AI Data Quality: How Clean Is Clean Enough?
Perfect data is not the standard, and waiting for it is how companies spend a year preparing and ship nothing.
Erin Moore
Fractional Chief AI Officer
Your data is clean enough when the error rate it introduces is smaller than the error rate you are replacing. That is the whole standard, and it is a much lower bar than "clean data" implies — which matters, because chasing perfection is how companies spend a year on data projects and ship nothing.
Set the bar against the current process
The comparison is not against perfection; it is against what happens today. If a human currently mis-categorises eight percent of tickets, a system that mis-categorises five percent on imperfect data is an improvement, and waiting for the data to support two percent may cost more than the difference is worth.
That framing requires knowing your current error rate, which most companies do not measure. Establishing it is the first useful task, and it is worth doing whether or not you automate anything — the same argument as in data readiness for AI.
The four failures that actually matter
Not all dirt is equal. In practice these are the ones that break projects:
Missing values in the deciding field. Tolerable in a field the model ignores; fatal in the one it keys on.
Inconsistent definitions. Two departments recording "active customer" differently is not a data-cleaning problem, it is an unresolved business disagreement, and no pipeline fixes it.
Stale records. Data that was true and no longer is. Harder to detect than missing data and more damaging, because it looks valid.
Duplicates. Inflates volume, distorts anything aggregated, and quietly double-counts the benefit you are claiming.
The cheap test
Pull a sample — a week is usually enough — and have someone competent grade it by hand. Count the rows you would have to discard and why. A morning of this tells you more than a quarter of platform work, and it produces the error rate you need for the comparison above.
If a third of the sample is unusable, the data work is the project, and pretending otherwise just relocates the failure downstream.
Where cleaning genuinely comes first
Two cases justify fixing the data before anything else: when the errors are systematic rather than random, because the system will learn the bias; and when the output touches customers or money, where the cost of being wrong changes the calculation entirely.
For anything internal and reversible, ship against imperfect data, measure, and improve. That sequence is also what the 90-Day AI Playbook recommends, for the same reason: a measured, imperfect result beats an unmeasured perfect plan.
Frequently asked questions
How clean does data need to be for AI? Clean enough that the errors it introduces are fewer than the errors in the process you are replacing. That is a far lower bar than perfect, and it requires knowing your current error rate.
What data problems matter most? Missing values in the deciding field, inconsistent definitions between teams, stale records that still look valid, and duplicates. Inconsistent definitions are the worst, because they are a business disagreement rather than a technical fault.
How do we test data quality cheaply? Hand-grade one week of real data. Count what you would discard and why. A morning of that beats a quarter of platform work for deciding whether to proceed.
When should we fix the data first? When errors are systematic rather than random, or when the output touches customers or money. For internal, reversible use cases, ship and measure instead.
Further reading
Tagged with:
Ready to Automate Your Business?
Let's discuss how AI automation can deliver measurable ROI for your organization in 90 days or sooner.