AI-Ready Data Is Not a Technology Project. It Is the Operating Model for the Intelligent Enterprise.

About the Original McKinsey Research

A Deep Examination of McKinsey's Argument

The Four Enterprise Technology Shifts Stalling AI Scaling

McKinsey organizes its analysis around four enterprise technology shifts that it describes as stalling AI scaling. The four shifts below are McKinsey's, rendered in plain English. The Market Signal notes that follow each one are our interpretation, not McKinsey's conclusion.

One. Unstructured data is becoming a harder enterprise challenge

Two. The risk surface area is expanding as AI retrieves, recombines and generates data

Three. AI is generating and reusing data faster than governance models can keep up

Four. Fragmentation is making chief data officers central to AI enablement

The Financial-Services Case Study

From Governed Artifacts to Reusable Business Data Products

McKinsey's wider article demonstrates the importance of treating the artifacts created as unstructured information is extracted, transformed and prepared for AI as governed assets rather than disposable outputs. We agree with that foundation. Our view is that the next step is to organize those governed artifacts around the decisions the business needs to make and turn them into reusable data products that carry business purpose, meaning, ownership, policy and evidence alongside the information itself.

Everything that follows in this section is the Data Tiles position rather than a McKinsey finding. The terms business data product, fused data product and active governance are ours. They describe how we believe the direction McKinsey identifies should be operationalized, and they are offered as an extension of that direction rather than a criticism of it.

A governed artifact is valuable. A reusable pipeline is valuable. A curated unstructured data product is valuable. But none of these, by themselves, produces a trusted business decision. In our model, a trusted business data product also requires a defined business purpose, an intended decision or outcome, business meaning, ownership, appropriate-use rules, decision-specific quality expectations, active access controls, visible lineage and evidence, and a reusable consumption model.

This is where our starting point differs. We do not begin with AI readiness. We begin with the business decision or outcome. The decision determines which information is relevant, which business definitions apply, which sources need to be connected, what fit for purpose means, what level of quality is required, which policies must apply, who owns the resulting product, who or what may access it, what evidence must be retained, and how success should be measured.

Business data products

A business data product organizes reusable, governed information about a core part of the business. Illustrative examples include Customer, Contract, Product, Policy, Transaction, Claim, Supplier, Asset, Employee, Risk, Case and Regulatory Obligation. These are examples rather than a fixed catalog, and they are not drawn from McKinsey's research or from named customer deployments.

A business data product carries the relevant business definitions, ownership, source relationships, quality expectations, access rules, sensitivity classifications, lineage, intended uses and evidence. It does not require physical centralization, and it is not an attempt to expose every available field. Latttice works with the enterprise systems and platforms an organization already has in place, and its purpose is to make the relevant information governed, understandable and reusable.

Fused data products

Business data products can then be fused around a specific business decision or outcome. Illustrative examples include a customer risk review fusing Customer, Contract, Transaction, Policy and Risk; a claims assessment fusing Claim, Customer, Policy, Asset and Evidence; a regulatory response fusing Regulatory Obligation, Policy, Transaction, Customer and Evidence; a credit or lending review fusing Customer, Account, Transaction, Contract and Risk; and supplier exception management fusing Supplier, Order, Shipment, Contract and Performance. Again, these are illustrative Data Tiles examples, not McKinsey material and not named implementations.

A fused product is not created by asking what data can be combined. It is created by asking what decision the business must make and what trusted information is required to make it. The decision determines relevance, quality, appropriate governance, intended use, access, evidence, ownership and the measures of success.

Reuse and compounding value

A governed Customer product may support customer-risk assessment, service prioritization, claims handling, churn analysis, fraud investigation, regulatory review and an AI customer-service agent. A governed Policy product may support compliance review, claims assessment, employee guidance, regulatory reporting and an AI policy assistant. A governed Contract product may support obligation monitoring, supplier risk, revenue assurance, customer review, exception handling and legal analysis. Each of these is illustrative, but the pattern is the point.

Reuse removes the need to recreate definitions, ownership, lineage, policies, quality expectations and evidence for every new initiative. The hundredth use case should not start from zero. McKinsey's own data-product research makes a related argument: data products should begin with business value rather than data improvement for its own sake, should serve multiple high-value use cases, and should create a flywheel in which reuse lowers incremental cost and shortens time to value.

