Field Note - Data and Analytics
Data Leaders on Building the Foundation Before Deploying the AI
We spoke with chief data officers and senior data leaders at twelve enterprises that have successfully scaled AI programmes. What they said about the decisions they made before the first model was deployed is not what most AI conversations focus on.
About This Field Note
Elixir conducted structured conversations with chief data officers and senior data and analytics leaders at twelve enterprises across financial services, healthcare, and manufacturing in the Gulf and India. All participants had successfully scaled at least two AI programmes to production. This field note captures the patterns in what they said about data foundation decisions made before AI deployment.
The Admission Most Data Leaders Make in Private
When we ask data leaders at enterprises with mature AI programmes what they wish they had done differently, the most common answer is not about the AI. It is about the data that came before it.
In most organisations, the AI conversation began before the data foundation was ready to support it. The pressure to demonstrate AI capability -- from boards, from competitors, from vendors, from internal advocates -- was faster than the pace at which data quality, data access, and data governance could be established. Pilots launched on data that was good enough to demonstrate the concept but not good enough to sustain the programme in production. The gap between pilot and production was wider than anticipated. The reason was usually the data.
The leaders we spoke with who had navigated this successfully had made one of two decisions: they slowed the AI programme down until the data foundation was ready, or they fixed the data in the specific domains where the first AI programme would operate before launching the programme. The ones who did neither had more expensive programmes that took longer to generate return.
What "Ready" Actually Means
"The data is not ready" is a statement that requires precision to be useful. Ready for what, specifically? The leaders we spoke with described data readiness in operational terms: the data that the AI system will use to make decisions is accurate enough, consistent enough, and available quickly enough that the system's outputs can be acted on without requiring systematic human verification.
This standard is higher than the standard most enterprise data currently meets, and lower than the standard of "all data problems fixed." Most of the leaders we spoke with built a pragmatic definition: they identified the data fields that the first AI system would directly use, assessed the quality of those fields specifically, and fixed what needed fixing in those fields before the AI deployment. Other data quality problems were deferred to later phases.
This selective approach to data readiness is consistently faster than trying to fix everything before starting. It is also more likely to generate the visible, measurable improvement in data quality that builds organisational will for the broader programme.
The Governance Piece That Gets Skipped
The most common gap in data foundations for AI, across the organisations we spoke with, was not technical. It was governance. Data governance -- the policies, processes, and ownership structures that determine how data is created, managed, and maintained -- is the unglamorous work that technical teams often defer in favour of the more visible work of building pipelines and training models.
The consequence of deferred governance is that data quality improvements made for an AI programme do not persist. New records are entered incorrectly because the processes that create them have not changed. Fields that were cleaned are re-contaminated because the systems that update them have not been brought into alignment. The AI system's performance degrades over time, and the diagnosis of why is more complex than it needs to be.
The leaders whose AI programmes have maintained their performance over time are, without exception, the ones who built governance alongside the technical foundation. They defined data ownership clearly: specific people are responsible for the quality of specific data domains, with explicit accountability for maintaining the standards the AI system requires. This is not complicated. It is also not the default.
-- Chief Data Officer, Financial Services (Gulf)"You can build a very sophisticated model on top of a mediocre data foundation. What you cannot do is get reliable outputs from it. The model does not compensate for the data. It reveals the data."
What They Would Tell Their Earlier Selves
We asked each of the leaders we spoke with what they would tell themselves at the start of their AI journey. The answers clustered around three themes.
The first was: start with the data audit, not the use case. Understanding the quality and structure of the data available is the prerequisite for choosing the right first use case, not something that happens after the use case has been chosen. The leaders who reversed this sequence typically found that their preferred first use case was blocked by data problems that could have been identified earlier.
The second was: invest in the data infrastructure before the model. The marginal improvement in model performance from using a more sophisticated model is almost always smaller than the marginal improvement from using better data. The organisations spending disproportionately on model capability while underinvesting in data quality are making a systematic error in where they place their resource.
The third was: make data governance boring and early. The organisations that waited until data governance was urgently needed -- when AI programmes were failing due to data quality -- spent more time, money, and goodwill establishing governance than the organisations that built it before it was visibly needed. Boring governance that nobody talks about is what good data infrastructure looks like. Dramatic governance interventions during an AI programme are what happens when it was skipped.
