For the past few years, the AI race has been defined by scale. Bigger models, larger context windows, more computing power and access to more data have promised better performance.
Now, as businesses move from experimenting with generative AI to deploying agents inside real workflows, I think the conversation needs to change. The next bottleneck is whether we can give AI the right context to make a decision, at the right moment, while clearly defining what it is allowed to do.
Enterprise decisions depend on relationships, history, current circumstances, policies and business rules. The question is becoming less about how much data we can give a model and more about what that information means for this decision.
More data is more useful when it is relevant
Enterprise data is abundant. The difficulty is making it relevant. A large organisation might have information about a customer spread across a CRM platform, billing system, service application, risk database and many other systems. Each record can be accurate in isolation, while the organisation still struggles to understand the person represented across those records.
An AI agent encounters the same problem. An agent reviewing a customer, claim, transaction or supplier needs more than a collection of records. It needs to understand that several records represent the same entity, how that entity connects to others, what has happened previously and which rules apply.
This is why I see the enterprise AI challenge shifting from data quality towards data relevance. Good quality data remains essential. Yet accurate individual records do not automatically create understanding. The scarce resource is the ability to identify the information that actually matters for a particular decision.
Gartner is now identifying the need for a context layer around AI agents, with semantics, operational state and provenance among its core components. In practice, this means establishing knowledge graphs to define relationships, based on ontologies for semantics, providing right-time access, providing right-time access to live business conditions (operational state), and maintaining full lineage tracking so every decision remains auditable (provenance).
The context window is different from real-world context
A context window tells us how much information a model can process. However, real-world context tells the model what that information means.
Consider a financial-services decision. Knowing that two customers share an address is one fact. Understanding whether they are members of the same household, directors of related companies, connected through previous transactions or entirely unrelated is something else.
That requires relationships, identity, history and business meaning. Context is therefore information organised around the reality an AI system is being asked to understand.
The future may belong to smaller agents
The prevailing narrative around agents is about increasing autonomy. Give an agent more tools, more capabilities and more responsibility, then let it complete increasingly complex tasks independently.
For high-stakes enterprise decisions, I think we should consider a different direction. To scale AI, we may need to shrink what each agent is responsible for.
A complex business decision can be broken into smaller steps. Some can be handled by deterministic software, some by rules, some by people and some by AI agents. An agent responsible for one tightly defined task can be given exactly the context it needs and constrained by explicit rules about what it can and cannot do.
Context is facts, relationships, rules and relevance
For an AI agent to make a useful decision, knowing the facts is only the beginning. Imagine an agent assessing an insurance claim. It may know the claimant, policy, claim history and circumstances surrounding an incident. It also needs to understand the relationships between those entities and the rules governing the decision.
Which policy applies? Which jurisdiction matters? What thresholds exist? Which procedures are mandatory? What information is current? What evidence supports the conclusion? I think of context as: facts + relationships + rules + relevance. Those components also need to be governed.
If context determines an AI-enabled decision, organisations need to understand where it came from, how current it is, how entities were connected, which rules were applied and why particular information was presented to the agent.
AI governance therefore increasingly needs to govern the context the model is allowed to act upon.
Context should become shared infrastructure
I do not think every AI project should build its own version of context. If an organisation has established a trusted understanding of its customers, suppliers, products, assets and relationships, that understanding should be reusable.
The same contextual foundation can support analytics, applications, operational processes and AI agents. Build the context once. Keep it current. Govern it properly. Reuse it wherever decisions are made. That makes context a shared enterprise capability rather than another AI feature.
Moving beyond AI theatre
The rush towards AI has created a temptation to demonstrate activity rather than outcomes. Organisations can announce pilots, launch copilots and deploy agents while still struggling to answer a basic question: what meaningful business decision has actually improved?
McKinsey’s 2026 research shows widespread AI adoption, while enterprise-wide scaling remains a work in progress. Adoption alone is therefore a poor measure of value.
A more useful approach is decision-centric. Start with a decision where better context could change the outcome and work backwards. Then give an AI system responsibility for the part it can perform reliably. That may feel less dramatic than building a super-agent. It may also produce more meaningful results.
The next advantage will be contextual
Foundation models and cloud infrastructure are becoming increasingly accessible. Organisations will continue to gain access to powerful models, sophisticated agent frameworks and enormous amounts of enterprise data.
The differentiator will be what those systems understand about the organisation itself. The advantage will come from delivering that unique web of entities, relationships, and rules at the exact moment a decision needs to be made.
The next generation of enterprise AI may therefore be defined less by who has the biggest model and more by who can give their models the right context, constrain their actions and make their decisions explainable.
For me, that is the real shift taking place. The race is moving from intelligence alone towards contextual intelligence.
Dan Onions
Dan Onions is Global SVP Data & AI at Quantexa. With over 27 years of experience helping organisations deliver business, data, and AI transformation, Dan combines deep technical expertise with a pragmatic understanding of organisational change, enabling enterprises to modernise technology platforms, build trusted data foundations, and adopt AI at scale to deliver measurable business outcomes.
