AI is good at a narrow judgment: does this look like a rush job, a warranty, a no-show risk. It is a bad place to put the actual decision. Quotes, refunds, and scheduling should land in deterministic rules you can read, test, and change without hoping the model feels the same tomorrow.
That split is showing up in the tools themselves. Laravel News put it plainly in the Decide with Jev series this month: the model can answer questions about a draft, but “the model should not pick what the app does next. Our code should.” In that episode, a small PHP class reads the AI’s answers and returns one label: ready for editor, revise, or needs review. None of those labels publish the draft. Even the best draft still goes to a person.
Artificial Prism’s week of 18–24 October 2026 logged the same idea from two other directions. OpenAI’s Decisions API is generally available. It answers bounded questions from text, images, or JSON, which is useful for routing and scoring without paying for a full generation. Microsoft launched Guardian Agents to monitor autonomous tools, saying frontier agents cannot be fully trusted. Satya Nadella urged separating intelligence providers from the authority to act.
Intelligence in one place. Authority in another. That is the design.
What this looks like on a real job
A homeowner in Mooresville uploads a photo and asks for a roof quote. The model can answer: material looks like architectural shingle, pitch looks steep, damage is visible along the valley. Those are judgments.
Your rules make the call: if confidence is below your cutoff, no number goes out, schedule a visit. If the zip is outside your Lake Norman coverage list, send the polite decline. If the job type matches last year’s comparable jobs, apply your markup table and email a range. The markup table is not a prompt. It is a table.
A refund request hits a Statesville retailer at 9 PM. The model can answer: the customer is describing a sizing issue, the tone is frustrated, the order date is in the message. Your rules make the call: orders under 30 days get store credit, over 30 days need a manager, anything over $500 needs a person. Laravel’s AI SDK 1.1, covered in the same October Laravel News archive, even treats a refund policy as a skill the agent loads on demand. The policy is still yours. The agent does not invent it.
A clinic in Cornelius gets a cancellation text. The model can answer: this reads as a same-day cancel. Your rules make the call: same-day under 4 hours triggers the no-show fee you already published, or it does not, depending on the policy you wrote. The model does not get to feel generous.
Why owners get this backwards
It is tempting to ask the agent “what should we do?” because that is how chat works. Chat is a terrible operations manual.
The model will be slightly different next month. Your refund window will not, unless you change it. If the quote, the credit, or the calendar slot is buried in generated text, you cannot test it, you cannot show it to your insurance person, and you cannot explain it when a customer in Troutman asks why their neighbor got a different number.
Laravel News’s Jev episode also flags two traps that show up in business rules: a missing answer is not the same as “no,” and a confidence of zero is not the same as “we have no confidence score.” Your code has to treat those as different paths. That is boring. It is also how you avoid refunding someone because the model shrugged.
How to split the work in your own shop
Write the questions the AI is allowed to answer. Keep them bounded. Yes/no, a category, a score, a short summary. OpenAI’s Decisions API is built for that shape: questions over text, images, or JSON, including many in one request.
Write the rules that consume those answers. If category is emergency and zip is in service area, text the on-call tech. If score is below 0.85, a person reviews. If the draft is “ready,” it still does not send until someone hits send, unless you have a written exception for a job type you fully trust.
Keep the rules in the application, next to the data. That is why we keep putting this in Laravel development rather than a pile of zaps. AI engineering here means the agent is a worker with a job description, not the owner of the checkbook.
AI optimization does not change this. Showing up in AI answers is marketing. The call on a quote is operations.
A one-page test
Take last week’s real inquiries. For each one, write: what question the model should have answered, and what your rule should have done. If you cannot write the rule in one sentence, the agent is not ready to act. It can still draft.
systemsevendesigns builds that split for owners around Statesville, Mooresville, and Lake Norman: AI on the judgment, your rules on the decision, a person on anything that moves money or a calendar. If you want to walk through quotes, refunds, or after-hours booking, bring the policy you already use. We will put it in code instead of a prompt.