The most common reason an AI project produces no value is not the choice of model. It is the quality of the data it trains on.
Signs the foundation is missing
The same metric has three different values depending on the report. Nobody can say with certainty what "active customer" means. Historical data has been changed retroactively, with no trace in a log. An important field is filled in for 40% of records.
In that context, a model will learn the noise rather than the pattern.
What we build first
- The metric dictionary. Every metric has a single definition, agreed between business and technology, and an accountable owner.
- Versioned pipelines. Data transformations live in the repository, with tests that run on every change.
- Automated quality tests. Freshness, completeness, uniqueness, distribution. When a check fails, someone is alerted before the wrong report reaches the board.
- Traceability. Any figure on a dashboard can be traced back to its source record.
Only then, the models
With the foundation in place, a model becomes an engineering problem rather than a matter of luck. We evaluate it on real data, define the acceptance threshold together with the business team, and put it into production only after a period of parallel running.
And if, after evaluation, a simple rule achieves 95% of the model's benefit at 5% of the cost, we recommend the rule. That is part of our job too.