AI Isn't Working Because Your Processes Are Broken
You can't spend five minutes on LinkedIn without seeing "AI experts" giving advice. But here's the reality: most companies are struggling to show actual impact from AI—the kind that moves the metrics that matter. Revenue. Costs. Market share. The stuff that reflects a company moving forward.
The diagnosis isn't complicated.
Set aside IT and legal blockers for a moment (I wrote about security concerns here). The simple reason AI fails is that your processes are broken.
That's it. Your workflows are inefficient, undocumented, or held together by tribal knowledge—and no AI will fix that. Worse, when you ask AI for a "marketing plan" without giving it real context, you get something vanilla, have no idea how to implement it, and then conclude AI doesn't work.
AI didn't fail. The process was never there to begin with.
The One-Sentence Test
When I was a college instructor, I had a simple test for students writing research papers: can you explain your paper in one or two sentences?
The students who could knew their research. They usually made an A. The ones who rambled, overcomplicated, or couldn't articulate the point? They didn't know the material as well as they thought.
Apply this to your business.
For any process—strategy development, campaign execution, monthly reporting—can you clearly explain what it does, how it works, and what the output is in a few sentences? If yes, move on. If no, that's your starting point.
Diagram Before You Automate
Take time to map your process. I prefer a whiteboard or a graph-paper notebook—something visual where you can see the workflow laid out.
Once you diagram it, the holes show themselves. You'll spot the bottleneck, the unclear handoff, the step that only works because one person "just knows" how to do it.
Don't take this lightly. Really think through the entire process.
Another story from my academic days: a professor gave an assignment to document how we would conduct research. I wrote, "I will look up resources in the online library."
He stopped me. "No. Start at the beginning."
He meant the actual beginning. What time do I wake up? How do I get to campus? Where do I sit? He wanted the full sequence—not because those details mattered, but because he was teaching us to think through entire processes instead of skipping to the "real" work.
The same applies in business. You need to understand every step before you can automate any of them.
Why This Matters for AI
AI is excellent at executing processes, streamlining them, and producing results faster than any human. But it requires proper input. You've probably heard this called prompt engineering or context engineering.
Context engineering is underrated. The better your inputs, the better your outputs.
Here's the difference:
Weak input: "Write a monthly report analyzing our sales data."
Strong input: "Write a monthly report for the executive team. Pull data from these three sources. Flag any metric that deviates more than 10% from the prior month. Structure insights as: what happened, why it matters, what we recommend. Tone should be direct, not hedged. Keep it under 1,500 words."
Same task. Wildly different results. And you can't write the second version if you haven't done the work to understand your own process.
Putting It Together
Once you've analyzed a process—diagrammed it, filled the gaps, executed it manually—you're ready to build an agent.
Take your workflow and write it out as a prompt. Longer than you think. Many of my agent prompts run 100-300 lines because I'm documenting the entire process. (Shortcut: photograph your diagrams and ask your AI model to extract them into sequential plain text.)
Review the prompt, submit it, and let the AI do its work.
A Real Example
I inherited a monthly reporting process that took 16 hours. Data came from multiple sources, analysis was manual, and the final output was a narrative report for leadership.
Before building anything, I ran the process manually for three months. I needed to understand every step—what data mattered, what patterns to look for, what leadership actually wanted to see.
Once I had it down, I built an agent in Claude. It pulls the data, surfaces trends, flags anomalies, and drafts the insights. I still validate the inputs and refine the final output, but the middle 80% runs itself.
16 hours became 3. The report now has visibility with leadership across multiple regions, and I've had requests to extend the approach to other teams.
That's the 10-80-10 model: 10% human input, 80% AI execution, 10% human refinement.
The Compound Effect
Once this practice clicks, building agents becomes second nature. Any redundant process is a candidate.
I now have agents for monthly reports, OKRs, annual reviews, email drafting by audience, alt text generation, schema markup, HTML conversion, page audits—the list keeps growing. (I've open-sourced several at elishaconsulting.com/github.)
So before you blame AI, take an honest look at your processes.
Start Here
Pick one process that frustrates you. Then:
Explain it in two sentences. If you can't, you don't understand it well enough to automate it.
Diagram it. Whiteboard, notebook, whatever. Find the gaps.
Run it manually. At least once. Prove it works before you hand it to AI.
Then build the agent. Not before.
AI doesn't fix broken processes. It scales them.
Fix first. Automate second. That's the order that works.
