AI Is Not Magic. It Reflects Discipline — or Lack Thereof.
Over the last three decades, I’ve led technology organizations ranging from a few dozen engineers to global teams of 300+ across continents. I’ve modernized enterprise systems, scaled digital commerce platforms, rebuilt legacy stacks, and now increasingly design AI-native systems.
Here’s the pattern I see repeating:
Someone says: AI, build this.
What they often mean is: I haven’t clarified what this is.
AI amplifies structure and punishes ambiguity.
If your thinking is vague, AI will confidently produce vague output. If your system boundaries are weak, AI will quietly erode them. If your organization lacks alignment, AI will multiply the chaos.
When this happens, the problem isn’t the AI model. The problem is discipline.
Prompts Are Not Strategy
There’s a narrative spreading right now:
You don’t need engineers anymore.
Just prompt it.
AI will figure it out.
That can be dangerous noise. Strategy without execution is noise. Execution without strategy is waste.
AI does not eliminate architecture. It makes architecture more important.
The Difference Between Prompting and Engineering
Here’s what most people do:
Write an enthusiastic prompt
Get output
Sometimes they iterate randomly and get an amazing first product, but then the quality degrades quickly
Here’s what disciplined teams do:
Draft the idea. Brainstorm with others, as well as AI
Allow your AI tool of choice to teach you the latest concepts and approaches
Think through all aspects. What would be bad? What would be good?
Critique it
Do I have assumptions that might be wrong?
What can I learn from AI that might do this better than me?
Tighten constraints
Shift roles and review as: Architect, Security expert, and UX professional
Consider costs
“Is there a way to have minimal costs and not sacrifice performance?”
Document all you learn and integrate it into your plan
Put plans and content into artifacts and reusable assets
Reference your reusable assets frequently in the process
Continuously improve your content and artifacts as your product matures and as you learn
A disciplined process is the difference between experimentation and engineering.
The 5-Step AI Validation Loop
When I work with AI, whether architecting a system or fixing a performance issue, I use a consistent loop:
Step 1 — Draft
Come up with an idea.
Write an outline of your vision.
Go from the beginning to the end of the project in your mind.
Be as detailed as possible.
Don’t be prescriptive about the architecture, but be open to new ideas and concepts.
Propose your vision to your AI tool(s) of choice.
Ask AI to propose something better. Not perfection, but to iron out the wrinkles and incorrect assumptions.
Step 2 — Critique
Switch roles.
Review your own plan.
“Where could this fail?”
Ask AI to do the same:
“What risks should I consider?”
“Is this solution biased?”
Step 3 — Tighten Constraints
“Reduce scope by 30%.”
“Break into logical long-term phases, then optimize for MVP.”
“Don’t attempt to get done in one session.”
“Eliminate compliance risk.”
Step 4 — Role Shift
“Now review this as a security architect.”
“Now evaluate cost impact.”
“Now critique UX friction.”
Step 5 — Lock
Summarize the refined plan into a reusable artifact.
This is not theoretical. This is how you build durable systems.
AI Is a Junior Engineer
A good mental model I’ve found is this:
AI is a brilliant junior engineer. Fast. Capable. Occasionally wrong. Often overconfident. Needs guardrails. Thrives with clarity.
You wouldn’t let a junior engineer redesign your core architecture without review. Don’t let AI do it either.
If something doesn’t seem right in what AI is saying, challenge it.
Make AI question its own assumptions.
Do deep research and iterate. You can even have multiple LLM models do their own research and compare results.
Different LLM models have different strengths and weaknesses. Get to know them one by one and then leverage them individually for their strengths.
Where This Breaks Down
There are three common failure modes I see:
1. AI adds features you didn’t ask for.
It’s creative and takes liberties when constraints are loose. Suddenly you’re asking, “Where did that page come from?”
2. AI rewrites working systems unnecessarily.
You think you’re done. Weeks later, features vanish. Something that worked perfectly is now broken.
3. AI generates mock data that becomes technical debt.
The system appears functional until real-world data exposes structural flaws. Mock data is good for some testing, but real data can be messy and throw curveballs.
All three are discipline problems. All three are solvable.
Engineering With AI, Not Around It
I approach AI the same way I approach enterprise modernization:
Clear ownership.
AI does not control releases. You do. Nothing ships without explicit approval.
Clear constraints.
No new features without confirmation. All changes must be additive unless explicitly authorized.
Clear architecture.
AI should not make structural decisions independently, but it should be encouraged to critique and challenge them.
Clear review loops.
“Go into planning mode.” Planning mode means no execution, only structured iteration until clarity exists.
Clear communication.
Summarize what was built. Capture lessons learned. Update system memory to avoid repeating mistakes.
Ask AI to repeat back the plan, where you are, and what is next, and adjust it continuously as you go.
AI is a multiplier. Multiplication only works if the base number is strong.
Closing Thought
The teams that win with AI are not the ones with the cleverest prompts.
They are the ones with:
Strong architectural thinking
Plans that anticipate risk and cost
Clear decision rights
Tight feedback loops
Leaders close enough to the work to remove friction
AI does not replace leadership, but exposes whether you have it.