On a Monday afternoon, I almost used AI the wrong way

Last Monday afternoon, I had coffee near Oakland’s Union Square with a restaurant consultant. She had received a notice saying scheduling and customer-service scripts might soon be handled by AI, with managers able to view chat summaries. Her first reaction wasn’t “Is this efficient?” but “Will customer complaints get blamed on me?” I’ve gotten stuck on the same kind of issue. While building a personal brand, I once used AI to summarize customer interviews. The original quotes and my assumptions got mixed together. Nothing went wrong, but the customer nearly felt misrepresented.

They skipped the AI debate and negotiated protections first

Restaurant workers in Oakland negotiated for clear limits: companies must explain what AI will and won’t do, while preserving a human appeal and correction process. Consultant Maria Sanchez later told me what happened. Sitting at a folding table beside the restaurant kitchen, she helped write protections into the agreement, including “no automatic employee penalties” and “sensitive content requires human confirmation.” She wasn’t a technical lead, but she joined the discussion. Solo studios and small teams can borrow the same approach: agree upfront on how data will be used, which decisions need human confirmation, and who fixes AI mistakes.

Replicating this today costs less than I expected

There’s no software to buy. The budget is $0, preparation takes about 60 minutes, and the technical barrier is simply being able to explain the work boundary clearly in writing. The first step is to open the AI tool I already use, create an empty document titled “AI Usage Boundaries,” and write three sentences: what I’ll let AI handle, what it must not handle, and who confirms the result. I’d show it to one client or colleague, then revise it based on their feedback. Not everyone needs this; if I’m not using AI yet, waiting is fine too.

Match the agreement to my stage

When I’m just starting out, I’d begin with one page focused on not pasting private client data directly into a tool. With one or two clients, I’d add “all external content must be reviewed by a human” and “source material is not automatically deleted.” While scaling, I’d also document version history, name the human reviewer, and create an appeal channel. There’s no need to turn it into a complex policy overnight. I’d start with the step most likely to go wrong in real work. I’m still filling in the gaps myself: when a boundary is unclear, I’d rather ask one extra question than save ten minutes and lose a night worrying about it.