What AI Can't Fix: The Limits of Automation Without Clean Data
An AI assistant trained on inconsistent records will produce answers as inconsistent as the records. The fix is not a better model.
A common request is an AI assistant that can answer questions about the business — stock levels, customer history, order status — pulled from existing records. This is achievable, but the quality ceiling is set by the records, not the model. If two staff enter the same kind of transaction three different ways, the assistant inherits that inconsistency and returns it as confident-sounding but wrong answers.
This is not an argument against AI. It is an argument for sequencing: clean up the fields that matter, agree on how they get entered going forward, and only then point an assistant at them. Skipping this step does not save time — it just moves the cleanup work later, after trust in the tool has already been damaged.
A useful test before any AI project: pick ten real records and check whether a competent new hire, reading only the data, could answer the question correctly. If a person cannot do it from the data, an AI model cannot either — it will simply do it faster and with more confidence.