A fused product may itself be reused where appropriate, but every reuse must preserve lineage, ownership, controls and clarity of intended use.

Active governance

Active governance is our position, not a McKinsey term. Governance must remain active during both the creation and the consumption of a data product. It should not exist only as catalog documentation, a policy document, a project sign-off or a one-time ingestion control.

A product should behave according to its governance whether it is consumed by a business user, a dashboard, an application, an API or an AI agent. Governance determines which rows and fields are available, who or what may access the information, which policies apply, whether sensitive information may be exposed, whether the product is appropriate for the decision at hand, what evidence accompanies the result and how usage is traced.

What Must Change Structurally

McKinsey identifies six enterprise disciplines that must evolve if AI is to scale. They are not new disciplines. They are existing ones designed for stable datasets and predictable workflows, now being applied to systems that retrieve, reason over and generate information in real time.

Observability. McKinsey argues that observability must extend beyond ingestion and pipeline health to context assembly, retrieval behavior, answer quality, citation integrity, content freshness and orchestration failures, and must cover search indexes, vector stores, APIs and newly generated artifacts. The point is that organizations must observe the full lifecycle through which outputs are assembled, generated and reused.

Data quality management. Quality must be maintained across source content, extraction, chunking, embeddings, retrieval and generation. McKinsey argues that quality now includes semantic integrity: outdated or superseded fragments must not continue to influence answers, updates must propagate predictably through indexes and embeddings, and quality must be tested at artifact level rather than only on the source document.

Metadata management. Metadata becomes a control layer for unstructured artifacts. McKinsey argues that it must make ownership, sensitivity, intent, allowed use, relationships and business entities explicit, and that it must apply to extracted objects rather than only to files and source documents. Unstructured artifacts should be connected to structured enterprise entities such as customers, contracts, products, assets, policies and transactions.

Data lineage. Lineage must track the source document and its version, extracted text, tables and images, segmentation and chunking, embeddings, retrieved fragments, prompt assembly, tools invoked, and relationships to structured data and business entities. McKinsey's point is that lineage shifts from tracing table transformations to tracing dynamic assembly across generated layers.

Governance and controls. Traditional storage-layer controls remain necessary but are no longer sufficient. McKinsey argues that governance must extend to retrieval, embeddings, prompts, memory, context assembly and generated outputs, and draws a distinction between compliant storage and compliant output. Sensitive information must be controlled while it is retrieved and generated, not only while the source file is stored. Our own concept of active governance, described later in this article, is our response to that same problem rather than a restatement of McKinsey's position.

Platform and tooling architectures. McKinsey calls for reusable and standardized foundations for extraction, parsing, chunking, embeddings, indexing, retrieval, guardrails, governed tools and reusable skills. The argument is not that tooling replaces governance, quality or ownership. It is that shared platforms allow those disciplines to be applied consistently at scale.

What Data Leaders Must Do Now

McKinsey closes with six concrete actions for chief data officers. Taken together, they describe a change in the job rather than only in the architecture.

One. Include structured and unstructured data in data products. Unstructured artifacts should be governed products rather than temporary pipeline outputs, with canonical schemas, entity alignment, quality thresholds that hold across transformations, artifact-level lineage, sensitive-content rules and consistent standards across AI applications.

Two. Establish shared foundation services. McKinsey argues for reusable enterprise services covering governed retrieval, tool and skill services, runtime policy enforcement, prompt and embedding controls, artifact-level lineage, observability and monitoring. Collectively these form a common control plane that individual applications consume rather than reinvent.

Three. Enable federated delivery on common infrastructure. Business units should be able to develop their own AI applications while using standardized extraction pipelines, shared metadata models, reusable embeddings, common guardrails and consistent infrastructure. Our interpretation is that this supports a federated model in which business units can innovate without recreating the underlying foundations or requiring every use case to begin with another enterprise-wide centralization program. This aligns with the Data Tiles no-rip-and-replace position.

Four. Manage derived artifacts as enterprise assets. Extracted objects, embeddings, indexes, summaries and generated representations should carry owners, versions, refresh cycles, audit trails, service expectations and retirement rules. McKinsey's argument is that they are not disposable technical byproducts.

