AI
AI in Practice: The Discipline Nobody Puts in the Demo
It wasn’t the model that made my team faster. It was learning where to push back and where to talk instead of type.
Cliff Simon
September 23, 2026
I spent an hour last summer with RevOps and finance leaders trading AI wins and screwups. Nobody was debating whether AI works. We had all seen it firsthand. But what we were stuck on was how to prompt it well, how to keep a team from doing something dumb with it, and how to know when to trust what it hands back.
Interestingly, none of these are model problems. They’re all discipline problems.
Talk to AI Like a Person, Not Type at It
I’m a lazy typer. I always have been. For months I was getting mediocre output from AI and assumed that was just the ceiling. Then I started using a voice-to-text tool called Wispr Flow and the quality of what came back changed overnight. Not because the model got smarter. Because I stopped typing three clipped sentences and started rambling the way I’d actually explain a problem to a person with the context, the constraints, the thing I tried last time that didn’t work.
That’s the part most people skip. A typed prompt gets edited down before it’s even finished because typing is slow and everyone wants to get to the point. Talking doesn’t have that filter. You give it more to work with, almost by accident, and more context is exactly what turns a generic answer into a useful one. Another operator on that call said something similar. His best prompts happen on a walk, thinking out loud, with the transcript cleaned up afterward rather than composed from scratch. Different tool, same discovery. The bottleneck was never the model’s capability. It was how much of the actual problem we were handing it.
Guardrails Aren’t Optional
My team isn’t on an enterprise AI plan. We’re on a team plan because I’m not paying for twenty licenses to cover a much smaller headcount. Without enterprise-grade data controls, I had to get explicit with my team about what goes into a shared AI instance and what doesn’t. I had to keep saying it, not just write it once and assume it stuck.
So every couple of weeks, we go back over it. What not to paste into the company AI account including customer PII, anything with a client’s financials attached, nothing that would be a problem if it showed up somewhere it shouldn’t. What’s fine to ask it to help draft versus what needs a human doing the actual thinking. It’s a standing conversation rather than a document nobody reads, because the real risk builds up in a hundred small decisions your team makes without you in the room.
This is the same governance problem that shows up at the infrastructure layer, just closer to home. A sales leader on the call made a point that stuck with me. A lot of companies are wiring up AI agents to their systems with no filtering on what those agents can actually see and calling the connection itself a finished product. An unfiltered pipe into your CRM or your drive is just plumbing, not governance. Whether that ungoverned access sits at the system level or at the level of what your own team pastes into a chat window, it’s the same failure with a different address.
Trust But Verify
The best stories on that call weren’t the wins. They were the mistakes because everyone had one and they were all the same shape, typically involving an output that looked completely credible and was quietly wrong.
One of the finance leaders on the call had built an automated research process that flagged which investors in a portfolio were actively using his product based on matching email domains. It came back showing a major investor with the highest adoption rate he’d ever seen. Turns out the tool had counted everyone’s personal Gmail address as a company domain and nobody there was actually a customer. Another executive on the call had something worse happen inside the business. A finance report built on AI-generated queries had been wrong for months because the underlying tool was silently capping results at a hundred rows and nobody had told it or noticed.
Neither failure was really about a bad model. Both were about someone trusting a confident answer without checking where it came from. That’s the actual skill nobody’s teaching fast enough. Not how to write a better prompt, but how to stay a little suspicious of a good-looking answer until you know what it’s built on.
And that skill has to be permanent, not a phase you grow out of. Every person on that call agreed the tools are getting better by the week, not the year, which sounds like a reason to relax the verification habit and is actually the opposite. A model that’s wrong in an obvious way is easy to catch. A model that’s wrong in a way that looks exactly like being right is the one that costs you a board meeting, and those get harder to spot as the output gets more polished, not easier.
None of this was really about the technology. It’s about what the people around it are doing differently like talking to it with enough context to get something useful back, telling a team clearly and repeatedly what it shouldn’t be doing, and staying skeptical enough to catch the answer that looks right but isn’t. Get those three things wrong and a smarter model won’t save you. Get these three things right and the model you already have is more than enough.
More signals
Subscribe