A customer writes that their order never arrived and attaches a screenshot showing a duplicate charge. Should the ticket go to shipping, billing or a reviewer who can see both issues? Google’s Gemini Decisions API, introduced on November 11, 2026, gives developers a way to ask those questions together and use the answers to route the ticket.
The API is powered by Gemini Decision 1. Available through the Gemini Developer API’s v1beta endpoint, it scores predefined answers and returns probabilities that an application can use in its own rules. That makes it useful when the immediate job is to choose a queue, check a condition or rate a piece of evidence.
How Gemini Decision 1 handles the same ticket three ways
A request can contain text, images, audio, video and PDFs, with up to 1,000 independent questions about the shared input. The input and all questions together must fit within the 262,144-token context window. Each question is evaluated independently, so the application decides how to combine the results. Google’s Decisions API documentation
| Question type | Example for a support ticket | What the API returns |
|---|---|---|
| Yes or no | Does the customer report a duplicate charge? | The estimated probability that the answer is yes |
| Multiple choice | Which queue should see this first? Options include billing, shipping, technical support and general triage. | The top choice and a probability for each option |
| Rubric score | How complete is the information needed for an investigation? | A weighted score across the predefined levels, such as missing, partial or complete |
These answers serve different purposes. A ticket can clearly mention a duplicate charge and still omit the order number. The application can flag the billing issue and send the incomplete case to a reviewer in the same workflow. It can also give certain signals priority, such as an explicit request to dispute a charge.
What a confidence score of 0.90 means
Imagine the API gives billing a probability of 0.90 for a ticket. For a multiple-choice question, the confidence score is the probability assigned to the top option. In a well-calibrated system, billing would be the correct destination for roughly 90 out of 100 comparable tickets scored near 0.90. You can check that by grouping past predictions into probability ranges and comparing them with the outcomes assigned by reviewers. Scikit-learn’s guide to probability calibration
The categories matter as much as the threshold. A multiple-choice question distributes probability across the options you provide. If every option is a specialist queue, a message about an unfamiliar issue still receives a top choice. Give general triage an explicit definition and route tickets with conflicting signals or missing evidence to a human. Keep the model’s role limited to assigning a queue and use your existing business rules for actions such as issuing refunds. Google’s guidance on choice questions
Where the API fits alongside rules and chat models
If customers already select a reason from a reliable form, a rule can route many tickets without a model call. Gemini Decision 1 is worth testing where the same issue appears in many forms, such as a short complaint, a call recording or a screenshot. A generative model can then handle work that needs a written response, such as drafting a reply that acknowledges both the missing delivery and the duplicate charge.
That division also applies to AI agents. An application can use a decision model to pick the next step and enforce its own rules about which actions the agent can take. DemystIA’s guide to delegating work to AI agents
What it costs and how to test it
Google lists a free tier and a paid rate of $0.10 per million input tokens, with zero output tokens billed. The shared material is counted once per request, and each question adds tokens for its instructions and criteria. At one call per ticket and an average of 2,500 billable input tokens, 100,000 tickets would cost $25 in Gemini API usage, before retries or other processing. Google reported 150–250 ms end-to-end latency in its own tests on inputs up to 16K tokens. Measure latency on your attachments and request sizes before using that range for planning. Google’s launch announcement · Gemini Developer API pricing
Start by replaying historical tickets with known destinations. Tune thresholds on one set and evaluate them on a separate set that includes ambiguous requests and unusual attachments. Compare with your existing routing rules. Track how many tickets the model can route automatically, how often those routes are correct and how often it misses issues that need urgent review. Then run it alongside the live queue without changing assignments and compare its choices with the support team’s. That gives you a way to judge whether the time saved on straightforward cases outweighs the work of reviewing exceptions.