Five. Govern semantic consistency across access paths and modalities. McKinsey's example is that a concept such as active customer should resolve consistently whether it is reached through SQL, warehouse analytics, keyword search, metadata search, vector search, semantic retrieval or an AI assistant. Achieving that requires linking structured records, unstructured artifacts, business entities and policy rules. Our interpretation, which we develop later, is that this is where AI readiness stops being a data question and becomes a question of shared business meaning.

Six. Measure readiness and reduce risk. McKinsey proposes four measures. Reuse: whether capabilities are created once and reused rather than rebuilt. Reliability: whether outputs remain accurate and traceable as content changes. Governance: whether controls operate where data-use decisions are made, including retrieval and generation. Scalability: whether growth lowers marginal time and cost while reducing duplication and complexity.

Where McKinsey and Data Tiles Align

Before setting out where we believe the signal extends further, it is worth being specific about where we agree, because the agreement is substantial.

We agree that data is a central constraint on scaling AI, and that structured and unstructured information must be governed together rather than run as separate programs. We agree that searchability is not usability, that AI-ready does not mean universally perfect data, that lineage must follow derived artifacts rather than stopping at the source file, and that governance must extend toward runtime and consumption. We also agree that data products are preferable to rebuilding for every use case, and that readiness should be measured through reuse, reliability, governance and scalability rather than through the volume of data made available.

Where We Believe the Signal Extends Further

McKinsey identifies the data-readiness requirement clearly and in useful technical detail. We believe the market signal extends into the operating model of the enterprise itself.

AI-ready data is not simply a state that an organization arrives at through a technical remediation project. It is an organizational capability, exercised continuously rather than achieved once. Enterprise information becomes genuinely AI-ready when the decision it is meant to support is clear, when business definitions are explicit, when policy is actively applied rather than merely documented, when ownership and lineage are visible, when quality is judged against the specific use case, and when the result can be reused by people, applications and AI alike.

It is worth drawing a distinction that is frequently blurred. Technically prepared data has been cleaned and made accessible, but it may still lack business meaning or governance context. Governed data has the controls applied. A reusable data product goes further, combining governed data with clear business meaning and ownership so that it can be consumed reliably across use cases. An organization can hold a great deal of technically prepared data and still be nowhere near AI-ready.

Follow that distinction far enough and it stops being a statement about data at all. It becomes a statement about the organization. An AI-ready organization is not one that has cleaned its data. It is one that can articulate what it means, agree on it across functions, govern it while it is being used, and make it available to any person or agent that needs it, repeatedly and without renegotiation. Data readiness is the visible symptom. Organizational readiness is the underlying condition.

Business understanding is the raw material of that condition, and most enterprises hold it in the least durable form imaginable. It lives in experienced people, in the unwritten rules of how a business actually works, in the distinction between the revenue figure the finance team uses and the one the sales team quotes. Humans reconcile these differences constantly and largely unconsciously. AI cannot. It will treat both figures as equally valid, and it will do so at a volume no review process can absorb. The market is therefore discovering that business understanding has to be captured, governed and made machine-consumable, which is a different project from any data program most enterprises have run before.

That is why we describe this as a change in operating model rather than another technology trend. Technology trends change what an enterprise buys. Operating model changes alter who is accountable for what, how work moves between functions, and what the organization treats as an asset worth maintaining. AI readiness meets that second definition. It reassigns responsibility for meaning to the business, relocates governance to the point of use, and turns reusable data products into shared enterprise assets rather than project outputs.

It also changes the relationship between engineering and the business. For two decades that relationship has been sequential: the business specifies, engineering builds, the business consumes, and any misunderstanding surfaces months later in a report nobody trusts. In a decision system, where agents act on context continuously, that latency is unaffordable. Meaning has to be encoded at the point of creation by the people who hold it, which is only possible if the operating model gives them a way to do so that does not require them to become engineers.

What accumulates from all of this is reusable enterprise knowledge. Every governed data product an organization creates is a piece of its own understanding, made explicit, owned and available for reuse. Enterprises that do not build this way keep rebuilding the same understanding, use case by use case, and mistake the resulting activity for progress.

AI-Ready Does Not Mean Perfect Data

It is worth being direct about a misunderstanding we encounter often. Some leaders interpret AI readiness as a mandate to cleanse, centralize or perfect every data asset in the enterprise before taking any meaningful action. This is neither realistic nor, in our view, correct.

