Here’s something nobody tells you when you’re buying an AI tool: the technology is probably fine. The question worth asking, before you sign anything, is whether you understand your own processes well enough to hand them to something that cannot guess.
A while back, we had a gap in our development process. Not a dramatic one. We weren’t missing entire phases or skipping discovery. We had a solid process for creating user stories through requirements gathering - use cases accounted for, scope defined, everyone on the same page. On paper, it looked complete.
The gap was in the acceptance criteria written against those stories.
It had always been there, quietly. Acceptance criteria was treated as an afterthought, and the cost showed up everywhere downstream. Developers were making judgment calls about whether a story was actually done. Testers were going back and forth trying to determine whether something passed or failed. Everyone was compensating, everyone was working harder than they needed to, because one piece of the process had never been formally defined.
We didn’t notice it as a gap because our team absorbed it. Good people do that. They adapt and fill the silence without even realizing they’re doing it. The process looked like it was working because the humans inside it were quietly holding it together.
That’s the thing about undocumented process gaps. They’re invisible right up until they’re not.
Now imagine pointing an AI agent at that same workflow. There are no judgment calls available to it. It cannot intuit what “complete” means when the acceptance criteria is vague, and it has no one down the hall to ask. The process needs to be articulated end to end, consistently, every time, and every crack your team has quietly papered over will show up.
This isn’t a reason to avoid AI. It’s a reason to treat an AI implementation as a process audit in disguise.
Three questions to ask before you automate
Is the tool capable?
Almost certainly yes. This is rarely the real issue. Most modern AI tools are capable enough for the task you have in mind.
Can you articulate exactly what you’re asking it to do?
Not roughly. Not in the way your team has always understood it. In writing, with no gaps, no assumptions that someone will just figure it out.
Is the process actually finished?
If the answer is no, the implementation won’t fail because the AI wasn’t good enough. It’ll fail because the process was never complete to begin with.
If you can’t answer question two with confidence, treat the AI implementation as a process audit first. Define what done looks like, write it down, and close the gaps your team has been quietly absorbing. Once that’s solid, automating it actually works.