-4B and 2.8 times faster than Jev in Microsoft's test |
| Price | $0.042 per million input tokens; output tokens free |
| Availability | Microsoft Foundry and OpenRouter |
| Consistency | Decision changed on 1.3 percent of perturbations; no flips when option descriptions were paraphrased or options reordered |
| Planned change | To be rebased on Microsoft AI and OpenAI models, no reason given |
These are Microsoft's own numbers from Microsoft's own tests, and heise points out that the blog repeatedly compares the model with GPT-6 Sol, a frontier model for complex language tasks, which makes that comparison of limited value.
The category is crowded. The Register counts more than 100 such models, from Jev, announced about three weeks before its 10 October report, to OpenAI's Decisions API in public beta, Cloudflare's open-weight Clef, Strands Decider 2B, Liquid AI d1, Perplexity's Decisions API, a Snowflake model, Surogate Rune and H2O-Lightning-4B.
Where They Fit And Where They Do Not
Decision models fit steps where software needs a quick, bounded judgement, and Microsoft lists routing, classification, prioritisation, verification and workflow control as the targets.
Uses named in Microsoft's post and heise's summary include routing a request to the right AI model by quality, cost and latency, validating an input against quality criteria, sorting customer feedback into categories, deciding whether an agent should continue, stop, retry or hand off, and picking the best action from a set when controlling a computer or robot.
They do not fit open-ended writing, and their limits are visible. They give numbers without reasons, heise notes, and its reading is that a decision model is steadier than many general models but less reliable than an algorithm or a decision tree, because a 1.3 percent flip rate means answers can still change. The Register adds that decision models can still make errors and points to research on how those errors might be magnified. Microsoft states a calibration goal, that a 90 percent prediction should be right about nine times in ten, so ask for calibration measured on your own data.
Five Questions To Ask Before You Buy One
Ask which base model sits underneath and under what licence, because Microsoft-Decision-1 is built on Alibaba Cloud's Qwen3.5-9B and the same question applies to every decision model. If your procurement policy limits where AI components come from, this answer decides the rest.
Ask when and how the base will change. Microsoft says it will rebase Microsoft-Decision-1 on Microsoft AI and OpenAI models without saying why, so today's results describe a base that is due to change; ask for the same tests on the new base before you rely on it.
Ask how it was tested. Request the benchmark set, who chose the comparison models, the flip rate for your own phrasings and a calibration check on your data, then run a sample of your real cases yourself.
Ask how decisions are logged and audited. The output is a probability with no explanation, so keep the inputs, the options and the scores for every decision, and agree a threshold below which a person decides.
Ask about price and latency in your own setup. Microsoft quotes $0.042 per million input tokens with free output and 85 milliseconds, which makes bulk routing and checks cheap, but check the numbers on your own traffic.
Read next: The EU Now Audits Your Cloud Vendor Directly | The EU Just Moved Your AI Act Deadline by 16 Months
