Automation Before Agents

A well-put formulation I heard recently, from someone who has built more of this than most: don't start with the agent. Start with the automation.

I've been thinking about it since, because it points at something I've seen again and again with clients without being able to put words to it.

Why People Don't Trust Agents

Most people don't trust a free-running AI agent. Not because they're resistant to change. Because they can't control it.

They can't see what it does step by step. They can't verify the result before it's done. They don't know if it made something up halfway through. And they still bear responsibility for what comes out. That's not tech fear. It's a reasonable demand for control from someone who has to stand behind the result.

Give that person an agent that runs on its own, and you haven't given them a tool. You've given them a risk they can't assess.

The Reversed Order

Flip it. First build the stable step -- a process that does one well-defined thing, the same way every time, and that can be verified. Hardcoded if necessary. Deliberately boring. It should be predictable, not smart.

Then let an agent be allowed to start it. Not to improvise it. To start it.

Now you have autonomy where it's harmless -- in the choice of when and what -- and control where it counts -- in the actual execution. The agent makes the decision to run. The machine does the job, the same way every time.

Have a hundred such steps and you've built something powerful without ever asking a person to trust something they can't control.

This Isn't a Contrarian Position

It sounds like a cautious stance in an industry that glorifies autonomy. It isn't. It's consistent with how the more thoughtful vendors draw the line: separate the predetermined flow from the free, self-directing agent. Use the free sparingly, where failure is tolerable.

What's provocative isn't the conclusion. What's provocative is how many people skip it, build the impressive autonomous thing first, and then wonder why no one in the organisation dares to use it.

Control is not the opposite of AI. Control is what makes someone willing to let AI in at all.

See also: It's Not the Speed That Kills You (series 26) and Trust Comes Before the Technology (series 27)