Skip to content
← All insights[SEO and AI]

How I use AI without sacrificing verification and quality

AI is very good at producing a first version quickly. Speed becomes dangerous when plausible output is treated as a correct result. I use it as an accelerator for research, options and repetitive work, while keeping verification as a separate and mandatory stage.

I define the task and the consequences of an error

An internal summary, a data migration and a production change do not carry the same risk. Before delegating, I establish which sources are allowed, which files may change, what result I expect and which actions require my approval. The harder an effect is to reverse, the stricter the controls should be.

I separate generation from establishing truth

The model can propose hypotheses, structure or code, but I do not use it as the source for facts available in documentation, a repository, analytics or the actual product. An important claim must connect to observable evidence, and missing data must be reported rather than filled with assumptions.

  • Requirements are checked in their source document or system.
  • Existing code is inspected before proposing new architecture.
  • Metrics are recalculated or traced back to the data source.
  • Official documentation takes priority for APIs and versions.
  • Invented content is labelled as draft rather than presented as fact.

I ask for evidence that can be inspected

A response such as ‘fixed’ is too vague. For code, I want the diff, tests and rendered behaviour. For analysis, I want sources, assumptions and calculations. For an interface, I want browser verification at the relevant sizes. Short, reproducible evidence is the difference between confidence and hope.

I keep product decisions with people

AI can formulate options and consequences, but it does not automatically know the company's priorities, promises made to clients or real risk tolerance. Decisions about what we publish, which data we use, which compromise we accept and which experience we want to provide remain owned by an identifiable person.

I use a short working loop

  • Inspect the context and define the boundaries.
  • Generate a small change that is easy to review.
  • Automatically verify what can be automatically verified.
  • Manually inspect meaning, exceptions and the final experience.
  • Correct the result and preserve a clear record of the decision.
AI can shorten my path to a version. It does not transfer responsibility for the result.

Quality does not come from avoiding AI, but from using it in a process that separates speed from truth. Clear limits, verifiable sources, small changes, tests and human review make acceleration possible without losing control. The standard for the result should remain the same, regardless of who or what produced the first version.