Our position is that data must be fit for purpose, not universally perfect. Readiness should be assessed in relation to the specific decision it will inform, the use case it will support, the level of risk involved, who or what will be consuming the information, the policy environment surrounding it, the evidence required to defend the outcome, and the business value at stake. A low-risk internal recommendation engine does not require the same level of data perfection as a system informing regulatory reporting or customer-facing financial decisions.

This aligns closely with Gartner's recent work on AI-ready data. In What Is AI-Ready Data? And How to Get Yours There, Rita Sallam argues that AI-ready data is not simply high-quality data. It is data qualified for a specific use case, enriched with metadata, governed, and managed continuously rather than prepared once. Gartner's guidance on evaluating AI data readiness makes a similar point, and its February 2025 press release on the risk that a lack of AI-ready data poses to AI projects underlines how often organizations discover this only after an initiative has already stalled.

Gartner's broader research reinforces that AI readiness is not simply about improving data quality. Organizations delivering successful AI outcomes consistently invest more heavily in trusted data foundations, governance, context and organizational readiness than organizations struggling to scale AI. Read alongside McKinsey's work on reusable governed data foundations, the direction of travel becomes clear. AI readiness is becoming an organizational operating capability rather than a technology implementation exercise. AI succeeds when trusted business context becomes reusable across people, analytics and AI, rather than being recreated for every initiative. (Gartner, 2026; McKinsey & Company, 2026)

This supports an incremental, but governed, approach. Organizations do not need to perfect every data asset before generating value from AI. That is not an argument for ignoring foundational quality problems, nor a suggestion that disconnected prototypes create enterprise-wide trust. The point is that organizations can begin with valuable business decisions while applying enterprise-grade governance, context and evidence from the outset, rather than retrofitting governance once a pilot has succeeded.

The Operating Model Problem

To understand why AI-ready data remains elusive for so many organizations, it helps to look at how most enterprise operating models are actually structured. In a great many organizations, business intent, data engineering, governance, analytics, AI development and decision-making sit in largely separate functions, each with its own priorities, its own timelines and, often, its own language.

The consequences of this separation are familiar to almost anyone who has worked inside a large enterprise. Business teams wait for data that technical teams have not yet had time to prepare. Technical teams receive requirements that are incomplete because the business context behind them was never fully articulated. Governance is applied through a separate approval process that happens after the technical work is already underway, rather than as part of how the work is done. The same business definitions get debated repeatedly across different projects because no single version has been agreed upon and made available for reuse. Data sets are rebuilt from scratch for each new use case because there is no established mechanism for reusing what already exists. AI teams, under pressure to move quickly, create their own retrieval and preparation processes rather than working from a shared foundation. Context accumulated on one project is lost by the time the next project begins. And the same trust problem, the same questions about where the information came from and why it can be relied upon, reappears for every new AI agent the organization builds.

Our view is that the intelligent enterprise needs an operating model in which engineering, governance and business expertise function as a coordinated system rather than as separate stages in a relay race. Engineers should continue doing what they do best: building and operating enterprise platforms, pipelines and infrastructure. Business experts should contribute what only they can contribute: meaning, purpose, rules, context and a clear sense of the outcome the information is meant to support.

Engineering alone cannot solve this, and it is worth being precise about why. Engineering can guarantee that a number is delivered accurately. It cannot determine whether that number is the right one for the decision at hand, whether the definition behind it matches the one used in the board pack, or whether the policy governing its use has changed since the pipeline was built. Those are judgments that live in the business. When the operating model gives business experts no structured way to encode them, they are made informally, inconsistently, or not at all, and AI then scales that inconsistency at machine speed.

This is why business ownership matters more than it did in the reporting era. A human analyst receiving a questionable figure applies skepticism. An agent does not. Ownership is the mechanism by which someone accountable for an outcome also becomes accountable for the meaning of the information supporting it, and that accountability is what makes trust auditable rather than assumed.

Governance has to become active for the same reason. Documented governance describes intent. Active governance applies policy at the point of consumption, which becomes a critical moment of control once retrieval and generation happen dynamically. Governance that is enforced only at onboarding, review or ingestion is invisible to an agent assembling an answer three steps downstream.

