Microsoft has launched Decision-1, a model that does not chat. It answers bounded questions in a fraction of a second and says how sure it is. A month ago this kind of AI barely existed. For Europe’s small businesses it may turn out to be the most useful kind yet.
FROM THE EDITORS
Every old post office had a sorting room. Letters came in by the sackful, and the sorters did not read them. They glanced at the address and threw each one into the right pigeonhole, thousands an hour, almost never wrong. The writing happened elsewhere. The sorting made the whole system work.
On 9 October Microsoft introduced its own sorter. Microsoft-Decision-1 is what the industry has started calling a decision model. You give it a question with fixed answers, such as “is this transaction fraud” or “which team owns this bug”, and it returns a probability for each option. Software acts on the answer directly. Nobody reads a paragraph first.
A category born in September
Decision models are new. TypeSafe, a San Francisco start-up founded by a former OpenAI researcher, came out of stealth in mid-September with a model called Jev and a $40 million seed round. Within weeks the field filled up. OpenAI opened a Decisions API. Cloudflare released Clef and Clef-flash with open weights under the Apache 2.0 licence. And SiliconANGLE reports that TypeSafe has since closed an $870 million round at a $7.5 billion valuation.
Microsoft’s entry is built for speed and price. In its announcement, Achint Srivastava of Microsoft’s Office of the CTO says Decision-1 is post-trained from Qwen3.5-9B for single-pass decision scoring. It costs $0.042 per million input tokens, and output is free. Microsoft reports a median response about 35 times faster than its comparison frontier model, and the highest accuracy in its own 36-benchmark comparison covering nearly 150,000 questions. It is available in Microsoft Foundry and through OpenRouter.
The number that matters: how sure it is
Speed is the headline. Calibration is the story. A calibrated model’s confidence matches how often it is right: when it says 90 per cent, it should be right about nine times in ten. Microsoft says this is a design goal, and commentator Aakash Gupta, reading the chart Satya Nadella shared, points to a calibration score of 92.2 out of 100.
Gupta puts it neatly with two new hires. One is right 83 per cent of the time and tells you exactly when she is unsure. The other is right 95 per cent of the time and sounds certain about everything. The first is the one you can trust with real work, because you know which answers to check. A model that knows when it does not know is a model that hands the hard cases to a person, and lets that person spend the day on them.
Microsoft also reports consistency. Decision-1 changed its answer on 1.3 per cent of slightly altered inputs, and not at all when the options were reworded or reordered. “Equivalent inputs should produce equivalent decisions,” the announcement says. For a sorter, that is the job description.
What it already does at Microsoft
The company’s own examples are modest and practical. Xbox Research used it to sort more than 10,000 pieces of player feedback into fixed themes, with quality it calls competitive with its frontier model at more than 14 times the speed and 200 times lower cost. The Copilot team uses it to check the quality of chat and agent responses. Engineers on call use it to find the right knowledge during incidents.
None of these replaces anyone. Each takes a pile of small, repetitive judgements off a person’s desk so the person can act on the result.
The sorting room for a firm in Ghent or Gothenburg
Every small business has its own sackful of letters. Support emails to route. Invoices to flag. Leads to rank. Reviews to tag as praise, complaint or question. Until now, using AI for each of those meant paying a chat model to write an essay and then reading it. A decision model makes each judgement cost almost nothing and take almost no time, so a firm can afford to sort everything, not just the urgent pile.
Two European details deserve attention. First, choice: with OpenAI, Microsoft, Cloudflare, TypeSafe and others in the field, no small firm needs to lock in. Second, control: Cloudflare’s Clef weights can be downloaded and run on your own servers, which is a straightforward way to keep sensitive data inside the EU if that matters to your customers.
What to try this week
- Find your sackful. Pick one repetitive yes-or-no or pick-one-of-five judgement your team makes dozens of times a day.
- Write the pigeonholes. List the fixed answers. A decision model is only as good as the options you give it.
- Use the confidence. Let the model handle the cases it is sure about and send the uncertain ones to a person.
- Check a sample. Compare a hundred of its answers with a colleague’s before trusting it with more.
- Mind the data. If the inputs include customer details, check where the model runs and under which terms.
The chat model taught small firms that AI could write. Decision models teach a quieter lesson: most of the work in a business is not writing, it is sorting. Hand the sorting to a machine that knows when it is unsure, and the people get their day back for the letters that really need them.
Sources
- Microsoft Command Line, “Introducing Microsoft-Decision-1”, 9 Oct 2026 — https://commandline.microsoft.com/microsoft-decision-1-model-foundry/
- Aakash Gupta (@aakashgupta) on X, 10 Oct 2026, commenting on Satya Nadella’s announcement
- SiliconANGLE, on TypeSafe’s seed round and Jev, 16 Sep 2026 — https://siliconangle.com/?p=847135
- Let’s Data Science, “Cloudflare Releases Open-Weight Clef Decision Models” — https://letsdatascience.com/news/cloudflare-releases-open-weight-clef-decision-models-c52886d6
- AI Weekly, “OpenAI launches Decisions API public beta on GPT-6 Luna” — https://aiweekly.co/alerts/openai-launches-decisions-api-public-beta-on-gpt-6-luna
Leave a Reply