AI Can Write Your Code, It Cannot Read Your Business

A founder showed us a feature last month that an AI coding assistant had written in an afternoon. It worked. It handled the happy path, the types checked out, and the UI looked clean. Then we asked what happens when a customer cancels a subscription mid cycle and expects a prorated refund instead of a full one. The AI had picked the more common pattern, full refund, because that is what most of its training data probably looked like. Nobody had told it the business only prorates after day ten. Nobody could have, because the assistant never asked.

That is the part of "AI writes code now" that gets skipped in the demos.

The model doesn't know your rules, it knows the average

Large language models are extremely good at producing code that matches the shape of code they have seen before. Auth flows, CRUD endpoints, form validation, checkout pages, they have seen thousands of each. What they have not seen is your specific pricing logic, your specific compliance requirement, or the one weird exception your ops team carved out eighteen months ago because a client threatened to leave. That knowledge lives in your head, in a Slack thread from last year, or in nobody's head at all anymore. The AI fills the gap with whatever is statistically normal, and normal is not the same as correct for your business.

This is why AI generated code so often looks finished and isn't. It passes review at a glance because the syntax is clean and the logic reads sensibly. The bug is not a typo you can spot, it is a business assumption baked in quietly, and it usually surfaces three weeks later as a support ticket or a chargeback.

Where this actually bites

A few patterns we keep running into with clients who lean hard on AI for shipping speed:

• Discount and refund logic that follows the "standard" rule instead of the one your contracts actually promise
• Permission checks that cover the obvious roles and quietly miss the one edge role finance insisted on
• Data validation that accepts formats your actual customers never use, and rejects the messy real world input they do send

None of this means don't use the tools. It means the review step matters more, not less, as code gets easier to produce. The bottleneck was never typing speed. It was always someone who understands the business checking that the logic matches reality before it ships to a paying customer.

What actually works

Treat AI output the way you'd treat a smart junior contractor who just joined and has never sat in a single meeting with your ops team. Fast, capable, occasionally very wrong about the one thing that actually matters, and cheerful about it because it has no idea it's wrong. Someone still has to own the business logic, write it down, and check the generated code against it line by line where it counts, not just skim for style.

That's the work we actually get hired for at Cedrix these days: not writing boilerplate, but making sure the software matches how the business really runs, including the exceptions nobody documented. If you want to see what that looks like on your own product, we build a working free MVP so you can watch the process before you commit to anything. Or just get in touch and tell us what broke.