AI CREATIVE REVIEW / ANALYSIS

OpenAI Decisions API: test one creative check before automating approval

The Decisions API is now generally available. Creative teams can use its bounded answers to flag assets for review, then measure mistakes before trusting the routing.

WASSAI editorial desk ·

Official OpenAI announcement graphic reading Build more responsive apps with the Decisions API on a purple and orange background.
Official OpenAI announcement artwork for the Decisions API. It is not an example of a reviewed campaign asset or a WASSAI test result. · Image source

OpenAI made its Decisions API generally available on 22 October 2026 after a public beta. It asks a model to answer defined questions with bounded results rather than write a long response. Developers can send text, images or structured inputs and request a probability, a choice among supplied options or a score against ordered levels. For a creator team handling many ad variations, that suggests a narrow first-pass check: identify which assets need human attention before they reach final approval. This is a developer API, not a button in a design editor. WASSAI has not independently tested it. [1]

Choose a question with a verifiable answer

Start with one campaign and one explicit rule. For example: “Is the pack clearly visible?” or “Does this feedback concern copy, imagery or the offer?” Supply the approved brief as context and define the permitted answers before sending anything to the API. OpenAI calls these predicate and choice questions. Its documentation also describes score questions for ordered rubrics. Keep the task small enough that a reviewer can independently label the same material. [1]

Collect 30 to 50 past assets or comments with known outcomes, including awkward crops, tiny text, conflicting requests and genuine exceptions. Have two people label them without seeing the model output. Then compare the API result with that reference set. Record false passes as carefully as false alarms: an overlooked claim, wrong pack or missing offer term can matter more than an extra review request. This sample size is a practical starting point, not a vendor benchmark.

Route work; keep approval human

A probability is an estimate, not a certificate of correctness. Pick a threshold from your own labeled examples and send uncertain cases to a person. For a choice question, include an “other” route so unusual feedback has somewhere to go. If the team cannot explain why an asset was flagged, inspect the input, instructions and approved reference before changing the threshold. Do not silently turn a routing tool into an automatic publish switch. [1]

Check the output at the actual placement size. A square ad and a vertical story can share the same design file while hiding different details after cropping. Keep the final check for legal wording, numbers, product identity and brand judgment with the people accountable for the work. Log each correction and update the example set when the campaign changes.

Measure whether the extra step earns its place

OpenAI says the generally available API adds native GPT-Live integration, prompt caching and support for up to 400 questions in one request. Its speed comparisons and customer results come from OpenAI and quoted adopters, so treat them as claims to verify in your own workflow. Time the full review loop, including setup, failed flags and human corrections. Keep the feature only if it reduces missed issues or review effort without creating new mistakes. [1]

Sources

  1. OpenAI: Build more responsive apps with the Decisions API · Published 2026-10-22; checked 23 October 2026.