The Trust Battery: why AI rollouts actually fail

Writing this at 5:47am. Cold brew gone warm. Wife and kids are asleep. Standard Sunday energy.

Last week an owner told me, with a straight face, “AI doesn’t work for my business. We tried it.”

I asked him what “tried it” looked like. Six months ago he turned on a tool. Told his team to use it for client emails. Didn’t review the outputs. Didn’t set checkpoints. Three weeks in, the AI generated something that pissed off a client. He pulled the plug. He’s been telling people “AI doesn’t work” ever since.

That’s not an AI problem. That’s a TRUST problem. And there’s a framework for it.

Monday play

Pick one AI workflow already live in your business. Place it on the Trust Battery (below). Be honest about which stage it actually deserves, not the one you wish it was at. If it’s at Pair mode but your team treats it like Autonomous, that’s your problem. Back it down two weeks. Log corrections. Move up only when the data says you can.

Tool worth a look

Memory in Claude and ChatGPT agents. Both shipped persistent memory in the last month. Your AI now remembers context across sessions, which means a workflow you’ve been running for 90 days is smarter than a workflow you started today. If you’ve got an agent that’s been “in pilot” for months, upgrade it. The same agent with memory is a different agent.

Prompt to steal

At the end of every workday, review the outputs you generated today. Rate each one 1-5 on usefulness. Flag any corrections the human made. Propose three updates to your system prompt that would have prevented those corrections. Output: a daily reflection note I can review tomorrow morning.

Drop this in a scheduled task. Your agents get smarter while you sleep.

Mini lesson

Treat every new AI agent like a new hire on day one. Pair mode first. Review every output. Earn trust before you give it the keys. Owners who skip this step are the same owners who tell me six months later “we tried AI and it didn’t work.” The tool didn’t fail. The onboarding did.

Mental model

The Trust Battery. Four stages of AI autonomy:

1. Pair (20–40%): Human reviews every output.
2. Async with checkpoints (40–60%): AI runs but stops for approval at defined points.
3. Audit (60–80%): AI runs end-to-end. Human reviews samples after.
4. Autonomous (80%+): Runs in background. Humans only pinged on exceptions.

Trust charges with clean execution. Drains every time someone has to fix the output.

Steal from a podcast

Builder Nityesh on a recent agent-building pod: “The killer move is a nightly reflection prompt that reviews the day’s outputs, rates quality, and proposes system prompt updates.” Set it to run at midnight. Wake up to a sharper agent. Compound that for 30 days and you’ve got something serious.

The number

20%. The trust level every new AI agent starts at. Not zero. Not 100. Twenty. It earns the rest by executing cleanly.

Heard this week

Anthropic just shipped agent memory to its managed agents (beta). Netflix and Rakuten are already testing it for customer-facing workflows. Persistent-memory agents are about to make every “pilot stuck in pilot” project relevant again. Pull yours off the shelf.

Talk soon.

Kevin