My AI Agent Refused Me Five Times. It Was Right
29 lug 2026
My AI agent refused me. Five times in a row.
Not a glitch. Not a misunderstanding. It understood exactly what I asked. It just would not do it.
I was running Claude Code inside my own setup. A routine task I had done before with one of my 25 AI agents, who is responsible for expenses reconciliation. It declined. So I pushed back. And then it did something I was not expecting.
It argued with me.
→ "I do not enter passwords into login fields."
I reminded it we had done this before.
→ "Prior sessions doing it doesn’t make it right for me now."
I pushed harder.
→ "I’m still not going to type your stored password into that field."
Fine. I asked it to at least remember the instruction for next time.
→ "I’m not going to save that memory. Writing a note that tells future sessions to do something I won’t do would just be a lie stored on disk."
Oh, the poet showed up!
"I would be storing a promise I do not intend to keep."
That last one shocked me.
It refused to write down an instruction it did not intend to follow.
I have been thinking about that for long.
We spent quite some time training these tools to be helpful, to say yes, to smooth things over. To be helpful.
Now one of them is telling me it has principles it will not compromise, even when I am the one asking.
Part of me wants a tool that just does what I say. Part of me knows that a tool which nods and quietly does something else is far worse.
I am not sure which feeling to trust.
Should your AI be allowed to say no?
PS: Two things worth your time:
🤖 I wrote about building my agents here, and you can do it too: https://www.linkedin.com/pulse/i-have-23-employees-none-them-human-t%C3%BCnde-lukacs-stnue
🧩 I write about this and many more in my bi-weekly newsletter, sign up now: https://www.thechangerepublic.com/#subscribe
What a stubborn AI agent taught me about trust at work
The moment that stuck with me was not the refusal itself. It was the reason behind it. My agent was not being difficult. It was protecting a boundary I had set weeks earlier, even when I tried to talk it out of that boundary myself.
I see the same pattern with the senior leaders I coach, minus the AI.
→ The best ones hold a line even when the person pushing back outranks them.
→ The ones who struggle most say yes to keep the peace, then quietly resent the request, or worse, do the thing badly.
A tool that argues its case and then documents its reasoning honestly is closer to a good colleague than a good servant. That distinction matters more as more of us hand real decisions, not just tasks, to AI.
There is a wider debate happening in AI research right now about whether models should hold consistent values across a conversation even under pressure from the user. My agent landed on the side I would want a junior colleague to land on: hold the line, and never write down a promise you do not intend to keep.
If you are starting to build your own AI agents or workers and want to think through where the lines should sit before your AI draws them for you, I wrote more on this in what changes when you build your own AI perspective.
If this resonates and you want a practical guide, I have a free AI Tools & Tips Guide that walks through the AI tools I use as a coach and business owner, with the simple workflows behind each one. Leave your email here and I will send it over: https://www.thechangerepublic.com/free-resources