Google launched Gemini Decisions API on November 11, 2026, giving customer-support and document-processing teams a new way to classify incoming material. The service answers predefined questions about supplied records with probabilities, rather than writing a conversational reply.
An API is a connection developers build into other software: this is an integration tool, not a ready-made office app. Google says it is available through the Gemini Developer API and Google AI Studio; Google Cloud support is still forthcoming, with no date specified.
Fixed answers, not a finished task
The Gemini Decision model accepts text, images, audio, video and PDFs. It can assess yes/no questions, choose from a fixed list or score material against defined rating levels. A request can contain up to 1,000 independent questions, provided the material and questions fit within its input limit.
For example, a retailer could use it to identify a message reporting a damaged order and suggest a support queue, following the customer-message example in Google’s documentation. Classifying a report of damage does not independently verify the damage or authorize a refund.
Questions in one request cannot see each other’s answers, so dependent steps still need application logic or another request. The model does not write explanations or extract a custom set of fields. For jobs involving several actions or a completed document, see our guide to AI agents for office paperwork.
What happens when none of the choices fits?
Google’s choice-question documentation identifies an important limitation: probabilities are distributed across the supplied options even when none is appropriate. A winning category is therefore not proof that the list covered the case. Its confidence score comes from those same probabilities, not a separate verification.
Include an explicitly defined “other” category and a separate route for human review. A check that asks whether the material belongs in the workflow can also help. Raising a confidence threshold alone does not fix an incomplete set of categories.
Google recommends setting thresholds using labelled examples from the application’s own traffic and retaining a human or second check for high-stakes actions such as payouts. Routing a request and approving an outcome should remain separate decisions.
Pricing and customer-data rules
Google’s current rate card lists a free tier and a standard paid rate of US$0.10 per million input tokens—the units used to meter the material processed. There are no output-token charges because the model generates no answer text.
Shared material is billed once within a request; each question adds its own instruction, criteria and formatting tokens. That makes grouping independent checks relevant to the bill, but does not make extra questions free. Integration work, retries and human review still contribute to the cost of a completed task.
Before uploading customer records, check which Google data-use terms apply. Under the ordinary unpaid-service rules, submitted content and responses can be used to improve products and reviewed by people. Google instructs users not to submit personal, confidential or sensitive material under those rules.
Paid-service rules exclude prompts and responses from product improvement, but allow limited safety-related logging. Those data-use rules also apply in the European Economic Area, Switzerland and UK, even for unpaid quota. Billing-linked projects and qualifying AI Studio accounts can change the applicable treatment: a free session does not by itself establish which rules apply.
For a business considering a pilot, the useful starting point is one well-defined sorting task, permitted sample data and a reviewer who can check mistakes—not automatic approval of the resulting action.