Gemini Decisions API: High Confidence Can Still Mean No

Google’s Gemini Decisions API gives applications typed probabilities instead of generated answer text. Before using it to route work, check the meaning of each field: for a yes/no question, high confidence can mean a confident no. A confidence-only approval condition can therefore invert your intended policy. Use the answer value, an independently enforced permission check and a tested fallback—not confidence alone.

What changed

On November 11, 2026, Google announced the Gemini Decisions API, powered by gemini-decision-1. The endpoint evaluates bounded questions over shared text, images, audio, video or PDFs without generating answer text token by token. It supports yes/no questions, selection from a fixed option set and ordered rubric scores. This is a new API contract to evaluate, not evidence that a classifier should control your production equipment.

The technical documentation specifies up to 1,000 questions within a shared 262,144-token context budget that includes all questions. Choice questions accept 2–255 options. Questions are evaluated independently; one cannot consume another question’s answer inside the same request. If classification depends on an eligibility result, implement that dependency in application code or a subsequent request.

Access, cost and deployment boundaries

The documented access path is the Gemini Developer API using v1beta. Google’s announcement offers an AI Studio no-cost tier for prototyping and prices the model at $0.10 per million input tokens, with no output-token charge. Shared input is billed once per request; each question contributes its own instruction and criteria tokens. Those are announced model charges, not a complete workflow budget: retries, upstream media handling, review work and downstream execution remain separate costs.

The launch says support in Gemini on Google Cloud is coming soon. Do not substitute that announcement for a verified regional Cloud endpoint, contract or processing-location commitment. For a German enterprise pilot, establish the exact service, account eligibility, approved data classes, retention terms and processing locations before uploading real support tickets or plant imagery. Legal and contractual conclusions require counsel; an API key is not procurement approval.

Google reports 150–250 ms end-to-end latency for its measured setup, with the detailed latency discussion covering contexts up to 16K tokens and nearly flat latency from 1 to 100 questions. That does not establish the same latency at the maximum context and question count. We have not independently benchmarked the endpoint. Measure your media mix, network path, concurrency and tail latency before committing to an interactive response target.

The confidence-polarity trap

The yes/no type is named noul in the API. Its noul value is the estimated probability of yes. Google defines confidence as max(noul, 1 − noul). Consequently, confidence describes certainty in the more likely answer, not the probability of permission, eligibility or safety.

Consider a deliberately synthetic response for the question “Is this request eligible for automated routing?” If noul is 0.01, confidence is 0.99. A rule that proceeds when confidence is at least 0.95 would approve an answer that strongly says no. The values here are constructed test inputs, not observed Gemini outputs. A local policy fixture executed for this article checks this polarity error, the positive case, uncertainty, missing output and denied permissions. It does not test model accuracy or calibration.

For a positively worded eligibility predicate, use the yes probability against a threshold selected on labelled examples. Keep the question wording and threshold version together. Negating the question—“Is this request ineligible?”—changes the direction of the decision even if the field names are identical. Do not hide that semantic change behind a prompt edit.

Confidence is also not a second independent safety signal: for a binary response it is computed from the same probability. Combining a yes-probability condition with its derived confidence does not provide two independent checks. Permissions and operational constraints must come from authoritative systems outside the model.

Closed choices still need an outside option

Google’s docs explicitly warn that a choice question distributes probability among the supplied options even when none fits. A winner among “billing”, “shipping” and “technical” is not proof that an employment complaint belongs in any of those queues. Add a clearly described other/review option and test an independent scope predicate; both remain model judgments that need evaluation.

For industrial service triage, use the endpoint to propose a queue, not to release a machine action. A scoped maintenance request can still lack a valid asset identifier or current work authorisation. The existing OT integration boundary explains why operational and safety controls must remain separately enforced. A high-probability route must not grant access to a tool or asset.

Ordered rubric scores need their own adapter. Google returns a probability-weighted level index, which can lie between levels. Different distributions can yield the same average; a middling score can represent ambiguity between extremes. Inspect the per-level probabilities when that distinction affects handling. Do not treat a rubric score as a probability or silently round it into a release instruction.

A decision contract to test before rollout

The following is a proposed acceptance checklist, not a Google-prescribed production policy. The numeric threshold used in the local fixture is arbitrary and exists only to expose incorrect branching. Production thresholds require task-specific labels, error-cost decisions and review-capacity limits.

Version the model identifier, question text, option descriptions, input construction and policy together. Retain structured reason codes, response status and the policy version under appropriate access controls. Avoid storing raw personal or industrial payloads merely to make debugging convenient. The RAG abstention guide covers a related evidence-sufficiency decision; this launch adds a different integration hazard: typed fields with different semantics must not share one generic confidence adapter.

Compare against the workflow you already operate

Against a generative model with structured output, the Decisions API removes answer-text generation and offers a direct probability interface. It does not replace extraction, explanations, tool arguments or multi-step reasoning. Against an existing deterministic routing rule, it adds an external model dependency and uncertainty; keep the rule if explicit fields already determine the answer reliably.

Run shadow evaluation first on an approved, labelled set of German and English requests. Include in-scope positives, clear negatives, out-of-scope messages, malformed media and adversarial text. Report wrong automatic routes among automatically routed cases, automatic coverage among eligible inputs, and review arrivals against staffed capacity. Separate the two languages and consequential request classes instead of accepting one blended accuracy number.

Then test the non-model branches: missing answer key, invalid probability, timeout, safety-filter rejection and disabled downstream permission. Google documents safety-filter blocks as HTTP 400 rather than an answers object. Your adapter must not turn that absence into success or repeatedly resubmit blocked material without an approved handling policy. The safe fallback for this proposed service-triage pilot is a bounded review path or an explicit unavailable result—not a plant-control command.

A narrow pilot, not automatic approval

The useful first deployment is one reversible routing decision with no direct actuation, a named review owner and a rollback switch. Confirm service terms and data handling, compare against the current rule, and release only when the complete policy meets predeclared error and workload limits. Vendor calibration claims do not establish calibration on your German technical vocabulary or your changing request mix.

If your team needs an implementation review, start with one question, its exact yes/no meaning, the current routing rule and the fallback owner. Through an AI automation consultation, we can scope the adapter tests and acceptance evidence before enabling automated routing. The deliverable is a decision contract that fails visibly—not a reassuring confidence number attached to every action.