Most companies that say they are AI just wrap ChatGPT, and you can do the same thing yourself.

Let me say the quiet part, because it changes how you should buy. Very many companies that say we are AI just build a wrapper around ChatGPT and call it AI. Under the hood everything flies to ChatGPT as-is, at best with a system prompt written in front of it. There is no magic. Which means you can often go and do the same thing yourself in your own ChatGPT, and the useful move is to figure out where that is true for you and just do it.

Take product descriptions. A proper PIM, product information management, easily turns into a fifty thousand dollar project, and most clients do not want that. But you can point ChatGPT at a column and have it write descriptions for the items listed there, which is effectively porting some of a PIM's functions into the thing without ever standing up a PIM. With a tool like n8n you drop in a file, run your GPT across the product-title column, pull out what is missing, and feed it more and more data over time so it gets sharper.

The same pattern scales into a real internal tool. Notion AI is essentially our internal GPT. I ask it something and it pulls the answer from across everything I have put in, to the point that from all the RFPs we have participated in, which I have described in there, I can ask it to compose a new one, and the more of them there are the more accurately it generates the next. It also has descriptions of every call I have had for roughly the last two years, so this thing will soon know more about me than I know myself. You can build a version of that yourself, just without leaning on the hosted AI element, by piping GPT in: at search time the front end returns the Elasticsearch index up top and GPT writes something underneath. And the second half of a product like that is connecting a chatbot, probably the same GPT, to your own parsed database, so when it does not know an answer it reaches into your data. That is retrieval, not magic.

We tried this thinking with Clay for working new markets faster, and the realization was simple: in Clay we were going to plug in the same OpenAI key anyway, so why not just try it directly in our own ChatGPT. Hand it a URL, have it describe the site, decide is this B2B or B2C, is this company interesting to us given what we do. Clay has huge potential, but you have to learn it and set it up, and if the cheaper path gets you most of the way, take the cheaper path.

Now the line I will not cross. Anything to do with AI, my first question is always, is it accurate, is it consistent, and the honest answer is almost always no. That is why you can train your own GPT on a defined dataset to push it toward accuracy, and why you check it rather than trust it. My own ChatGPT usage has grown several hundred percent in a couple of months, I google less because Google gives you links and GPT gives you an answer, and I have custom instructions in there so it answers the way I want. But the value is never in saying we are AI. It is in knowing exactly which boring task you can hand to a model today, and whether its output is consistent enough to actually use.