Sean Nathaniel is CEO of Upland, a provider of enterprise intelligence that turns knowledge, content, and data into AI-powered outcomes.
In most of my conversations with other executives about AI, two discussions are happening in two different places. In one, the business is asking why its AI investment hasn’t produced the return it expected. In the other, security and legal are asking what that same AI can see and whom it might show it to. Rarely are the same people in both rooms.
They should be, because they’re asking about the same thing.
The common line is that data is the fuel for AI. While that’s true, fuel doesn’t power anything until it’s refined. For AI, context is the refinery: the layer that turns raw knowledge, content and data into answers that are accurate and appropriate to share.
I’ve written before about the three asset classes enterprises feed AI (knowledge, content and data) and the semantic and context layer that connects them. This article focuses on that layer itself, and why it determines both the return an organization gets from AI and the security of the information AI touches.
Why The ROI Problem And The Security Problem Share A Root
Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. IBM’s 2025 Cost of a Data Breach Report found that among organizations reporting a breach of their AI models or applications, 97% lacked proper AI access controls.
I’d read those two findings together. One is about value, and the other is about risk, but both trace back to the same gap: AI that can reach information without understanding what it means, who it’s for or when it applies. When that understanding is missing, AI produces answers that are wrong, answers that shouldn’t have been shared or both. Organizations usually respond with separate fixes owned by separate teams, when a single foundation would address both.
What Context Actually Is
Context is the information about your information. In practice, it comes in three layers.
The first is meaning. This is the metadata that sits on top of the data and tells AI what it’s looking at: what type of asset it is, who authored it, when it was last validated, which domain it belongs to and how it relates to other assets. A contract is just text until metadata tells AI which customer it covers and which policy governs it.
The second is permission. This is the security that tells AI who can access information and how it can be used. Most enterprises have spent years building access controls inside their source systems, but those controls were designed for applications, not for AI. When content is indexed and retrieved for an AI assistant, its permissions often stay behind in the original system. The document is still locked down where it lives, while its contents are available to anyone who asks the right question. Governance has to travel with the information itself, as part of its context.
The third is relevance. These are the rules and conditions that dictate when information applies: effective dates, version history, jurisdiction, customer segment and which policy supersedes which. A pricing guideline that’s accurate in one region and wrong in another is a relevance problem, not a data quality problem.
When one layer is missing, the others can’t compensate. Metadata without permissions produces accurate answers delivered to the wrong people. Permissions without relevance rules produce secure answers that are out of date.
We’ve seen this play out directly. A global law firm rolled out Microsoft Copilot across its Microsoft 365 environment, but its most valuable material (case files, contracts and research) lived in its iManage document management system, out of Copilot’s reach. Attorneys couldn’t ask something as basic as which version of a contract was current. Off-the-shelf plugins could connect the systems but lacked the semantic indexing needed for accurate answers. In a law firm, access can’t be loosened for convenience; client confidentiality depends on who sees what. The firm connected the two systems through an indexing approach that carried document context and security permissions with the content itself. Attorneys got answers drawn from both environments while confidentiality controls stayed intact, and adoption followed.
Four Steps To Building The Context Layer
1. Test your permissions from the user’s side. Trace what happens to access controls as content moves from source systems into your AI environment. Then query the AI as users in different roles, including roles that should be restricted, and document what comes back. The gaps usually appear where content was copied into an index without its permissions.
2. Set a minimum context standard for AI access. Before content comes within AI’s reach, require a classification, an owner, an access policy and a last-validated date. Content that doesn’t meet the standard waits until it does. Setting this bar forces the enrichment work most organizations defer, and AI-assisted classification can apply it at scale.
3. Document the rules that determine relevance. Effective dates, superseded versions, regional or regulatory scope and intended audience often live in people’s heads or in system configurations. Capture them as metadata AI can read, starting with the domains where a wrong answer is most costly.
4. Give the context layer a single owner. When AI value sits with the business and AI risk sits with security, the layer that serves both tends to go unfunded. Assign an accountable owner and track answer accuracy alongside access exceptions on a shared scorecard.
Where The Return Comes From
Most enterprises have access to the same AI models. What separates their results is the context those models run on: what an organization’s knowledge, content and data mean, who is allowed to see them and when they apply. That context is specific to each business and takes years to build, which is why it can’t be purchased off the shelf.
Data is still the fuel. The organizations that treat context as the foundation, rather than an afterthought, are the ones whose AI earns the trust of both the employees using it and the leaders responsible for protecting the business.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?