Every client says this, and it is rarely the blocker
“Our data is a mess” is the most common sentence in a first call. It is usually true and usually not disqualifying — extraction, matching and validation exist precisely because real data is untidy.
What matters is which kind of mess you have.
Mess we handle by design
- Inconsistent formats — dates as text, numbers with stray characters, mixed units
- Free text where a field should be, including notes fields used for a second purpose
- Duplicates, matched and merged with a review band
- Legacy records created by people who left, under conventions nobody documented
- Customer shorthand that means something specific to one account
Mess that genuinely blocks
If the information was never recorded, nothing recovers it. A model can read a badly formatted delivery date; it cannot invent one that was only ever agreed on the phone.
- Missing entirely — the field exists in the process but not in any system
- Contradictory — two systems disagree and nobody owns the answer
- Unrecorded decisions — the rule lives in one person's judgement with no pattern to learn from
What we do about it during discovery
We profile the data before quoting: counts per entity, completeness per field, distinct values in fields that should be constrained, and the outliers. That takes days and it changes the plan more often than any other input.
It also produces a list of what needs fixing at source, which is frequently cheaper and more valuable than anything we build on top.
Fix capture, not history
Clean what is actively used, archive the rest, and put validation at the point of entry so the mess stops being created. One good validation rule prevents more than a year of cleaning scripts corrects.
Businesses often spend months cleaning records nobody will look at again. We will tell you when that is what you are proposing.