Will AI replace data analysts?

What models genuinely do well now

• Writing the query or formula from a description. The syntax barrier — the reason people who understand their data still cannot get answers out of it — is substantially gone.
• Judgment-shaped cleanup at scale. Deciding that "Acme Corp.", "ACME Corporation" and "acme corp" are one company. Rules do this badly; models do it well.
• Reading unstructured sources. Invoices, statements, free-text fields.
• The first pass on an unfamiliar dataset. "What is in here" is answered faster by a model than by a person opening it.

What they are still bad at, and why it is structural

• Knowing that the data is wrong. An analyst who knows the business notices that a region's numbers doubled because a sales office was reclassified. A model sees a doubling and explains it confidently. This is not a capability gap that more training fixes — it requires knowledge that is not in the data.
• Choosing the question. The valuable move is usually reframing what was asked, and asking for it is not something a model is positioned to do.
• Being trusted. Someone has to sign the number. Accountability does not delegate to a system that cannot be accountable.
• Knowing what was not measured. Every dataset has a survivorship problem, and it is by definition not visible in the dataset.

What actually changes about the job

The part of analysis that was craft — remembering the syntax, joining the tables, wrangling the export — is getting cheap. The part that was judgment — knowing which number is wrong, which question is the real one, and what the data does not contain — is getting relatively more valuable, because supply of the first part collapsed and demand for the second did not.

The uncomfortable version: this is bad news for anyone whose value was the craft part, and it is arriving faster than most career advice assumes. The people who are fine are the ones who could already tell you why a number looked wrong before they checked.

What we built accordingly

TableDI 2 is a desktop app for file work you redo every period: it keeps your sources, rules and delivery as a job, so next month you drop in the new files and run it again. Everything runs on your own machine. Its AI is optional, runs on your own API key, and is placed at the craft layer only: drafting a job from a sentence, explaining an exception, suggesting a rule. It never runs a period. The run itself is plain rules, so the same files give the same result every time, and when the AI drafts a job it sees the column names and the first 50 rows, not the whole file.