Curated Data

Bringing together the right information with the right people will dramatically improve a company’s ability to develop and act on strategic business opportunities

— Bill Gates, Business @ the Speed of Thought [1]

What is curated data?

Curated data is enterprise data intentionally prepared for AI use, delivering reliable, organization-specific insights rather than fragmented or conflicting results.

The value of data in the age of AI forces a wider view of what counts as data. For most, the word “data” has historically conjured images of lakes, tables, rows, and SQL queries. While this structured data is still vital, it’s no longer sufficient. AI-Native organizations must now treat unstructured content, such as documents, unstructured customer feedback, and slide decks, with equal seriousness.

Across this wider field, two categories of data matter most, and the distinction between them runs through everything that follows. One improves how products are built; the other gives the products themselves their edge:

• Data that powers product development: the data people and agents draw on to make product development faster, cheaper, better aligned to intent, and more responsive to feedback.
• Data that powers products: the differentiated, proprietary data built into the products themselves, the basis on which their AI features compete in the market.

Models are available to everyone; what sets an organization apart is its own data. Curation makes critical data available to models and the humans collaborating with them responsibly and effectively. Four characteristics make data curated:

• Complete: people and agents can access all the knowledge needed to answer a question or complete a task.
• Accurate: the information is verified and trustworthy.
• Maintained: the data stays current.
• Securely accessed: the right people and agents can reach the right data, and no more.

How does curated data power product development?

The Outcome-Driven Product Development cycle runs on data. Regardless of the cycle’s stage, the quality of the data that informs it will either maximize or compromise its effectiveness. Figure 1 illustrates the relationship between different types of curated data and the cycle.

Intent data captures the product vision, strategy intent, and roadmap data that define current expectations. ART outcomes must leverage this data to identify the focus of progress in the coming year. Priorities are weighed against it, ensuring that the work that most advances the strategy comes first. Outputs are built to align with intent, with each feature, enabler, or experiment tracing back to the intent that justifies it.

Context data describes the conditions in which products operate. This data covers markets, customers, competitors, and regulations that affect the product. While intent focuses on the ‘why’ of the product strategy, context conveys understanding of ‘where’ it must succeed and the prevailing conditions. Outcomes must be grounded in it, priorities shift with it as windows open and constraints arrive, and outputs are shaped to fit the conditions they will meet.

Specification data provides a detailed definition of the expected outputs and the policies that must be adhered to. Agents utilize this data and are accountable to it. This specification is vital, even if it only encompasses a single stage of the process. It transforms intent and context into buildable and verifiable details. In an AI-native organization, the specification serves as the key artifact that agents create, build upon, and are reviewed against while producing outputs

Product performance data is evidence of how a product behaves in use. Performance is typically measured across four dimensions: business outcomes, user engagement, customer satisfaction, and technical performance. It provides objective data to support economic prioritization of PI outcomes and outputs. Further, it powers the feedback cycle that validates or invalidates hypotheses and enables live tracking of key results. Positive outcomes can generate feedback signals for amplification, while negative ones enable early truncation of misguided intent. Successful hypotheses will be reflected in performance data measuring customer and business value, as well as ROI.

Customer feedback data is the direct signal customers give about a product, gathered primarily through their actual use. It joins performance data at the metrics stage, supplying the reasons behind the numbers. Feedback not only shows changes in how a product is used, but also why customers responded as they did.

Each of these data types has a greater positive impact and is more powerful when connected to the others than when used alone. When an outcome traces to the intent that set it, the performance that measures it, and the feedback that explains it, a change in any one of them affects the others, and people and agents reason from a coherent whole rather than isolated fragments. Establishing and maintaining those connections is central to what it means to curate the data the cycle runs on.

How does curated data power products?

Curated data is crucial for product development, giving organizations a competitive edge with exclusive information that rivals lack. This data is viewed as an asset, and companies are increasingly protecting it rather than sharing it for free. Many resources that were once freely available online are now behind paywalls, as their owners recognize their value and prefer to safeguard and leverage them.

Its value shows most clearly in the products built on it. For example, a support assistant, grounded in a company’s help articles and past tickets, resolves problems that a general model cannot. A fraud model trained on a payment network’s own record of transactions and disputes scores risk in milliseconds, on patterns a newcomer has never seen. A clinical model built on a provider’s own linked patient and genomic records surfaces treatment options no one else can offer. Each rests on data proprietary to the organization and not reproducible by others.

Differentiated data need not already exist within the organization. Sometimes the strategy is to create it by designing a product that, when used, generates data no competitor has. Luis von Ahn’s early work showed the idea at its sharpest: his ESP Game turned image labeling into a game people played for enjoyment, producing a labeled dataset as a by-product that Google later licensed [2]. Platform products have turned this into a strategy of their own. A social platform views the data flowing through it as an asset that it owns and can develop further. As more people use the platform, it becomes a powerful engine for generating data assets that grow in value.

For a Product Manager deciding where AI can set a product apart, three questions prime the work.

What differentiated data do we own, or could we build, that others do not? The starting point is the data the organization has that competitors cannot easily match. Richness is part of the test: how deep the data runs, how specific it is to the product’s customers and segments, how current it is, and whether it is genuinely better than what a rival could assemble.

What could we do with it? Differentiated data is only potential. The question that turns data into an advantage is what problems could it solve and what new capabilities could it create if used imaginatively? The same dataset can sit unused or become a feature no competitor can offer.

