Can AI Make Decisions for Your Business?
Can AI make decisions for your business? Yes. How decision models and confidence scores work, what to hand over in a fab shop, and what stays with you.
Yes. AI can make routine business decisions today: routing an RFQ, sorting a drawing, flagging a risky job. New decision models return a confidence score with every call, so AI handles the clear cases within limits you set and sends anything uncertain to a person. Owners keep the judgement calls, like price and which jobs to chase.
In the last five weeks, TypeSafe AI, OpenAI, Cloudflare, Liquid AI and Microsoft have all launched AI models built for one job: making decisions. Not writing emails. Not chatting. Deciding.
So, can AI make decisions for your business? It already can, and it’s getting better fast. Let AI guide you and let it do the routine calls, with human guardrails where they’re needed. You stay the owner, and owners make the judgement calls.
Can AI really make decisions?
Yes, and your shop already runs on countless small ones. Which job does this email belong to? Is this drawing a revision or a new sheet? Does this packing slip match the PO? Is this RFQ worth a look?
Each one takes a person a few seconds to a few minutes. Together, they eat whole days. Think of a good lead hand: they don’t ask the foreman before every cut, because the calls are routine and they know the limits. The foreman steps in on the unusual stuff. That’s exactly the role AI can play now.
Is AI accurate enough to trust with decisions?
For the kind of decisions a shop makes, increasingly yes. But the evidence is mixed, so here it is straight.
The good news: when the best models summarize a document that’s in front of them, the 2026 Stanford AI Index reports that the top 15 models’ hallucination rates “range from 1.8% to 5.4%.” That’s a test of summarizing, not sorting or routing, so treat it as a rough guide, not a promise. But it sets the right expectation: even very reliable means ~1 miss in 20 to 50, which is why low-confidence calls go to a person.
The catch: on a tougher accuracy test, the same report found hallucination rates across 26 top models “range from 22% to 94%.”
Read those two together and the lesson is clear. AI is far more reliable when it’s reading what’s in front of it, like an email, a drawing or a PO, than when it’s guessing from memory. Shop decisions are the first kind. Give the AI your documents and your rules, and it’s working with the drawings on the bench, not trying to remember them.
How smart are the best AI models now?
Smart enough that “can it?” is the wrong question, and they keep improving every few months. But no model is perfect, so put the guardrails in your system, not just in the model: limits, logs and sign-offs belong in the workflow.
What are decision models?
Most people know AI that writes: ChatGPT-style large language models (LLMs). Decision models are a newer kind that doesn’t write at all. You give it a situation and a fixed set of answers, and it returns a choice, a score or a yes/no, each with a probability.
The simplest way to put it: the LLM writes, the decision model decides. An AI agent can use both: a decision model to sort and route, an LLM to draft the email.
Think of a go/no-go gauge versus a tape measure. The tape gives you a reading you have to interpret. The gauge just tells you pass or fail, instantly, every time. Decision models are gauges for information.
| Model | Maker | What it decides |
|---|---|---|
| Jev | TypeSafe AI | Choices, scores and yes/no, each with a probability |
| Decisions API | OpenAI | Yes/no odds, choices and scores on text and images |
| Microsoft-Decision-1 | Microsoft | Yes/no, multiple choice and ratings |
| Clef | Cloudflare | Odds for every option, on text and images |
| d1 | Liquid AI | One-pass answers on text and images |
TypeSafe says Jev is “two orders of magnitude faster and more efficient” than LLMs on these tasks. Some of these models you rent, others you can run on your own hardware; our post on open vs closed AI models covers the trade-offs.
We’ve seen this work first-hand. For Brassex, a custom furniture maker, we built a production management agent that reads client order emails and gathers what the shop needs to build each piece; today it “handles every order from the first email to final delivery,” and the team got 20 hours a week back.
Why does the confidence score matter so much?
Because it tells you when to let the AI act and when to call a person.
Microsoft puts it well: “Applications use confidence to decide when to act, defer, or ask for review, so a 90% prediction should be right about nine times out of 10 on representative cases” (Microsoft). TypeSafe makes the same point from the other side: “If a model can do a task 95% of the time but doesn’t say when it’s in the 5%, it can’t automate that task.”