Reusable data products then become enterprise operating assets rather than project deliverables. Every reuse avoids repeating the definition debate, the governance review and the preparation work, and the cost of the next decision falls. Data products, in this model, become the governed interface between the enterprise information environment and the decisions, applications and AI systems that need to use it.

Taken to its conclusion, this is what it means to say that trusted business context is becoming enterprise infrastructure. Not infrastructure in the sense of servers or storage, but in the sense that it is shared, funded, maintained and depended upon by everything built on top of it.

The Signal We Are Seeing: AI-Ready Data Is Becoming the Enterprise Operating Model

We believe the next wave of enterprise differentiation will not be determined primarily by which models an organization has access to. Model capability is converging quickly, and competitors will have access to broadly comparable models within a similar timeframe. Advantage will instead come from the quality of enterprise context available to those models, the speed at which that context can be assembled for a new decision, the ability to enforce policy at the moment information is used, and the ability of people and agents to work from the same understanding of the business.

The first generation of enterprise AI taught organizations how to build intelligent agents.

The next generation must teach them how to build intelligent organizations, in which every agent, and every person, shares the same trusted understanding of the business. That, more than any individual model or platform choice, is the capability that will separate the organizations that scale AI from those that remain permanently stuck at the pilot stage.

What Enterprise Leaders Need to Build Next

If AI readiness is an organizational property, the response cannot begin and end with the technology roadmap. Enterprise leaders need to decide how business meaning will be created, who will own it, how it will be governed, and how it will be reused across decisions, applications and AI. That requires a different management system, not simply a better data platform.

The first requirement is a shared language for the business. Every enterprise contains competing definitions of customers, revenue, performance, risk and value. People have historically reconciled those differences through experience, conversation and judgment. AI cannot depend on that informal process. Leaders must make the meaning of the business explicit enough to be shared across functions and durable enough to be reused by machines.

The second requirement is a new working relationship between business and technology.

Business teams cannot remain passive consumers who submit requests and wait for technical teams to interpret them. Equally, engineers should not be expected to infer purpose, policy and meaning from incomplete requirements. The intelligent enterprise requires a continuous relationship in which engineering provides reliable platforms and integration, while business experts actively define context, ownership, appropriate use and intended outcomes.

Governance must also change position. It cannot remain an approval stage that sits before or after delivery. It must travel with information and remain active wherever that information is used, including dashboards, applications, APIs and agents. A policy recorded in a catalog but absent from the moment of retrieval is not governing an AI system. It is documenting an intention the system may never see.

Leaders must then make reuse an explicit enterprise discipline. Every governed definition, policy and data product should reduce the work required by the next decision and the next AI agent.

Finally, organizations need a better measure of progress. The number of connected sources, prepared datasets, pilots or agents does not demonstrate AI readiness. The more meaningful measure is whether the enterprise can make important decisions faster, with consistent meaning, visible evidence and appropriate control. Decisions, not data volume, are the unit of value.

If McKinsey's paper identifies what enterprises need to build, the obvious next question is how that operating model can be implemented in practice. This is where our own work begins. Rather than viewing AI readiness as a technology initiative, we have focused on how business teams and engineering teams can jointly create trusted business data products that become reusable enterprise assets.

What Data Tiles Is Building in This Space

We did not begin by designing a product. We began with a question about operating models. If trusted business context is going to become enterprise infrastructure, what has to exist so that context can be created once, governed continuously and reused by every person, application and agent that needs it? Traditional enterprise architectures were built to move and store data rather than to carry meaning, ownership and policy alongside it.

Latttice is our Data Product Workbench. It is designed to work with the technology and data platforms organizations already own, rather than requiring every source of information to be copied into a single new centralized platform before any value can begin. That choice matters to the operating model argument: an approach that depends on centralizing the enterprise before it can produce trusted context simply relocates the problem. Latttice instead enables business and data teams to work together to create trusted, governed, fit-for-purpose data products, connecting enterprise data to business purpose, meaning, governance and consumption.

Business data products sit at the heart of that model because they are business-led and jointly governed, and because they provide a practical mechanism for carrying business meaning, ownership, policy and evidence together in a form that survives reuse. A dataset can be copied. A dashboard can be rebuilt. A data product, when it is genuinely business-owned, carries the reasoning behind it, which is what allows a second team, or a second agent, to use it without starting the interpretation work again.