What stands between possessing that data and using it safely? Owning useful data is not the same as being able to act on it. The most valuable data is often the most sensitive. Records from an interactive voice response system, for instance, can be rich in customer signals and dense with personally identifiable information that carries real risk if mishandled. Making that data an asset that an agent can draw on means curating it so sensitive material is protected and surfaced only where it belongs.

How do we prioritize what to curate?

Prioritizing what to curate starts with assessing the problem or opportunity’s significance and the feasibility of obtaining the curated data needed to address it. Typically, a specific set of data can be relevant to multiple problems. Therefore, prioritization focuses on identifying patterns: the recurring issues within the organization’s objectives, the data those issues rely on, and the efforts required to prepare that data for use in a timely manner.

Feasibility mostly depends on how well the data is currently curated. Some data is already curated and ready to draw on. Some is well-organized internally but limited in accessibility and integration. And some barely exist in a usable form. The state of the data sets the cost of pursuing the opportunity that needs it.

Curation can be intensive, so it is not always the first move. When an application looks promising, it is often wiser to test it before committing to an investment, running an experiment, or building a prototype that validates or rules out the opportunity using a fraction of the data. Once the opportunity proves worthwhile, acquiring and curating the data becomes a matter of crafting and prioritizing the appropriate enablers for the architectural runway.

For data that drives product development, a natural starting point is intent data. This includes the vision, desired outcomes, strategy, and execution intent that define what the organization aims to achieve. Because of its foundational importance, this is the natural place to begin.

The balance between the two types of data is a strategic choice that changes over time. An organization that is working to accelerate its AI -Native capabilities focuses more on the data that drives product development. In contrast, an organization that is already proficient in AI tends to prioritize the data that supports its existing products.

This naturally evolving balance offers the advantage that the risks associated with data used for product development are typically much lower than those associated with data used in actual products. The skills, supporting policies, and pipelines developed during the curation of product development data transfer naturally to product data, reducing risks in the process.

What are the essential qualities of curated data?

The data quality model in the international standard for software product quality requirements and evaluation (ISO/IEC 25012)[3] provides an extensive list of data quality attributes. These may be broadly aggregated using the categories introduced earlier in the article, as illustrated in Table 1.

Category	Qualities
Complete: People and agents can reach all the knowledge a question or task requires; agents, unlike people, cannot fill gaps with judgment	Completeness: all expected values and relations present Consistency: free from contradiction across sources Understandability: interpretable, with clear labels and units
Accurate: The information is verified and trustworthy, and its lineage can be followed back to its source, which matters more than ever as AI generates synthesis that must be separated from fact	Accuracy: correctly represents the true value Credibility: regarded as true and believable Precision: exact to the level of detail needed Traceability: an audit trail of access and change
Maintained: The data stays current; at the speed agents act, stale data produces stale answers just as fast	Currentness: kept at the right age for use Recoverability: retains its quality and operation through failure Portability: moves between systems without losing quality
Securely accessed: The right people and agents can reach the right data, and no more, with agents now among the consumers that access must account for	Accessibility: reachable by those who need it Confidentiality: reachable only by the authorized Compliance: adheres to the regulations in force

Across all four sits a fifth quality the rest depend on: curated data must be owned. Data ownership is not a new discipline, and many organizations have long struggled with it, but it becomes an imperative once people and agents act on data at speed. 

The recognized data-management body of knowledge (DMBOK)[4] separates accountability from day-to-day care: a data owner holds approval authority over a domain of data, while data stewardship is the work of keeping it to standard. What is new is the reach: as far more of an organization’s material becomes curated data, it will need owners and stewards for kinds of data it never thought to govern before.

How does AI help with data curation?

Producing curated data has always been challenging. Even transferring structured data from source systems to a warehouse involves time-consuming processes like profiling, modeling, and governance. Today, the demand is even greater. An AI-native organization aims to gather more data from various sources, and as it advances in its AI journey, its need for data intensifies.

The good news is that AI can carry much of the load. Models can scan a new source for personally identifiable information before it is brought in. Agents can work through a body of data to identify and formalize the metadata that makes it understandable. Much of the mechanical effort of curation, the profiling, the tagging, the classifying, is work that AI and automation can now take on.

Curation is not just a mechanical process. It heavily relies on judgment, and some decisions can be challenging. Questions arise, such as which sources are trustworthy, what information is considered outside acceptable boundaries, and where the data may be confidently wrong. To prevent the workload from becoming overwhelming, it is essential to delegate as many of the mechanical tasks as possible to AI and automation. Additionally, we should invest in features and tools that support decision-making without making the judgments themselves.

Used this way, AI is turned on the very discipline that feeds it: the data that powers AI is curated, in part, by AI. That is how an organization catches up with the work in front of it and keeps up with the work to come.

---

References

[1] Gates, William, and Collins Hemingway. Business @ the Speed of Thought. New York: Warner Books, 1999

[2] Luis von Ahn, Laura Dabbish. Labeling Images with a Computer Game. https://dl.acm.org/doi/epdf/10.1145/985692.985733. See also Law, E., & von Ahn, L. Human Computation. Morgan & Claypool Publishers. 2011.

[3] ISO/IEC 25012:2008, Software engineering — Software product Quality Requirements and Evaluation (SQuaRE) — Data quality model.

[4] DAMA International, DAMA-DMBOK: Data Management Body of Knowledge, 2nd edition. Technics Publications, 2017.