10 min read - OpenAI Decisions API: Where Bounded AI Fits in a Workflow

OpenAI made its Decisions API generally available on 22 October 2026. The dedicated endpoint accepts text, images or structured JSON and returns a bounded answer: a probability, a choice from supplied options or a score against ordered levels. The answer format is narrow. The business consequence may not be.

That distinction should shape adoption. A misrouted support ticket can wait a few minutes; a payment, account suspension or hiring decision can affect money, access or a person's prospects. Both are choices, but they need different automation policies.

The useful buyer question is therefore smaller than “Where can we add AI?” Which recurring judgement has ambiguous input, a fixed set of answers, measurable error costs and a safe fallback?

The release narrows output, not responsibility

OpenAI's 22 October announcement says the API is generally available with GPT-6 Luna. It supports up to 400 questions in one request, prompt caching, images in individual question instructions and regional processing in the United States and Europe. OpenAI reports responses up to ten times faster than GPT-6 Luna through the Responses API. That speed figure is a vendor comparison, not a promise for a particular application.

The developer guide defines the three answer types:

• predicate estimates the probability that a condition is true;
• choice selects one value from options supplied by the application and returns probabilities plus a confidence field;
• score evaluates ordered levels and returns a probability-weighted score.

Typed output makes application code easier to inspect. It does not make the selected answer correct, nor does it decide what the application may do next. Product owners still own the labels, evidence, thresholds, fallback and effect of each outcome.

Choose the mechanism before tuning the model

Teams often start with a model demonstration and look for work it can absorb. Reverse that order. Write the operating question, identify the current decision-maker and compare four mechanisms.

This division prevents a probabilistic service from replacing a rule that already works. It also keeps generation and classification separate. OpenAI's guide directs developers to Structured Outputs when they need an object that follows a custom JSON schema, and to function calling when a model needs to request a tool action. Decisions fits questions with fixed answer types.

The boundary is more important than the provider. Our model-routing guide covers task, quality and latency choices across models. Here, the first choice is whether a model belongs at the decision point at all.

Price the two kinds of error

Accuracy alone hides the failure that matters. Suppose a system decides whether an incoming invoice needs manual review. A false positive adds review work. A false negative may let a material anomaly pass. The same accuracy can conceal very different costs depending on which cases the model misses.

Write an error ledger before setting a threshold:

Convert those effects into a threshold policy. Low-consequence routing might allow an automatic choice above a tested threshold and send the rest to a general queue. A material action may use the model only to prioritise review, never to execute the action. Record who can change a threshold and the evidence required.

OpenAI's guide says teams should set routing, filtering or review thresholds with labeled examples from their own application and account for false-positive and false-negative costs. A confidence field is therefore an input to policy. It is not a certificate of accuracy.

Build the evaluation set from the real queue

A tidy demonstration set rewards tidy behaviour. Production inputs contain incomplete sentences, copied email chains, scans, internal abbreviations and categories that no longer fit. Pull the evaluation set from the workflow the endpoint would enter, subject to the organisation's data rights and retention rules.

Each case needs the input, expected label, reviewer rationale and business effect of a wrong answer. Preserve disagreements instead of averaging them away. If two experienced reviewers apply different labels, the category definition may be the defect.

Slice the results by the differences that can change performance: language, document source, customer segment, input quality and rare class. A France-Morocco operation might serve requests in French, English or Arabic; testing one language would not establish quality in the others. That is a test-design example, not a claim about this API's language performance.

Our evaluation-driven development guide explains how to keep a golden set and regression gate. For a bounded decision, add calibration checks: when the service returns similar confidence, does the observed success rate remain similar across time and slices?

Run beside the workflow before entering it

Start in shadow mode. Send a copy of eligible inputs to the endpoint, store the answer with the model and policy version, but let the existing process remain authoritative. Review disagreements against the current baseline.

