Data quality is holding insights back

By Abigail Stuart, Founding Partner, Day One Strategy

Data quality, not AI, is what’s holding pharma insights back

I’ve spent 20-plus years trying to understand human behaviour, and that work has always shown me the same thing: the insight is only as good as the data it’s built on. For example, research that sets out to understand why HCPs prescribe what they do, but only looks at rational, conscious drivers, will always fail to capture the emotional and unconscious signals that actually underpin behaviour.

That’s one of the key themes running through our second annual AI in Pharma State of Play report , published last week, based on 20 in-depth interviews with insights and analytics leaders across 10 pharma organisations, conducted this summer.

A key finding, not surprisingly, is that AI adoption in pharma insights teams is now widespread. Access is no longer the issue. The state of the underlying data is – and it’s one of the things stopping teams from turning that access into real commercial value.

And there are really two challenges here:

1. The first is whether AI can access and use the data companies already have.
2. The second is whether that data is rich enough to support the decisions being asked of it.

First, make the data useable

One quote from our interviews summed up the problem:

“The model is the easy part. The harder work is getting the data ready for it.”

That problem is particularly apparent when it comes to market research data. Company data lakes are usually well resourced with structured sales and digital data. But the rich qualitative and quantitative data that companies invest in can stay trapped in slide decks and transcripts, often in formats no model can easily read. Duplicate and inconsistently named files came up as the single biggest barrier to adoption – bigger than any limitation of the AI itself.

As one Senior Global Customer Insights Manager told us:

“Meta-analysis is a very good keyword… we are sitting on a treasure that we are not really kind of living.”

Making that treasure accessible is crucial. But that leaves a much bigger problem. Once AI can reach the data, the next question is whether the data itself captures enough of the human reality to be useful.

Then, make the data worth using

Making existing research accessible is only half the opportunity. We also need to think differently about the research we are creating now, particularly for brands where there is little historical data to draw on.

Some leaders told us that because their biggest, in-launch brands carry the most commercial pressure, teams are understandably cautious about experimenting with AI on them. AI use can therefore end up focused on lower-priority, end-of-lifecycle brands, where the risk is lowest.

But we think there is a bigger opportunity earlier in the brand lifecycle. A brand in development has little historical data of its own, so the primary market research being commissioned today becomes especially valuable. It isn’t just answering one question. Done properly, it becomes

part of the data foundation for AI, built to answer today’s question and potentially hundreds of future ones. That’s a very different way of thinking about the value of a single research project.

A study published this week backs this up. A Columbia University team tested AI digital twins against the real people they’re built to represent – nearly 1,800 people, twins trained on 500-plus of their own prior survey answers. The twins barely beat an AI given none of that person’s history at all, and got it wrong in predictable ways: flatter answers, heavier reliance on stereotypes, more optimism than the real humans.

What this suggests is that five hundred survey answers may still leave gaps in what actually drives a decision – the mental models, the unconscious and emotional responses behind a choice, the things people can’t easily articulate. That’s the gap richer, deeper human research is designed to close. A synthetic model of a customer, a patient or a brand is only as good as that underlying depth, which is the case for investing properly in primary market research now, rather than treating it as a box to tick before the “real” AI work begins.

It’s a variation on the lesson I opened with: the differentiator was never having AI. It’s having human data underneath it that’s deep enough to trust.

Where to start

In the report, we set out a full checklist for becoming data-ready. Three places to begin:

1. Triage your highest-value data first. Start with the datasets that matter most rather than trying to include every file at once.
2. Create one shared data standard for every agency and vendor. Set clear guidelines for how research is named, stored and structured.
3. Design research for future AI use, not just today’s brief. Think upfront about the behavioural, emotional and contextual data future models may need, as well as how the research should be structured and tagged for reuse.

Read the full report here: https://dayonestrategy.com/resources/ai-in-pharma-state-of-play-report-2026/