You Can’t Automate a Workflow That Won’t Sit Still

The fastest way to waste money on AI is to automate a process nobody can describe, own, or hold still long enough to improve.

I am working with a business that originally hired me to help build a tool around its advertising work. The idea was to bring ad platforms, copy, and post analytics together, then use that information to recommend useful updates.

That was the first version of the problem. It did not remain the problem for long.

The scope shifted from managing advertising performance to understanding budgets. Then it shifted toward tracking vendor and client performance. It is currently moving again.

Ownership changed too. I was hired by a minority owner who wanted to move quickly and explore what AI could do. Soon after, most of my work moved to the majority owner, who is more hesitant about AI and less willing to spend time defining how it should participate in the work.

The tool was moving because the target was moving

Shortly after that ownership shift, I realized the engagement had developed serious scope creep. I had to sit down with the owner, narrow the work, and put the boundaries in writing.

Eventually, requests reached the point where I regularly had to say, “That is outside this month's scope.”

I thought the conversation might cost me the client. Instead, they kept renewing. That has allowed us to continue making progress, but the instability has had a real cost. I have discarded significant work, implemented features only to remove them, and waited a week or more for decisions needed to continue. Deadlines have remained fluid because the way the owners work is fluid too.

AI can move quickly. That does not help when the business has not decided where it is going.

“What would a good result look like?”

One current request is an AI-assisted end-of-month recommendation analysis. Before building it, I asked a few basic questions:

• What would a good outcome look like?
• Who will use the recommendation?
• What decision should it help that person make?
• How should the information be presented?
• What would make the recommendation useful?

The response so far has been silence.

That silence is not a prompt-engineering problem. It is not a model problem. It is a workflow-definition problem.

If we cannot describe the decision an output must support, we cannot evaluate whether the output is good. If we do not know who owns the process, we cannot create a reliable handoff. If the goal changes every few weeks, each new feature risks becoming expensive rework.

Optimize the work before you accelerate it

A changing business is not automatically a broken business. Priorities move. Owners learn. New information changes the plan. The mistake is pretending a moving process is stable enough to automate.

Before building a tool, adding an agent, or automating a recurring task, the business should be able to answer five questions:

1. What problem are we solving?
2. What result should the work create?
3. Who owns the process and its decisions?
4. What does the workflow actually look like today?
5. How will we know the new version is better?

When those answers are missing, more technology does not create clarity. It operationalizes the ambiguity that is already there.

Sometimes the most valuable AI work is recognizing that the workflow is not ready for AI yet, then helping the business make it ready.