So you set the line. Above it, the AI acts on its own. Below it, the call lands on a person’s desk with the facts attached. Start the line high and lower it as the AI earns trust, the same way you’d give a new hire more rope after a few good months.
What decisions can AI make in a fab shop?
Here are four, from easy to bold.
Routing RFQs. An RFQ lands. A decision model reads it and answers three questions: which estimator, how urgent, and does it match the work we do? Clear cases route themselves. The odd one goes to the chief estimator.
Sorting drawings and documents. Is this a revision or a new sheet? Shop drawing or erection drawing? Mill cert, packing slip or invoice? AI files it to the right job automatically and flags anything it isn’t sure about.
Flagging risky jobs. Before a bid goes out, AI scores it against your own red flags: unusual coatings, tight milestones, a GC you’ve had payment trouble with. It doesn’t kill the bid. It makes sure nobody misses the warning.
AI recommends, you decide. Bid/no-bid is the bold one. AI pulls the scope, schedule and risk flags, compares the job to the kinds you win and make money on, and gives you a recommendation with its confidence. You make the call in minutes, not hours. The decision is still yours; the homework isn’t.
For a dozen more, see our business process automation examples.
Which decisions should stay with a person?
The judgement calls. Owners, this is your job, and AI makes you better at it:
- The price. AI prepares the number. You set it.
- Which jobs to chase. Strategy is yours.
- Engineering answers. Anything stamped stays with qualified people.
- Sensitive customer conversations, about money, problems or schedule changes.
- Anything the AI flags as low-confidence.
One more thing to keep in mind: more automated decisions means more total mistakes, unless you check. It’s the flip side of Jevons paradox, the idea TypeSafe named Jev after: when decisions get cheap, you make many more of them. So log every call, spot-check a sample, and watch the low-confidence pile.
How do you get started with AI decisions?
Pick one decision your team makes dozens of times a week, like routing incoming emails. Run the AI alongside your people for a few weeks without letting it act. Compare its calls to theirs. When it agrees on the clear cases and flags the tricky ones, set your confidence line and let it run.
That’s how shops that reorganize around AI pull ahead, as we showed in the real benefits of AI in business. It’s not the tool. It’s letting the tool take the calls it’s good at.
Want help handing off your first decisions?
We’ll find the routine calls eating your team’s day and build the workflow automation and AI agents to make them, with limits and sign-offs where they matter. Book a free 30-minute consultation.
Sources
- Stanford HAI, The 2026 AI Index Report: Responsible AI chapter
- TypeSafe AI, Introducing System One Models & Jev (Sep 15, 2026)
- OpenAI, Decisions API public beta (Oct 6, 2026)
- Cloudflare, Introducing Clef (Oct 1, 2026) and Clef-omni (Oct 9, 2026)
- Brassex case study
- Liquid AI, Open d1 (Oct 7, 2026)
- Microsoft, Introducing Microsoft-Decision-1 (Oct 9, 2026)
Frequently asked questions
Can AI make decisions on its own?
Yes, for routine calls inside limits you set, like which job an email belongs to or whether a packing slip matches the PO. Good setups use a confidence score: above your threshold the AI acts, below it the call goes to a person.
How is AI used in business decision-making?
Two ways. It makes the small, repetitive calls itself, like routing, sorting and checking. And it prepares the big ones, like bid/no-bid or pricing, by pulling the facts together and giving a recommendation, so the owner decides faster with better information.
What are decision models?
A newer kind of AI model that doesn't write text. You give it a situation and a fixed set of answers, and it returns a choice, a score or a yes/no with a probability attached, which software can act on directly. Examples include TypeSafe AI's Jev, Microsoft-Decision-1 and Cloudflare's Clef.
What are some examples of decision models?
TypeSafe AI's Jev, OpenAI's Decisions API, Microsoft-Decision-1, Cloudflare's open-weight Clef family and Liquid AI's open-weight d1 models. All were released between September and October 2026.
Which decisions should stay with a person?
The judgement calls: final prices, which jobs to chase, engineering answers, sensitive customer conversations and anything the AI flags as low-confidence. Everything routine can run on its own within the limits you set.