Rather than treating governance as documentation produced after a data product has already been built, Latttice activates governance during both the creation of a data product and its ongoing use. That is what we mean by operationalizing governed business context: definitions, sensitivity, ownership and policy are not sitting in a catalog describing the data product, they are part of how it behaves when it is consumed.

This is also how trusted business data supports AI delivery. When the information required for a decision is already governed, fit for purpose and linked to lineage and evidence, a major source of AI delivery risk has been addressed before the first agent is built. When people can rely on a data product for a decision, AI can be given access to the same governed context. That does not eliminate every model, workflow or adoption challenge, but it does address one of the most persistent barriers to reliable enterprise AI.

Lenz is our AI Factory. Lenz enables organizations to build AI agents on top of the governed business context that trusted data products provide, so that the agent inherits the meaning and the policy rather than reconstructing them through prompt engineering. Put simply, Latttice creates and governs the trusted business context, and Lenz enables AI to reason and act using that context. The separation is deliberate. Context should outlive any individual agent, and agents should be replaceable without the enterprise losing its understanding of itself.

We want to be clear that we do not believe Data Tiles has solved every aspect of the AI data readiness challenge that McKinsey describes. What we believe is that Latttice and Lenz offer a practical way of operationalizing important parts of the operating model this article describes, particularly the parts concerned with connecting business meaning to governed, reusable data.

The AI Operating Model

Together, Latttice and Lenz support what we describe internally as an AI operating model built around trusted, fit-for-purpose data products. The cycle behind that model begins with the identification of a specific business decision or outcome the organization wants to improve. Relevant enterprise information, structured and unstructured, is then connected to that outcome. Data products may be created from raw, curated, governed or otherwise prepared enterprise sources. Business context and definitions are applied so that the information means the same thing to everyone who uses it. Governance and access policies are activated as part of this process rather than added afterward. A reusable data product is created from this combination of data, meaning and governance, and published so that it can be consumed appropriately by the people, applications, analytics tools and AI agents that need it. Usage and outcomes generate evidence that can be used to improve the product over time, so that the product evolves as needs change rather than being discarded and rebuilt from scratch for the next use case.

This cycle connects directly to the broader argument we have made throughout this article. An enterprise that can move through this cycle repeatedly, and can reuse what it builds each time, is an enterprise that is genuinely becoming AI-ready, in the fullest sense of that term, not merely in the narrower technical sense that a data readiness audit might measure.

A Broader Market Movement

It is worth stepping back to note that McKinsey is not making this argument alone. McKinsey approaches the issue through AI data readiness. Gartner approaches it through use-case-specific AI-ready data, metadata and governance. IDC approaches it through enterprise adoption and operationalization. Their terminology, scope and emphasis differ, and none of these sources formally endorses the others. Read together, however, they point in a similar direction, and that convergence supports our own interpretation: enterprise AI progress now depends far less on selecting better models and far more on building trusted, governed, reusable enterprise information and the operating models that sustain it.

That convergence is precisely why we publish this series. Individually, each of these publications is a data point. Read together, they describe a market movement, and it is the movement rather than any single paper that enterprise leaders should be planning around.

Conclusion

McKinsey's research makes an important and well-evidenced case that data readiness is a central constraint on scaling enterprise AI. We agree with that conclusion, and we believe the evidence McKinsey presents deserves serious attention from every executive responsible for an AI program.

Our contribution to this conversation is to argue that AI-ready data is not a preliminary technical task that gets completed before the real work of AI begins. It is a living enterprise capability, one that must be created, maintained and improved continuously by business and technical teams working together. The organizations that succeed in scaling AI will not be the ones that complete a single data readiness project. They will be the ones that build a durable operating model through which trusted business context can be continuously created, governed, reused and improved.

The intelligent enterprise will not be defined by how many models it deploys. It will be defined by whether its people and its AI share a trusted understanding of the business.

That, in the end, is the market signal we believe McKinsey's research points to.

The question is no longer whether AI will transform the enterprise. It already is. The question is whether your organization has an operating model capable of giving every person, and every AI agent, the same trusted understanding of the business.