Technical diagnosis before implementing AI

The first question is not “which AI do we use?”

Most projects start with the tool: model, framework, agent, vendor. Understandable, but usually the wrong order.

Before adding AI, you need to know which business piece is stuck. If the process is unclear, the data is unreliable, or nobody knows who decides, AI will not fix it. It will only make disorder move faster.

What a diagnosis should inspect

A serious technical diagnosis is not an endless audit. It is a fast, actionable read of five layers:

1. process: what happens today and where time is lost;
2. data: what exists, in what format, and with what quality;
3. systems: which tools already exist and how they connect;
4. decision: who validates, approves, or corrects;
5. risk: what happens if AI is wrong.

With that, you can decide whether you need AI, classic automation, process redesign, or technical debt removal.

Data before prompts

Many teams ask for “a good prompt” when they actually have a data problem: duplicate documents, inconsistent fields, manual exports, undisciplined CRMs, spreadsheets nobody governs.

If the base looks like that, an LLM can shine in a demo and fail in production. Not because the model is bad, but because the context is bad.

In legacy system modernization, AI often becomes useful after the flow is cleaned up, not before.

Decisions before automation

Another critical question: which decision do we want to improve?

It is not enough to say “automate support” or “build a sales agent”. You need the decision:

• which ticket should escalate;
• which lead deserves follow-up;
• which invoice looks anomalous;
• which document blocks an operation;
• which customer needs a human response.

Once the decision is clear, you can design a reasonable human-in-the-loop. Without it, you review everything or automate too much.

The useful output

A good diagnosis should end with a process map, prioritized opportunities, required data, non-AI quick wins, AI candidates, minimum architecture, effort estimate, and risks to avoid.

That is enough to make an executive decision. You do not need an 80-page report. You need to know what to move first.

If everything is already clear, build an AI-native MVP. But often the ROI appears earlier: discovering that the problem was not the model, it was the process.

If you want AI in a real process, diagnose it first.