The Context Advantage

When companies upgrade their AI models, something curious happens. The model becomes smarter, benchmarks show improvement, and demonstrations become increasingly impressive. Yet, the actual business results sometimes barely shift.

For instance, the AI might still pose questions that the organisation had already addressed months ago. It may suggest actions that are against company policy, treat a strategic client similarly to a regular one, or recommend a setup that the engineering team dismissed long ago. The root of the issue isn’t necessarily a lack of intelligence. It’s a matter of not understanding the business context.

This distinction is vital because the most advanced AI models are now accessible to nearly everyone. Your rivals can access the same cutting-edge intelligence that you can. However, they cannot tap into your specific customer history, business knowledge, organisational memory, or grasp your decision-making processes.

AI models are becoming common, but context remains unique. This is potentially where the next competitive advantage in enterprise AI can emerge.

Let’s consider a straightforward enterprise inquiry: Can we offer this customer a larger discount?

An advanced model can analyse discount options, but to answer that question correctly, it may need to know the customer’s lifetime value, existing contract details, payment history, profit margins, approval thresholds, and perhaps insights from a recent conversation where the account manager learned that a competitor had engaged in the deal.

It’s not that the model lacks intelligence; rather, it lacks context.

Humans usually don’t make significant decisions based on a single piece of information. We tend to synthesise a myriad of data, experiences, relationships, history, and situational awareness, often without realising it. AI agents should possess a similar capability.

This is why enterprise AI is shifting focus from an era emphasising prompt engineering to one increasingly prioritising context engineering. Satya Nadella recently described this new marketplace as simply, “a world of context engineering.”

Prompt engineering focuses on the question: What should I communicate to the model?

On the other hand, context engineering delves deeper: What must the model comprehend before it can take effective action?

This distinction becomes crucial as AI transitions from merely answering questions to executing tasks.

In conversations about enterprise agents, “context” often gets reduced to merely giving a model access to more data. This definition is overly simplistic.

I like to view enterprise context as a layered structure: DATA → SEMANTICS → KNOWLEDGE → WORK → MEMORY → JUDGMENT

Data informs the agent about past events: orders, transactions, inventory, claims, and customer records.

Semantics provides insight into what that data signifies for the business: defining a strategic customer, calculating margins, determining asset allocations to production lines.

Knowledge includes what the organisation knows: policies, contracts, procedures, research, and accumulated expertise.

Work describes the current situation: who is involved, discussions that have taken place, and existing commitments.

Memory recalls previous events: past interactions, decisions made, exceptions, and outcomes.

Judgment helps in deciding the next steps.

Memory is particularly important as it allows agents to accumulate context through experiences. Microsoft Foundry now distinguishes between session, user, and procedural memory, enabling agents to go beyond just remembering conversations, but instead remembering preferences and learning how tasks are accomplished over repeated interactions.

Sometimes, improving an agent noticeably doesn’t require a larger model. A enhanced memory might suffice.

While not every agent needs every layer, the more significant its responsibilities, the richer and more carefully managed its context must typically be. This transition recognizes that context evolves from a mere retrieval challenge to an essential aspect of enterprise architecture.

This transition is now evident beyond a single technological platform. Gartner is increasingly identifying a context layer as a critical architectural foundation for agent-based AI, integrating semantics, operational states, provenance, and other enterprise signals to ensure reliable agent actions. Meanwhile, Anthropic, focusing on AI engineering, argues that context should be seen as a finite resource that needs careful management rather than simple expansion.

Microsoft’s latest direction is especially noteworthy through this lens.

Work IQ, Fabric IQ, Foundry IQ, and Web IQ are progressively forming a shared intelligence layer across Microsoft’s offerings.

Work IQ sheds light on how work is undertaken, while Fabric IQ clarifies business semantics and relationships. Foundry IQ offers reusable enterprise knowledge and agent retrieval capabilities, and Web IQ enhances this knowledge with up-to-date external insights.

The architectural implications of this shift are more significant than the product names. Rather than having each new agent rebuild connections, retrieval processes, semantic definitions, and permissions, context becomes a reusable foundation that numerous agents can draw from.

Microsoft captured this change particularly well: As agents become the new application model, context is critical infrastructure.

For years, businesses have been told that their data presents a competitive edge. Sometimes it does.

