A plain script is the right tool when the rule is exact, like "amount over 5,000" or "field is empty". An AI decision model fits when the answer needs judgment but is still a yes/no or one of a few options, and a full AI agent is worth paying for only when the output itself is text or the task takes several steps.
Google just entered that middle category with Gemini Decision 1, as Simon Willison reports. It answers three kinds of questions: yes/no, pick from a list, and score against a rubric. It costs 10 cents per million input tokens with free output, the same as OpenAI's GPT-6 Luna decision model. Jev is cheaper at 4.2 cents but handles only text. Gemini takes text, images, audio, video and PDFs, and returns a probability instead of a flat 1 or 0.
What this changes for small business automation
Lead scoring, routing support tickets to the right person, and checking whether an invoice scan has the required fields don't need a generated paragraph. The probability is the useful part. Above a threshold the item goes through automatically, and uncertain cases go to a human. The same pattern fits PDF screening like in a tender analysis pipeline I built, where "is there a penalty clause" is a yes/no question. One caveat: Simon tested four questions on one video, so run about a hundred of your own real examples before you trust any threshold.
Why it matters: Before building an agent, check whether the task is really a choice. If it is, one decision model call plus a few lines of code is cheaper and much easier to test.
Source: Simon Willison