AI and Data Quality: Foundation Work

AI does not tolerate bad data better than the previous generation of systems. It tolerates it worse, because it produces a fluent answer from it instead of an error.

A broken report looks broken. A model given the same broken input writes a confident paragraph, and nobody notices until someone acts on it.

What actually goes wrong

Duplicates. The same customer three times because three systems created them independently. Every count is wrong and every summary is subtly off.

Stale fields. A phone number updated in one system and not the other. Advisors learn quickly to ignore signals drawn from a stale field, and that habit is worse than having no signals, because it also kills the good ones.

Missing context. A field that is empty and a field that means zero look identical to a system reading it. They mean opposite things.

Definitions that drift. Two teams filling the same field with different intent, for years, and nothing flagged it because both were internally consistent.

The five things worth doing

Profile the tables that matter, so you know your actual null rates, duplicate rates and format breaks instead of guessing. Turn the worst findings into automated tests that run on every load and fail loudly. Give every critical domain a named steward in the business. Fix at the source rather than in the report. And publish a score per domain so improvement is visible and fundable.

That last one is the one people skip, and it is the reason quality work never gets funded twice.

Where it interacts with AI specifically

Retrieval amplifies whatever is in the store. Two versions of a policy and the system quotes whichever it found first, confidently, with a citation that makes it look verified.

Extraction inherits the format breaks. A date field with four formats in it produces four kinds of downstream error, and the model will happily parse all of them into something plausible.

The honest sequencing

You do not need perfect data to start. You need to know which two things are worst, and to fix those before you build something that will quietly launder them into answers people trust.