Using AI to Solve Programming Problems

March 10, 2026 · Scott D. Smith · Originally published on LinkedIn

A friend and I were discussing his next project for a restaurant chain to leverage AI and customer data. After a couple of beers, left over pasta and more than a few ideas, we both came to the same conclusion.

AI is not the correct tool to use for this project.

Simple reality was that standard python code would be the best approach, so that's the direction he took and it worked well for the client.

Yes. He needed the project to return clear and correct data the same time every time. Yes. The client was enamored with AI and had brought the project to him as an "AI Project."

Our conversation saved him at least a week of frustration and failure in retrospect.

We have reached a point of AI Overkill in 2026.

As a consulting veteran and former CISO/CFO, I'm seeing a dangerous trend:

Businesses are deploying AI "black boxes" to solve problems that a 20-line Python script could handle with 100% certainty.

Worse. Clients are predetermining what should and should not be an "AI Project" rather than letting the development team determine the approach.

Let's be clear. I LOVE artificial intelligence and vibe coding. Anything that solves problems for my clients and I faster is wonderful and I embrace it. However, it is still early days in the AI corner of the world. Most clients simply are not ready or capable of supporting it from an ML/Ops and privacy perspective. Others do not understand that it is not "magic" and solves all problems at lower cost.

The Fallacy

If your business logic can be defined by a set of rules, it belongs in a script/code, not a prompt.

• AI is Probabilistic: It predicts the next token and its predictions are often wrong.
• Code is Deterministic: If you write a function in Python or Java to parse a JSON schema or calculate a tax rate, it works the same way every time.

The Economic Reality Check

I recently reviewed a "GenAI" invoice reconciliation tool.

• The AI Path: $4,500/month in token costs + 15-second latency + occasional "math hallucinations."
• The Python Path: $20/month in compute + 200ms latency + 100% accuracy via Regex and a rules engine.

The "AI solution" was 225x more expensive and significantly less reliable.

The CISO/CFO Perspective

From a governance standpoint, "Prompt Engineering" is not engineering, it's fragile.

1. Explainability: "The model weights shifted" is not a valid defense in a NIST or FFIEC audit.
2. Technical Debt: Every AI wrapper you build is a dependency on a third-party model you don't control. A Java class is an asset you own forever.
3. Data Leakage: Stop sending proprietary logic to cloud LLMs just because you're too lazy to map a data structure.

The 90/10 Rule

True senior architects are pivoting back to the 90/10 Rule:

• 90% of your logic stays in structured, rules based code (Python/Java).
• 10% of your stack, such as "fuzzy" problems like sentiment or summarization, can go to AI.

Well written code remains easier to explain, audit and sustain. Take the time to analyze what approach and tools you will require.