‘We need to clean all our data first.’
False. You first need to know which data the intended use depends on. A big clean-up beforehand delays everything and rarely serves the use that will matter.
Updated on 24 October 2026
What is going on.
The idea seems like common sense: AI feeds on data, so the data must be clean before you start. Many companies therefore launch data quality, master data and de-duplication projects, and put AI off until they are finished.
These projects rarely finish. Data keeps arriving, with its flaws, while yesterday’s is being cleaned. And the effort is spread across all the data, whereas a given use draws on only a small part of it. It is better to start from the intended use, identify the data it depends on, and focus the effort there, at the source if possible.
What we observe.
- Cleaning without a use means cleaning without a criterion.You know neither what level of quality to aim for nor which data matters.
- Data degrades again.If the cause is not dealt with at entry, the clean-up has to be done again.
- Today’s models tolerate some imperfection.They can read heterogeneous documents, provided their results are checked.
What you can do right now.
- Put on hold any clean-up project that does not serve an identified use.
- For your first use, measure the quality of only the data it needs.
- Fix the cause of the flaws at entry rather than their effects after the event.