Tool Reviews

Former OpenAI Researcher Creates AI That Only Chooses From a List, and Three Giants Jumped In Within 16…

Decision APIs replace text with typed choices and probabilities, giving developers reliable signals for routing, retries, and production workflows.

Former OpenAI Researcher Creates AI That Only Chooses From a List, and Three Giants Jumped In Within 16…

Key Takeaways

  • Decision APIs such as Jev, OpenAI’s Decisions API, Cloudflare Clef, and Strands Decider return typed choices with probabilities instead of free-form text.
  • This approach helps developers build more reliable AI features for agents, automation, moderation, routing, and workflow decisions.
  • Probabilities give applications a second signal for deciding whether to act, retry, ask for clarification, or escalate to a human.
  • Builders should check output schemas, calibration, fallback behavior, latency, pricing, and data handling before adopting these tools.
  • The broader trend is AI infrastructure moving from chat-style outputs toward decision layers that produce values software can directly use.

![Abstract friendly illustration of typed choices and probability bars](data:image/svg+xml,%3Csvg%20xmlns='http://www.w3.org/2000/svg'%20viewBox='0%200%20800%20450'%3E%3Crect%20width='800'%20height='450'%20fill='%23f7f9fc'/%3E%3Ccircle%20cx='180'%20cy='225'%20r='70'%20fill='%233b82f6'/%3E%3Ccircle%20cx='400'%20cy='225'%20r='55'%20fill='%2310b981'/%3E%3Ccircle%20cx='620'%20cy='225'%20r='40'%20fill='%23f59e0b'/%3E%3Crect%20x='130'%20y='320'%20width='100'%20height='12'%20rx='6'%20fill='%233b82f6'/%3E%3Crect%20x='350'%20y='320'%20width='80'%20height='12'%20rx='6'%20fill='%2310b981'/%3E%3Crect%20x='570'%20y='320'%20width='60'%20height='12'%20rx='6'%20fill='%23f59e0b'/%3E%3C/svg%3E)

Former OpenAI Researcher Creates AI That Only Chooses From a List, and Three Giants Jumped In Within 16…

Meta description: Jev, OpenAI’s Decisions API, Cloudflare Clef and Strands Decider return typed choices with probabilities instead of text, helping builders get predictable outputs.

Tags: decision APIs, structured outputs, OpenAI’s Decisions API, Cloudflare Clef, Strands Decider

For developers building AI features that need reliable actions, a roundup from Towards AI names Jev, OpenAI’s Decisions API, Cloudflare Clef and Strands Decider as four decision APIs that return typed choices with probabilities instead of plain text. The shift matters because text is easy to read but hard to trust in production: a model can answer with a sentence, a tool name, an invalid label, or a vague refusal. A typed choice narrows the model’s response to a known set, while a probability gives software a score it can use to route, approve, retry, or ask a human.

Why typed choices matter more than chat text

Large language models are good at producing fluent language. Fluent language is not always the right product output. If a customer support agent needs to decide whether to create a ticket, ask for more detail, or close a case, a paragraph can be more work than help. The application then has to parse the paragraph, guess the intent, and risk getting the decision wrong.

The tools listed in the report point to a different pattern: the model is asked to choose from a constrained set of outcomes. In practice, that could be an enum, a boolean, a category, a tool name, or another structured value. The source does not specify each tool’s supported schema, but the shared idea is that the API returns a decision the program can use directly.

Former OpenAI Researcher Creates AI That Only Chooses From a List, and Three Giants Jumped In Within 16…

This is useful for AI agents, workflow automation, content moderation, routing, and any feature where the next step depends on a clean label. It also reduces a common failure mode: the model sounds confident while producing a value that the backend cannot accept.

Probabilities turn a model output into a decision

A typed choice answers “what should the model select?” A probability answers “how much should the system trust that selection?” The combination is what makes these APIs different from ordinary text completion.

A probability can drive simple rules. A high score might let the system act automatically. A medium score might trigger a clarification question. A low score might send the case to a human. That does not mean the score is always perfectly calibrated, but it gives developers a second signal beyond the label itself.

For users, the practical benefit is fewer brittle integrations. When an AI feature returns a known value plus a confidence signal, the surrounding product can behave predictably: retry a task, show a fallback, log the uncertain case, or keep the user informed. When it returns a free-form sentence, the product team often has to write extra parsing rules, and those rules can break as the model’s wording changes.

What builders should check before adopting these tools

The report names the APIs, but it does not provide enough public detail to treat all four as interchangeable. Before adding one to a workflow, check the official documentation for the exact output schema, supported types, error behavior, latency, pricing, data retention, and whether probabilities are exposed consistently across requests.

Former OpenAI Researcher Creates AI That Only Chooses From a List, and Three Giants Jumped In Within 16…

Developers should also ask how the system handles ambiguous inputs. A decision API that forces a choice may be useful, but it can also hide uncertainty if the model is pushed to pick an option when none is appropriate. Good design usually includes a fallback: a “no match,” “unknown,” or “needs review” value, plus a way for the application to escalate.

Calibration matters. If a score of 0.9 is used to approve actions automatically, the team should know whether that score corresponds to 90% accuracy in real conditions. Without evaluation data, probability can still help, but it should not be treated as a guarantee. The source does not list benchmark numbers for these tools, so teams should run their own tests with representative data.

What to watch next

The broader signal is that AI infrastructure is moving beyond chat boxes and toward decision layers. Agents, copilots, and automated workflows need outputs that applications can act on, not only read. Jev, OpenAI’s Decisions API, Cloudflare Clef and Strands Decider all fit that trend by returning typed choices with probabilities instead of text.

The next thing to watch is how openly each product defines “decision” and “probability.” If the APIs publish clear schemas, integration examples, and reliability notes, they become easier to adopt. If they remain vague, builders may need to test them directly before putting them in front of real users.

For now, the useful takeaway is simple: when an AI feature must choose, choose a pattern that returns a value the program can use. Text can describe a decision; typed choices with probabilities can be the decision.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.