However, simply having data and effectively utilising that data with an agent are fundamentally different matters. Many enterprises aren’t short on information; they are overwhelmed by it — scattered across databases, documents, emails, meetings, applications, and people’s memories. The benefit doesn’t stem from linking an agent to everything. It comes from transforming information into usable context: maintaining business significance, understanding relationships, distinguishing reliable knowledge from convenient facts, respecting permissions, and retrieving relevant data when decisions are made.

The enterprise with the most data might not create the best AI. The one that turns its context into a usable format might.

There’s also a pitfall to avoid. If enhancing AI capabilities depends on context, the instinct is to provide the agent with as much context as possible. This aim is misguided.

Context requires attention and computational resources. Irrelevant information clutters the process. Outdated information can lead to mistakes. Conflicting context may impede reasoning rather than enhance it.

The goal should not be maximum context, but rather minimum sufficient context — the smallest, most trusted, relevant, and permission-aware collection of information needed to yield the correct result.

In Beyond Tokens, I presented the concept of an Intelligence Budget: How much intelligence is required for this outcome?

Similarly, enterprise agents increasingly need a companion concept: The Context Budget. How much context is necessary for this decision?

A scheduling agent may need access to calendars and preferences, while a pricing agent requires customer economics, margins, commitments, and policy approvals. A clinical agent demands a completely different level of detail, accuracy, and governance. Every unnecessary piece of context incurs a cost, and every missing essential piece introduces a risk.

This gives us a valuable framework for considering enterprise agents: Right Intelligence × Right Context → Right Outcome

Allocate more intelligence where outcomes justify it. Provide more context where decisions demand it. Avoid doing either indiscriminately.

Another significant aspect of the Context Budget is permissions.

Enterprise context belongs to individuals. An agent that retrieves the correct answer while disregarding whether it has permission to access that information fails to function intelligently; it creates a security concern.

Therefore, enterprise context must stay identity-aware and permission-aware as it transitions from source systems to agent reasoning, which is especially crucial as agents develop persistent identities and greater autonomy.

An agent’s permissible knowledge is just as important as what it is able to do.

Autonomy defines the scope of action, and context delineates the boundaries of understanding.

The initial generation of enterprise GenAI often centered on RAG: linking documents to a model to extract pertinent excerpts. While this represented a significant advancement, an agent-driven enterprise cannot merely consist of numerous independent RAG systems, each rediscovering the organisation’s customers, products, policies, vocabulary, and permissions. That risks further technology fragmentation.

The opportunity lies in shifting from agent-specific retrieval to shared enterprise context.

Which knowledge should agents inherit? How should business entities be uniformly defined? Which sources are authoritative? What memories should last, and which should be discarded?

And perhaps the most challenging question: Who owns enterprise context?

At present, accountability is dispersed among data teams, application teams, knowledge management, security, business functions, and increasingly, AI teams. In an agent-driven enterprise, context may need to be handled as a product, rather than a side task for many.

Imagine two competing companies using precisely the same state-of-the-art model. One links it to disconnected documents and a handful of applications, while the other possesses robust data foundations, shared business semantics, governed knowledge, interconnected workflows, and organisational memory. Even with identical access to AI, their intelligence systems diverge significantly.

Agents in the second organisation have a more profound understanding of customers, better interpret signals, and act with greater awareness of the organisational culture. This advantage compounds over time. Every workflow generates more context. Every interaction enriches organisational memory. Each decision forms an additional signal for future agents to utilise.

As new models appear, they lose value. Context, however, gains worth with each day of operation.

This insight might be the most important strategic shift. Anyone can lease intelligence, but no one can lease your context.

The first wave of enterprise AI has demonstrated the potential of powerful models. The next phase might reveal that these models are just one piece of the puzzle.

An agent requires intelligence to reason, but it needs context to comprehend your business.

As frontier intelligence becomes increasingly attainable, enterprises may realise their most sustainable AI asset was never the model itself. It’s everything the organisation learned before the model arrived: its data, semantics, knowledge, relationships, history, workflows, and judgement. Those organisations that can transform these assets into trusted, reusable context will give every agent they create an advantage from the start.

So perhaps the strategic question has evolved from merely: Which model should we use?

To: What does our organisation know that our competitors don’t, and can our AI effectively utilise that knowledge?

Because, in a world where everyone can access intelligence, context may be what makes that intelligence truly unique to you.

Microsoft — FY2026 Q2 Earnings Conference Call

Microsoft — FY2026 Q4 Earnings Conference Call

Microsoft Learn — What is Foundry IQ?

Microsoft Foundry — Foundry IQ: Build Smarter Agents Faster with Unified Knowledge and Serverless Retrieval