Somebody in your office has already tried it. They opened one of the AI tools, pasted in a proposal or a customer email, and got back something that read better than what they had. That was the week of wow. Then it drifted. Somebody used it for a bid narrative and it made up a spec section. Somebody asked it about a customer and it answered with confidence about a customer it had never heard of. By the end of the month it was a thing two people used for emails, and the platform vendor's new AI feature sat in the menu like the reporting module nobody opens.
Here is the flip: that was not the tool failing. That was the tool doing exactly what it is, in a shop where nobody had been told what it is.
I ended the last post on a question, where does your operation lose the most minutes to a step nobody would miss, and I said finding that step is the work. One thing comes before it. Before you put any AI tool on your work, you have to know what you are putting there. Not how to use it. What it is.
What it actually is
Strip the branding off and every one of these tools is the same kind of thing underneath. It reads text and writes text. It has read more than any person alive, and it knows nothing about your business unless something hands it that, every time you use it. It does not know your crews, your pricing, which customer pays late, or that the Henderson job is a week behind. It does not remember Tuesday on Wednesday unless something was built to carry the memory over. And it does not know when it is wrong. It writes the wrong answer in the same voice as the right one, because to the tool there is no difference. It is finishing a sentence either way.
Everything you were sold on top of that, the AI feature in your platform, the assistant in your estimating software, the chat window with your logo on it, is a wrapper around that same thing. The wrapper decides what the tool gets to see, what it keeps, where your information goes, and what it is allowed to say when it has nothing to say. The models are the same ones I use. What differs, shop to shop, is what got built around them.
The guy you already know
You have worked with this tool before. It was a person. The estimator who answers every question in the same confident voice whether he has looked it up or not. You learned to run that person. You did not ask him what was true. You handed him the drawing and asked him what the drawing said.
I have been that person. Writing the first post on this site, I went to cite a number I had carried around for a long time, that six in ten software rollouts at small companies fail. It does not exist. Not a study, not a survey. A sentence I had absorbed and started saying in the voice I use for things I have measured. The tool does that at scale and at speed. Once you understand it is that guy, you already know how to run it: hand it the file, ask it what the file says, and check the answer against something you know cold.
Read the live system, not the sales page
The rule I hold before any tool goes on the work is short. Ground truth first: understand what the tool actually is and does, from the running thing, before you trust any document about it. The demo is a document. The feature page is a document. Software people call the same rule observability: believe what the running system reports over what the design said it would do. Operators have an older version. You do not put a new machine on a job off the brochure. You read the plate, you know what it is rated for, and you know who is allowed on it.
Three questions do most of that work for an AI tool, and none of them is technical.
What does it see? When your estimator asks it about a bid, what was actually handed to it, the package or the question? If the answer is that it just knows, it does not.
What does it keep? Does it remember your last conversation, and if so, where does that memory live and who owns it? Same answer as the first post on this site. Own it, do not rent it.
What does it do when it does not know? Put it on a thing you know cold, with a hole in it, and watch. If it never once says it cannot tell, it is guessing, in the voice it uses for facts.
That last one is the gate I hold my own builds to. I built a tool that reads bid packages, and before it shipped I put a package through it that ran to over two thousand pages. It read the whole thing, and it named the two pages it could not read. That was the last gate, because it is the hardest thing to get the tool to do and the only thing that makes the rest of the read worth trusting. A tool that says what it could not see is worth more than one that flatters you.
Where the structure comes in
The week of wow ends the way it does because nothing was built around the tool. Nobody handed it the file, so it made one up. Nothing carried Tuesday into Wednesday, so every conversation started from zero. Nothing told it what to do when it did not know, so it did what it does. That is not a tool problem. It is structure, and structure is what the rest of this series is about: what gets settled first, what compounds, what your people need to understand instead of copy.
How I run that read on a specific operation is the work, and it is the part an engagement earns. I am not hiding it. I am telling you where it lives.
Where this leaves you
I am not telling you to stop using it, and I am not telling you to trust it. I am telling you it is a specific kind of thing, and once you know what kind, most of the decisions about it get easier, because you have made them before about people.
So before the first build, one read. When your people use it, what does it actually see, and what did you assume it already knew? If the honest answer is that you assumed it knew your business, that is not a failure. It is the most common assumption in the industry right now, and the first thing a build has to fix. The next post in this series is about what to settle once you have that read, and in what order.