Tables, dates and forms: where generative AI slips
aieducation
A table can look perfect and still be wrong. Here's why generative AI slips on structure, and how good products catch it.
Ask an AI for forty expenses and you may get three date formats, a missing row twenty-seven, and a total that doesn't match. The cause is the same one behind its number mistakes: generative AI writes one token at a time. For teams doing generative AI for business, it's worth understanding where structure breaks.
Why tables fall apart
You see a grid. The AI writes one long line, with small separators between cells, and that line is folded back into a grid. If one cell is skipped, the row shifts and everything after it lands in the wrong column.
To keep row thirty-eight aligned, the model has to look back at a header it wrote long ago and work out which column it's in. Nothing enforces the grid, so the longer the table, the easier it is to slip. Dates drift too: without a clear rule, every date style looks likely, so the format can change halfway through.
Reading tables is also tricky. Columns in a PDF or a spreadsheet screenshot can merge, or numbers can jump rows, so check key numbers against the source. For more on this, see how to check AI-generated numbers.
How well-built products handle it
Top models are much better now, and many platforms use structured outputs to force an exact template. That keeps the shape right, but it doesn't verify the content.
Good products split the job:
- The AI fills in the fields: date, item and amount from each receipt.
- Ordinary software builds the table.
- A format check runs before you see anything, covering row counts, date formats and totals.
- If a check fails, the table goes back to the AI with a specific fix. This is the same idea behind tool use and code execution.
In short: AI for words, code for the exact part.
Getting cleaner tables from a chatbot
- Name the exact format, including the date style.
- Keep tables short, or ask in batches.
- Ask for CSV and open it in a spreadsheet.
- Count the rows and check the totals.
Before choosing an AI product, ask whether it forces a fixed format, whether software builds the table, and whether it checks row counts and totals first. Let AI fill it in, and let software line it up.