Published 20 October 2026 Updated 21 October 2026 By Hannah Press How many of the decisions that you made today actually needed you? A good chunk of the decisions any of us make at work are small, routine choices: which email to answer first, which…
Published 20 October 2026
Updated 21 October 2026
By Hannah Press
How many of the decisions that you made today actually needed you?
A good chunk of the decisions any of us make at work are small, routine choices: which email to answer first, which request can wait and, if you’re me, what time to crack open the second Coke Zero of the day—and yes, that’s a very important work decision. Whatever your routine decisions look like, AI
is starting to take them off your plate.
For the past two years, AI companies have raced to build AI models
that reason longer to solve harder problems. Decision models do the opposite. Given a question with a fixed set of possible answers, they quickly pick one (an option, a point on a scale or a yes/no) and attach a confidence score—a number for how likely the answer is to be right. When the confidence score falls below a threshold the developer sets, the decision can be routed to a human.
The latest AI trends, brought to you by experts
Get curated insights on the most important—and intriguing—AI news. Subscribe to our weekly Think newsletter. See the IBM Privacy Statement
.
Thank you! You are subscribed.
Your form was submitted. An IBM representative will call or email you shortly.
Where decision models came from
Last month, TypeSafe AI introduced the “System One” decision model category with Jev
, a model built to make fast decisions that software can act on directly. OpenAI, Microsoft and others followed within weeks, and people are already finding creative ways to put the models through their paces
.
A recent Mixture of Experts panel
said that while the technology behind decision models isn’t anything new, how it’s being used is.
Kaoutar El Maghraoui
, Principal Research Scientist at IBM, explained that a typical large language model (LLM)
writes its answer one token (a word or part of a word) at a time, feeding each one back in before choosing the next. That loop is why responses take time and why formatting sometimes breaks.
A decision model reads the input once and goes straight to an answer. “This is kind of a systems design correction,” she said. “They’re just refocusing the models for the right tasks.”
Watch the full episode of Mixture of Experts
Small decisions, bigger stakes
Volkmar Uhlig
, CTO and VP, Data Platforms at IBM, credited the category’s recent surge in popularity to the rise of AI agents
, which can complete tasks on behalf of humans. Agents running in the background need a quick, predictable answer to decide what happens next. With decision models, “suddenly we are in the world of programming and not anymore in the world of human speech,” Uhlig said. People chatting with an LLM want “a little bit more meat around” a yes or no response, so the two kinds of models serve different, complementary jobs: decision models make the quick calls inside software, and LLMs handle the conversation.
That’s also why the decision model’s confidence score matters so much. Once a workflow is running, decisions that clear the threshold can be acted on without a person ever seeing them. In a recent IBM Institute for Business Value survey
of 2,000 CEOs worldwide, respondents said 25% of operational decisions are currently made by AI, without human intervention. By 2030, leaders expect this to nearly double to 48%.
A lot depends on the model flagging when it might be wrong. But as we know, AI isn’t 100% accurate, and neither are humans. Nick Gallagher
, Staff Writer at IBM Think, recently reported on AI forecasting models
and said in the episode that even the best human forecasters struggle with “knowing the degree to which they’re sure about a particular answer.” The value of a decision model, he said, is being transparent when it isn’t sure, so a human can step in and make that judgment.
When confidence can be trusted
Mixture of Experts host David Zax
, in a conversation outside the episode, framed it as a management question. Would you rather manage a new hire who’s sure of themselves 90% of the time but right only 60% of the time, or one who’s right less often but knows it and checks in? Ideally, we can count on our teammates to know what they don’t know and to ask when they’re unsure. That’s a fair standard for these models, too.
The catch is whether the score can be believed
. LLMs are known for being overconfident. As Gallagher put it, they’ll say, “I’m totally sure about the fact that the Giants will win the Super Bowl this year.”
Decision models are designed to be calibrated
, meaning their confidence should match how often they’re right. If the score is off, being confidently wrong leads to more mistakes and more time spent fixing them.
Call it the new hire test. Does the model reliably flag the decisions it’s likely to get wrong? And is someone still checking its work?