OpenAI's new Decisions API returns an answer that software can act on without first generating a paragraph. The company introduced the public beta on October 6, using GPT-6 Luna, and advertises roughly ten times the speed of its Responses API for typed decisions. That is OpenAI's claim. The Founders Report has not benchmarked it.
For a SaaS founder, the interesting use is a small task with a clear destination: deciding which queue should receive a support request, for example. Whether a request belongs in billing is a different problem from composing a good reply to it.
First check whether classification is actually delaying the customer. If requests wait hours for an available agent, saving a fraction of a second on routing may change very little. A pilot makes more sense when the application is waiting on that decision and the answer fits a fixed set of destinations. If existing rules already route the requests correctly and quickly enough, leave them alone.
Start with a decision you can describe precisely
According to OpenAI's documentation, the API supports checks that return a probability, choices from a supplied set, and scores against defined levels. OpenAI directs developers to Responses for custom structured objects or tool calls instead.
Consider a hypothetical support inbox where the billing team handles payment changes and a separate technical team investigates a product that stops working after a customer renews. Both messages contain billing language, but the second customer needs someone to investigate the failure rather than explain the renewal charge.
Define the route for that second case before you evaluate the model. Otherwise the test will measure disagreement inside your own team as though it were a model error.
Also give the application somewhere to put an uncertain answer. A general review queue is a legitimate destination. Silently forcing every request into a confident category makes the demo look cleaner and the inbox harder to operate.
Measure what happens after the answer
OpenAI's evaluation guidance recommends task-specific tests using representative data. For this inbox, that means including confusing messages and cases your staff have previously rerouted. A set of obvious requests will tell you very little about the work people actually need help with.
The following is a proposed evaluation, not a report of tests we ran. Have someone who handles the inbox assign the expected destination for historical requests your company is permitted to process, after removing personal information that the test does not need. Keep a separate test set. Using the same requests to tune the routing and judge it would leave you without evidence that it can handle messages it has never seen.
Run the candidate alongside the existing route. Compare the proportion of requests that land in the right queue, then look at the mistakes individually. A misrouted routine enquiry and a missed cancellation request should not have the same business weight merely because each counts as one error.
Record elapsed time from the application receiving the request to the queue accepting it. Report typical routing times and how long the slower requests take. A fast model cannot fix a slow connection to the queue, and an average can hide the cases where a user sits waiting.
Include rerouting work in the cost comparison. Somebody has to fix the mistakes. A cheaper model response can make the process more expensive if staff must reroute more requests.
Keep classification separate from permission
In this example, the model is helping route a message. It does not need permission to issue refunds or change a subscription. Keep those actions behind their own checks.
A refund request belongs with billing. That does not mean the customer qualifies for a refund. Combining both questions in an informal prompt makes it harder to understand which decision caused the result, especially when the message is routed correctly but the requested payment should be refused.
Teams considering a larger automation should document the workflow before automating it. For this smaller change, write down who handles uncertain requests and how a mistaken route gets corrected. Those instructions are part of the product, even if they never appear in a launch announcement.
Before approving a migration, have the support lead review the errors and timing results. Include the staff time spent correcting mistakes before deciding whether the new route saves work. The speed claim alone cannot answer that.