Surveys keep finding the same thing: most organisations do not consider their data foundations ready for AI. A 2026 Deloitte survey found around 72% lack unified, accessible data. The wrong conclusion is that you need a data programme before you can automate anything. The right one is that you need enough data quality for one workflow at a time.
Three questions per workflow
- 01Access: can a system read what a person reads to do this job? If a human opens four screens, the automation needs all four sources, not three.
- 02Consistency: does the same thing get recorded the same way? Not perfectly, but predictably enough that the exceptions are a list rather than a shrug.
- 03History: do you have enough past examples to test against? A few hundred real cases with known-correct outcomes is usually enough to measure quality honestly.
If all three are yes for one workflow, you can start on that workflow, whatever the state of the rest of the business.
Good enough beats clean
Waiting for clean data is how a year passes with nothing shipped. Modern extraction handles far more variation than the rules-based tooling that shaped everyone's expectations. Duplicates, inconsistent naming and missing fields are workable when the system is designed to flag uncertainty rather than guess. What is not workable is data nobody can reach.
Sequence it the other way round: the first automation usually reveals exactly which data problems cost money, which makes the case for fixing those and only those.
The one thing worth doing first
Build a test set. Take 200 to 500 real historical cases with known-correct outcomes and put them somewhere stable. That single artefact lets you measure a system honestly, compare approaches, catch regressions when a model changes, and settle arguments about whether quality is improving. Teams that have one ship; teams that argue from anecdotes do not.