Everyone has heard “don’t automate a bad process.”
If your deployment process is manual, undocumented, and requires three people to remember the secret handshake, putting an AI agent in front of it doesn’t fix it. It just makes the bad process run faster and fail in new ways.
We all know that.
What we miss is this: when you give AI a goal without a process, it will invent its own process. And that process is probably bad.
I see this constantly with AI coding agents. You tell it “add authentication to this app” or “migrate this database” or “automate these incident reports.” You didn’t give it a process. You gave it an outcome. So it makes up the steps:
- What to check first
- What order to do things in
- What counts as “good enough”
- When to stop and ask vs. when to keep going
- How to validate that it actually worked
That invented process is trained on the average of the internet. It doesn’t know your constraints, your risk tolerance, your compliance requirements, or how your team actually operates. It doesn’t have to live with the consequences next Tuesday when it breaks.
AI can help you design a process. That’s useful. Let it suggest one. Let it poke holes in yours.
But you have to own the process.
In practice, here is what works:
1. Start with a process you own. Either write it yourself, or take an existing good process your team already uses. Be explicit. Steps, inputs, outputs, validation, rollback.
2. Let AI critique it. Ask “what am I missing? What would fail? How would you make this more resilient?” Some suggestions will be good. Some will be bad. You decide.
3. Lock it in. Once you have settled on the process, write it down where both you and the agent can see it. Checklist, doc, CLAUDE.md, runbook, whatever.
4. Make AI follow it to the letter. This is the harnessing part. Don’t let the agent freelancers wander off because it found a “better way.” If it wants to deviate, it should propose a change to the process and get your sign-off. Otherwise, it follows the process exactly.
The new failure mode isn’t “we automated a bad process.” It’s “we let AI quietly invent a bad process and then we automated that.”
If you don’t define the process, AI will. And you will end up debugging a workflow you never approved, built on assumptions you never validated.
Decide the process first. Then harness the AI to execute it.