AI without the theater.

AI is easy to get excited about. Turning that excitement into a number that moved is harder. Many businesses are still figuring out how to close that gap. The work takes more operational judgment than technical skill.

What theater looks like

The signs are familiar: a chatbot the team ignores, a pilot that never ships, an "AI-powered" announcement with no measure behind it, or a tool bought before anyone named the job.

Picture an owner buying an AI receptionist after a polished demo while insurance paperwork keeps piling up in a queue. The tool may work, but it solves the wrong problem. It soon becomes another subscription because nobody named the job first.

A good exchange with ChatGPT or a polished online demo shows only a small part of what focused use looks like. It does not tell an owner which workflow is worth changing or how to judge the result.

Start with the operation

Begin with the recurring work that costs the business time or revenue. Once the job is clear, you can judge whether AI belongs in the solution and what the result needs to be.

Four questions before you build or buy anything

We run our ideas and vendor proposals through the same filter.

1. Is it a real, repeating workflow? A workflow that runs fifty times a week has room to repay the work of changing it.
2. Does it move a number? Name the revenue recovered or staff time saved before the project starts. Without a measure, you cannot tell whether it worked.
3. Is a human in the loop where the decision matters? Models can draft and prepare the work. A person should approve decisions that carry real consequences until the system has proved itself.
4. Can you turn it off? You need a clear off switch and a way to return to the old process.

The first two questions rule out most weak ideas.

Split the work deliberately

Models are useful for reading messy language, drafting a first version, and pulling facts from documents. Plain software should handle fixed rules, calculations, dates, and the steps that must run the same way every time. A reliable system usually combines a few focused model calls with ordinary software.

Building custom software around a model is faster and more affordable than it was two years ago. The feedback loop can take hours rather than quarters. The scarce part is the experience to build a reliable system and the judgment to know where the model should stop.

What this looks like in practice

Two years ago I barely touched AI. Today I use it to build, plan, and operate. rafl, the NFC tap-to-enter giveaway platform I have been building for two years, has six apps run by two people. I use the same approach in my day job: models for the language work and ordinary software for the rest.

Start with one workflow

The workflow is the strategy. Find your most expensive repeating workflow and improve it before buying a general AI tool.