A practical pilot has four stages:

1. Freeze the question, choices and label definitions. Reject inputs outside that boundary rather than forcing them into the nearest class.
2. Backtest on labeled historical cases. Report each error type and slice, not a single average.
3. Run live in shadow mode long enough to capture ordinary variation. Record latency, failures, reviewer disagreement and changes in input mix.
4. Allow a reversible production route for the lowest-consequence band. Keep the manual queue, kill switch and audit record active.

Move the threshold only through a recorded review. A prompt change, new choice, model update or altered input source should trigger regression tests because each can change the distribution behind yesterday's result.

NIST's Generative AI Profile recommends post-deployment monitoring that covers appeal and override, incident response, recovery, decommissioning and change management. NIST presents the profile as voluntary risk guidance. It does not certify an OpenAI integration or prescribe a universal threshold.

Measure the whole path, not the endpoint alone

OpenAI currently lists uncached input at $0.10 per million tokens for GPT-6 Luna on the Decisions endpoint, with discounted cache hits, charged cache writes and no output token charge. The documentation also notes regional-processing and long-context multipliers. These terms were reviewed on 23 October 2026 and can change.

Operating cost also includes integration, labeling, human review, retries, monitoring and correction. Measure cost per correctly handled case and time to an accepted outcome. A cheap call that creates a review backlog is an expensive workflow.

The same applies to latency. OpenAI's deployment checklist recommends representative evaluations and comparison of task success, latency, token use and cost per successful task. Network time, fallback calls and a person's review can erase an endpoint-level gain. Test from the regions and systems that will carry real traffic.

Separate residency, retention and permission

The current Decisions documentation says regional processing is supported in the US and Europe, while data residency, Zero Data Retention and HIPAA use have eligibility, agreement and configuration conditions. Those are separate controls. Selecting a region does not establish that a particular organisation has Zero Data Retention, nor does a product option prove a legal or regulatory outcome.

Before a pilot uses real records, document the fields sent, the purpose, permitted region, retention setting, access path and deletion process. Confirm the active account configuration and contract rather than copying a statement from a product page into a security assessment.

Make the go or no-go record short

Approve a limited rollout only when the record answers these questions:

• Which exact input and answer set are in scope, and what is explicitly excluded?
• What are the observed false-positive and false-negative costs by important slice?
• Which confidence band may act, which band goes to review and what happens on failure?
• Who owns labels, thresholds, monitoring, incidents and rollback?
• Which provider terms and data controls were verified for the active account?
• What change forces a fresh evaluation?

Keep rules or human review when categories remain disputed, errors cannot be reversed, or the endpoint saves no end-to-end time. Proceed with a narrow route when the answer space is stable, the error tradeoff is acceptable and shadow results beat the current process on the measure the business actually cares about.

The Decisions API makes bounded model output easier to buy and integrate. Production authority still has to be earned one workflow at a time.

Explore Exceev's AI and engineering services if your team needs to map a decision boundary, build an evaluation set or test a reversible production route.

Sources and limitations

We opened and reviewed every source below on 23 October 2026.

• OpenAI's general-availability announcement is the primary source for the release date, supported inputs, current capabilities, regional-processing statement and vendor latency comparison.
• The Decisions API developer guide defines the answer types, question limit, threshold guidance, usage reporting, current pricing and eligibility caveats for data controls.
• OpenAI's API deployment checklist recommends representative evaluations and measuring success, latency, token use and cost per successful task.
• NIST's Generative AI Profile supplies independent, voluntary guidance on monitoring, override, incident response and change management. It does not evaluate the Decisions API.

We did not run the endpoint, inspect OpenAI's model or reproduce its latency and customer performance claims. We did not verify a customer's contract, region, data controls or production workload. The mechanism matrix, error ledger, shadow-pilot design and go or no-go questions are Exceev's analysis, not requirements from OpenAI or NIST. This article is technical and operational guidance, not legal advice or a claim that a bounded answer produces a safe or compliant business decision.