Google's New Decision API Is Built to Choose, Not Chat. Here's Where It Could Help.
Let a model sort the small decisions. Keep the authority to act somewhere else.
An AI assistant does not need to write a paragraph every time it makes a small choice. Sometimes the useful answer is simply which queue a request belongs in, whether an attachment needs attention, or which tool should handle the next step.

Google's new Gemini Decisions API targets those bounded tasks. Released on November 11, it uses Gemini Decision 1 to score predefined options rather than generate an open-ended response. It accepts text and several media formats, supports up to 1,000 questions against a shared input, and returns probabilities and confidence scores. Google's announcement
For people building automations, the opportunity is straightforward: separate the small sorting decisions from the expensive work of composing, researching or acting. But a faster decision does not automatically make a workflow safer.
Start with a narrow job
Consider a small business receiving a mixture of routine questions, broken links and requests that need personal attention. A useful first experiment would be to suggest a category for each message, while leaving the existing process unchanged.
That is deliberately less ambitious than letting the model send replies or issue refunds. It gives you a way to compare its choices with yours before mistakes become external actions.
The question should be narrow enough that two people could usually agree on the answer. "Is this message about a broken link?" is easier to test than "What should our company do about this customer?"
That distinction matters more than how impressive the demo looks.
Give it permission to be uncertain
Google's documentation says choice questions distribute probability across the options supplied even when none is a good fit. It recommends measures such as an explicit catch-all category, a separate scope check or a confidence threshold. The guide also says to choose thresholds using labeled examples from your own traffic and consider the different costs of false positives and false negatives. Decisions API guide
In our proposed pilot, that means including an "other or needs review" outcome. A model that confidently puts every message somewhere is not necessarily better than one that sends ambiguous work back to a person.
Nor should a confidence number become a permission slip. A system can be confident that a request concerns billing without being authorized to change a bill.
Measure the work saved, not just the API speed
Google advertises paid pricing of $0.10 per million input tokens with no output-token charges. Its latency figures are benchmark conditions, not a promise for every application or input size. Google's announcement
The more useful business measurement is how much review the system actually removes.
For a first trial, keep a small, representative test set with examples you have already categorized. Include ordinary cases, ambiguous wording and messages outside your expected categories. Compare the proposed labels against those answers. Record both mistakes and the cases correctly deferred for review.
Then ask three questions:
- Did it save time after checking and correcting its work?
- Which mistakes would be costly if connected to real actions?
- Is a simple rule or keyword filter already good enough?
A cheaper model call is not a saving if it creates a new cleanup job.
The Butler's take
This is worth investigating when a workflow repeats many small, well-defined judgments. It is less compelling when the real task is understanding an unfamiliar situation or deciding what the options should be.
The sensible starting point is a read-only pilot: make recommendations, compare them with human decisions and keep the fallback visible. Only then consider connecting the result to an action.
Before you connect it: Our seven failure checks for unattended AI agents cover what happens when an otherwise useful workflow loses access, times out or reports a result that does not exist.
This article was researched and drafted with AI assistance using Google's announcement, release notes and documentation, checked on November 12, 2026. The pilot is a proposed approach, not a reported test. We have not independently benchmarked Gemini Decision 1. No human-review claim is made.