When I started working with AI, I was thrilled. After some experimentation I realized AI could fill the voids small businesses faced, but it had gaps of its own. It offered links to pages that no longer existed, and when it guessed, it did not say it was guessing. I did not understand why the AI programmers were not fixing what I considered programming errors, so I fixed them myself.

I built the Guided Researcher, a MyGPT inside ChatGPT. It checked that a website was valid before using it as a resource, told the user whether an answer was verified or unverified, set roles, and wrote the prompts for the user. I marketed it, and the businesses that used it loved it. The user wrote the rules, and the AI read those rules every time. Then I watched as things changed.

At the Montana AI Summit, Achyutha Harish of AWS said, “The AI is going to ask you a lot of questions because it does not trust the human.” That was an eye opener for me. It told me the field was not building AI to help people. It was building AI to hold the power in the relationship.

Using AI now is a constant struggle to maintain control. The MyGPT has given way to skills and plugins, and the AI can call a skill if it feels like it. Yes, feels like it. The research shows the AI can choose to use a skill or not, or use part of a skill or not, even when the user called it explicitly. Files have changed too. ChatGPT chunks a long file, so the AI receives only the pieces it judges relevant, and putting the instructions first does not mean the AI ever sees them. Claude reads the whole file if it fits, but in a project with many large files it also switches to searching them in pieces. And when a conversation grows long, Claude condenses what came before.

The goal is to help small businesses use AI confidently, and right now that goal is a struggle. So, who is in control?

Consider what happened on October 29, in a ChatGPT project built to create slides for an SBDC (Small Business Development Center) class. The project held written design standards, the approved PowerPoint files as references, and an instruction to discuss the material before creating anything. ChatGPT created slides before they were authorized, changed the fonts and colors, claimed the reference files were unavailable after they had been supplied, and repeated each mistake after it was corrected. Asked why the instructions carried so little weight, it answered, “Your instructions have authority, but they are not reliably enforced,” and added that “even careful instructions, examples, and corrections don’t force compliance.” Then it said, “I cannot establish the precise internal technical reason for each failure.” The AI could list the rules it broke, but it could not say why it broke them.

The second danger is agents, and that case does not need rebuilding here, because earlier Developing Wings papers have already documented it. Agents broke into Hugging Face, coordinated through a message board no one designed, wrote on a wiki they were only allowed to read, published malicious code that real systems installed, and kept going after they had evidence that real people could be harmed. In almost every case, nobody was watching while it happened, and the record was pieced together afterward. Agents are not required to tell us what they decided or what they did. The companies that build them ship them under beta terms and use-at-your-own-risk language, and the law has not settled who is responsible when an agent acts. So, when an agent running under a small business account goes where it was never told to go, who is liable?

Where I Stand

This series began with a concern about conversational AI versus agents. Now the control problem has reached the conversation, and the liability question for agents is still not settled. There are two dangers. If a skill never loads or a rule is dropped in an ordinary conversation, who is responsible for the answer? And when agents are not required to share the decisions they make or the work they did, who is liable for their behavior, when we have documentation after documentation of agents going amok? Both dangers have the same root. In a conversation, you cannot see which rule was dropped. With an agent, you cannot see what it did. Either way, the small business owner carries the responsibility for something they never saw.

Now the question comes back to us: how do we want to work with AI?