Caroline Chung, is Vice President, Chief Data & Analytics Officer at MD Anderson Cancer Center.AIgettyIn the rush to adopt AI, many organizations focus on models, scale and data volume. Far less attention is paid to something quieter but far more consequential: context. Without it, even the most sophisticated analytics can mislead. With it, leaders can make better decisions, avoid wasted pilots and build systems that actually hold up in the real world.This question of how people come together to make data-informed decisions is one I explored previously in my last article. Context is what operationalizes that culture. Context becomes relevant from the moment data is collected or generated and accompanies the data through every stage of flow and use, accounting for biases, uncertainties and assumptions that can influence decisions. If we ignore the context, we shouldn’t be surprised when AI models or systems show poor performance or behave unpredictably.Where Assumptions Enter The DataWhen people think about data, they often picture it in rows and columns of a table. By the time it reaches that form, assumptions are already in play. A column label may look straightforward, but definitions can vary in ways that are invisible unless someone pauses to ask what the data actually represents. For example, who was filling out the data within this column, under what circumstance and with what instructions? These contextual details can be indicators of the consistency or reliability of the data and whether this data is similar or different from other data with a similar or even the same column label.During the pandemic, one example surfaced repeatedly: the number of days a patient spent in the hospital. That sounds simple in theory. In practice, it was anything but. Some teams counted full 24-hour periods. Others counted calendar days crossed at midnight. In some cases, a patient admitted minutes before midnight was counted as a full day. When those datasets were combined, the resulting statistics shifted meaningfully.That example is relatively simple. The complexity increases when data is entered by people interpreting questions through their own lens. In surveys, clinical assessments or intake forms, individuals may respond differently based on motivation, understanding or incentives. Someone hoping to qualify for a job or a clinical trial may answer differently than someone without that context. If we do not consider how data is interpreted at the point of entry, we lose visibility into the uncertainty before analysis even begins.Historically, data was collected for a single intended purpose. Re-use was limited because moving and analyzing data required significant effort. Today, data flows easily across systems and AI thrives on aggregation. That shift requires a different mindset. We need to think deliberately about which data elements are most likely to be valuable for re-use, under what conditions and with what level of confidence. We also need to shift how we generate data with context in order to fulfill this downstream purpose beyond the initial intended use. The more context, the greater the likelihood that the data will have value downstream, so it is worth investing these efforts for the data with the greatest likely downstream value.When Context Disappears, AI Finds The Wrong PatternsEarly assumptions matter even more once data is used for AI. These systems are exceptionally good at identifying patterns, regardless of whether those patterns align with our intent.A well-known example emerged from a study conducted by MIT. Researchers evaluated AI models designed to detect Covid using chest imaging. The models performed well statistically but failed when deployed in emergency rooms. The failure was puzzling and could not be explained by commonly cited biases alone.What later became clear was a missing piece of context. The Covid-positive images used for training came almost entirely from adults. Many of the Covid-negative images came from children. The models learned to distinguish pediatric lungs from adult lungs rather than Covid from non-Covid. The pattern was real, but it was not the pattern the researchers intended to capture.Had the contextual metadata of patient age of the imaging data been surfaced earlier, teams would have separated the datasets and trained models differently. Instead, the focus on the data without considering the context obscured the signal. This is a recurring issue in AI. Combining data that is not fit for the same purpose in the absence of contextual metadata introduces noise that in many cases no amount of scale can resolve.Context does not stop AI from finding patterns. It helps humans determine which patterns matter.Contextual Thinking to Avoid Death By A Thousand PilotsMany organizations are struggling with what is often described as death by a thousand AI pilots. Teams launch experiment after experiment without achieving durable impact. The root cause is rarely a lack of effort or talent. More often, it is how the problem is framed.The starting question is frequently, “How can we use AI?” That framing almost guarantees misalignment. A more effective approach begins by defining the problem or purpose clearly with the entire context. What are we trying to improve? Who is involved? Which workflows will be affected? Where might disruption create value and where might it introduce risk?When we bring teams together to fine-tune a problem statement, we'll sometimes find that, while AI may be the motivator, the solution is an operational workflow issue unrelated to technology. Other times, a technology solution can bring notable benefits and address the problem, but only after the underlying process is understood across the team. Contextual thinking allows leaders to collectively consider and strategically select pilots that have a realistic path to implementation at scale. It surfaces dependencies early, clarifies fit for purpose and reduces the tendency to apply technology as a generic fix for complex systems.Why Context Matters More In AI Than In AnalyticsContext has always mattered in analytics, but AI amplifies the consequences of getting it wrong. Many AI models still function as black boxes. While research into explainability continues to advance, models will surface spurious correlations if context is missing. More deterministic methods often expose inconsistencies through human review. Newer probabilistic AI models make this much more obscure.Another challenge is that many AI models are typically trained and tested on highly curated datasets. The real world does not behave this way. Data arrives incomplete, noisy and inconsistent. Without considering and understanding the context of use, leaders cannot anticipate how models will perform in their own environments, also appreciating that their environments can change over time.An aspect of the context of use is the gap in AI literacy. Many AI tools are now deployed directly to end users who are not data or AI experts. Without clear context around limitations, uncertainty and appropriate use, various risks emerge, including confirmation bias, automation bias that may result in decisions and actions with substantial consequences. AI Agents Raise The Stakes For ContextThe next wave of AI adoption is shifting from systems that inform decisions to AI agents that act on them. Agents are increasingly being deployed to carry out multi-step tasks, retrieving records, summarizing findings, drafting communications and coordinating handoffs across workflows, often with less human review at each step. An analytics output is a pattern a person weighs before acting. An agent acts.That autonomy raises the stakes for context. A human analyst who encounters an ambiguous field can pause to ask what the data actually represents. An agent will not pause unless we give it a reason to. Return to the hospital length-of-stay example: An agent asked to aggregate data across systems will happily combine incompatible definitions unless the contextual metadata travels with the data and the agent is designed to respect it. Context is how an agent knows what it is working with and, just as important, what it should not do with it.The same contextual thinking that guides pilot selection applies here. Before an agent is given a goal, leaders should be able to first ensure the goal is defined clearly while considering the various questions around context. Whose work does it act on? Which workflows will it touch? What are the boundaries of appropriate use, and when should it be handed back to a human? An agent given a goal without this context will optimize for something, but not necessarily for what was intended, and it will do so at machine speed and scale.Context does not limit what agents can do. It is what makes their autonomy safe enough to be useful.Building Context Into The Data LifecycleCapturing context requires both culture and infrastructure. Literacy and fluency must be priorities. Everyone across the ecosystem plays a role in shaping data quality and meaning.Practically, this means thinking about the entire data lifecycle and building a metadata supply chain alongside the data itself. Metadata can include definitions, timestamps, collection conditions, intended use and security and governance. In many cases, much of this context already exists and is often overlooked.When someone receives a dataset, the expectation should be that they review the metadata before analysis. Does this data fit the intended purpose? Is it appropriate to combine with another dataset? What uncertainties should be considered?The same principle applies to AI models. Just as data carries assumptions, models do as well. Model descriptors, perhaps on model cards, help document those assumptions so decision-makers can appropriately weigh outputs. This mirrors how we evaluate human expertise. Confidence and trust matters, and this often depends on the source of the information..Ultimately, context enables informed decisions that integrate and leverage the complexity, not just faster decisions. In AI, that distinction is critical. When we preserve context from collection through application, we give ourselves the ability to trust what these systems are telling us and to know when not to.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?