Last Friday, I ran a remote client day from Hangzhou and hit one very specific problem
A freelancer friend sent me more than 80 client requests and asked, “Which ones should I take?” I filtered them manually by industry, budget, and deadline. Twenty minutes later, I had still missed two. I then spent more time turning my criteria into a prompt and checking everything again. The whole workflow took nearly an hour. I got stuck here too: the issue wasn’t that the AI wasn’t smart enough. I had mixed up two separate jobs—understanding the problem and making the decision.
This time, I tried a model built to return a decision in one pass
Google released Gemini Decision 1 and the corresponding Gemini Decisions API. It isn’t meant for chatting or drafting long articles. It scores options: it can make a pass-or-fail judgment, select multiple answers, or score against a rubric. Text, images, audio, video, and PDF files can all be inputs. One call can handle up to 1,000 questions and returns confidence scores. Google announced it at AIE CODE 2026, with Sundar Pichai and Logan Kilpatrick presenting. Cloudflare’s Michelle Chen also shared her experience using production models at the same event.
For us, the more practical use case is this: a person first defines the options and criteria, then the model applies fixed rules. It could prioritize consulting leads, score 50 pieces of copy against a brand guide, or extract assessable details from incoming attachments. It fits situations where “the standard already exists, and I just need a fast classification.” It doesn’t replace decisions around client relationships, pricing strategy, or complicated trade-offs.
Replicating it today doesn’t cost much, but the preparation can’t be skipped
The official price is about $0.10 per million input units, with no additional output charge. A typical small-batch decision will probably cost less than $1. The real time cost is organizing the rules. I spent roughly two hours on my first attempt, breaking client requests into three parts: budget fit, delivery time, and risk level. The technical barrier isn’t high, but I still had to connect through Google’s developer console and prepare a small program to submit questions. If code is completely outside my comfort zone, I can ask a technically inclined colleague or service provider to help with that part.
My first step wasn’t calling the model. I opened the Gemini Decisions API documentation and checked whether my account and region had access. Then I used the test page to submit three options, one scoring rubric, and five samples. I’d only move real client data in after the model’s results were broadly consistent with my own judgments. Not everyone needs this either: if I’m handling only a few messages a day, manual sorting may be more reliable, and waiting is perfectly fine.
Here’s how I’d roll it out by stage
Just getting started
If I’m only seeing a few batches of leads each month, I’d start with a spreadsheet and judge them by hand. Once similar requests exceed 30 items, or each batch takes at least 30 minutes, I’d run a small classification experiment. Otherwise, I’d be automating for automation’s sake.
Working with 1-2 clients
If I already have regular clients, I’d turn the delivery standards we’ve agreed on into a scoring rubric. The model could handle the initial screen, while I keep the final review. I’d begin with low-risk tasks, such as checking assets for compliance or sorting support tickets. I wouldn’t let it decide whether to renew a contract.
Scaling up
If the team handles large volumes of inquiries, resumes, or content reviews each day, I’d keep manual spot checks and record every misclassification. Once performance has been stable for two weeks, I’d consider processing larger batches. I’d also retain the original materials and decision rationale so every result can be traced. AI can shorten the time I spend on repetitive screening, but client promises and final responsibility still belong with a person.