Startup pitch: “We use AI to automate X.” Investors: “Wonderful. Here is €2M.”

Six months later: an AI budget of €120,000 — data scientists plus compute — and a model at 78% accuracy. Benchmark of a 200-line Python script: 74%.

Difference: four points. Cost: €120,000.

(If you think your AI project is different — it might be. The odds are against you.)

What eight audits showed

Project one, support ticket classification. Solution sold: fine-tuned LLM. Solution that works: regex plus ten if/else rules. Saved: €60,000.

Project two, product recommendations. Solution sold: custom neural network. Solution that works: collaborative filtering, fifty lines of scikit-learn. Saved: €40,000.

Project three, fraud detection. Solution sold: real-time ML model. Solution that works: simple statistical thresholds — amount above X, unusual geolocation. Saved: €80,000.

flowchart TD
    A["8 AI projects audited"]
    A --> B["6 projects
a Python script was enough"]
    A --> C["2 projects
ML genuinely required
(advanced NLP, vision)"]
    B --> B1["Total saved: €340,000"]
    B --> B2["Time-to-value ÷ 6"]
    C --> C1["Approved only AFTER proving
simple rules were insufficient"]
    style B fill:#1d3b26,stroke:#4ade80
    style C fill:#4a3b00,stroke:#fab400
    style B1 fill:#1d3b26,stroke:#4ade80

The problem is never the complexity of the problem

It is the complexity of the solution being sold. Because a complex solution justifies a large budget, impresses the board, and requires a specialised team — which is to say, job security.

Three questions I ask before recommending AI

Have you tried the dumbest thing that could possibly work? In six cases out of eight, the answer was no.

What accuracy gap is actually acceptable? If 74% is enough, why pay €120,000 for 78%?

Who maintains this model in eighteen months? If your data scientist leaves, the model dies with them. If it is plain Python, any developer can keep it alive.

The bridge to constraint

When you cannot afford the sophisticated solution, you start with the one that works. And it is often enough.

Sophistication is a luxury you buy once simple has stopped working — not before.

AI is a powerful tool. But most problems do not need power. They need clarity.

Start with the solution that fits in 200 lines. If that fails, then reach for the model.

Sources

• Field observations — AI audits, 2024-2026