Can Gemini Decisions API help sort your enquiries?
The inbox is full. Your team reads each message to decide who should handle it, then the customer can get help. Google’s new Gemini Decisions API is worth considering if that sorting takes time every day. Start with suggested labels for a person to review. If a short form or a simple rule already gets requests to the right place, keep it.
Sort the request before answering it
Google announced the Gemini Decisions API on November 11. It evaluates questions with defined answers: yes or no, a choice from a list, or a score against criteria you define. It can assess text, images, audio, video and PDFs. It returns scores rather than composing a customer reply.
TLDR AI’s November 12 edition brought the launch to our attention; The Rundown AI also linked it that day. For an owner, the opportunity is fewer interruptions while messages reach the right person.
Consider it when people describe the same kinds of requests in different words. A message mentioning a new delivery address may need someone who handles an existing booking, even if it never uses the word “booking”. If your team cannot agree where the message belongs, settle that first.
Google’s November 11 announcementDiscovery: TLDR AI, November 12Also covered: The Rundown AI, November 12
An illustrative example: one message, two jobs
Imagine an event-hire business receiving this invented message: “Please move tomorrow’s chair delivery to the side entrance. Can you also quote for twenty more chairs next month?” This is an illustrative example, not a client case or a test result.
Sending this message only to the delivery team or only to the quoting team leaves one job waiting. A proposed workflow could instead ask two separate questions: does the message request a change to an existing delivery, and does it request a new quote? The delivery request goes to the person arranging tomorrow’s delivery. The quote request becomes a separate task.
Those labels do not confirm that the entrance is accessible, that stock is available or that the customer has accepted a price. They tell people what needs attention. Keep the original message attached so neither person has to reconstruct the request.
Google’s documentation says questions in one request are evaluated independently. Choice questions also distribute probability among the options even when none fits. Include an explicit “other” route and send uncertain or unrecognised requests to a person.
Google’s decision types, independent questions and fallback guidance
Judge the trial by the work left over
The confidence score estimates whether the selected answer is correct. It is not a measured accuracy rate for your business. Google recommends setting thresholds with labelled examples and weighing the cost of different mistakes.
Begin with clearly synthetic examples covering ordinary requests, mixed requests, missing details and messages outside your services. Agree the expected destination before running them. This is a proposed test, not one we have performed.
Passing these examples would only prepare you for a supervised trial. The useful result is less sorting work after you include the time spent correcting mistakes.
- Count messages that reach the right person without a correction, alongside those that are sent to the wrong place.
- Check mixed requests separately: a correct label for one need can hide a missed second job.
- Keep uncertain cases visible in a review list, with someone responsible for clearing it.
- Compare the time spent checking and correcting labels with the sorting time you hoped to remove.
What you would actually be buying
This is a developer API, a way for software to call the model. It does not automatically add sorting to your inbox. Someone must connect the messages, categories and destination tools. Google lists it on the Gemini Developer API using the v1beta interface; its launch post says Google Cloud support is still coming.
At publication, Google lists a free tier and a paid rate of US$0.10 per million input tokens, the units used to measure the material processed. It bills no output tokens. Shared input is charged once within a request; each question adds its own instructions. The cost of connecting your tools, monitoring the workflow and checking results is separate.
Check the project’s active rate limits in AI Studio and confirm that your intended connection supports the new endpoint before commissioning it.
Current Gemini Decision 1 pricingAPI availability and limitationsCheck project rate limits
When waiting is the better choice
For a small inbox, sorting by hand may be enough. Wait if categories change constantly or the same person will read every message anyway. A clearer enquiry form may save more effort than adding another connection.
If you are unsure whether the task needs AI at all, start with our guide to choosing between AI and ordinary automation.
Sources
Choose one sorting decision to improve.
Describe the categories and show forvertz a made-up example of an awkward request. We can help scope a useful first workflow without using private customer records.
Discuss your enquiry workflowUpdated