The risk with AI at work isn't that it's wrong. It's that it's wrong quietly.

Originally posted on LinkedIn, September 2026. First of a four-part series that ends in the Career Companion Hub.

I spent this month building a client-engagement tracker with Claude Code — multi-client, role-scoped, the system of record for how I run advisory work. Partway through, a button stopped responding. No crash. No error. Nothing in the logs.

Two failures were stacked underneath it.

The environment was silently blocking the confirmation dialog that button depended on, so the code was reading "cancel" every time without ever asking me.

And beneath that, the real one: every form in the application was dead. Log time, submit an expense, add a person. All of it rendered perfectly and saved nothing.

That's the failure mode nobody puts in the adoption deck. Not obvious hallucination you can catch on sight. Output that looks finished, survives a glance, and quietly does nothing.

Which changes what you should be measuring.

Most organizations are grading AI on throughput — drafts produced, hours saved, tickets closed. Wrong instrument. The number that matters is where unverified output enters a decision, and who owns it when it turns out hollow.

Three things I run because of that:

Delegate aggressively, verify structurally. The model owns research, drafting, modeling, and build. I own the checks. Here that meant 60 automated tests asserting the one thing I actually cared about — that one client's administrator can never see another client's data.

Specify like a contract, not a wish. Midway in, I found that "start a fresh workspace" was wiping every existing engagement. No model was going to catch that. It required knowing my own operating model: concurrent clients, each with a scoped admin approving their own team's time and expenses.

Put the human signature where the liability sits. Research, synthesis, first drafts — delegate. Entity formation, IP, anything I'd sign my name to — mine.

I came to this from sales, not engineering. What I keep finding is that most failures are systemic and unnamed — a funding structure, a governance gap, an accountability nobody assigned. AI adoption is the same shape of problem wearing newer clothes. Teams buy the model and skip the question of who is accountable for what it produces.

The bottleneck was never the technology. It's whether anyone in the room can specify precisely and distrust a clean-looking result.

Read the discussion on LinkedIn → https://lnkd.in/p/gzQ6mS-S