Thought Leaders
Context: Building a Trustworthy Data Foundation for Agentic AI
Add Unite.AI to your preferred sources on Google
Context is key to how we communicate. Remarks taken out of context can appear to mean the exact opposite of what was intended. Meet someone out of context, and we can struggle to recognize them. Context is the glue that stitches together human relationships and allows us to find meaning in our social interactions.
The same is true of AI agents. Context is what enables agents to remove ambiguity – to tell, for example, the difference between Apple (the technology company), Apple (the record label) and apple (a fruit).
This disambiguation is crucial because enterprise AI makes decisions and acts at lightning speed; flagging transactions, approving credit, and triggering alerts. If agents don’t understand the entities involved in their workflows, they will continue to act, but you can bet your bottom dollar that they’ll make mistakes.
Given that some estimates suggest a typical large enterprise will use as many as 1,600 agents by the end of this year, these mistakes will quickly scale.
The Complexities of Resolving Digital Identities
Agentic AI is rightly causing much excitement, but until agents can truly understand the entities in their datasets we risk building castles on sand.
This isn’t an easy problem to solve. A single customer, for example, can appear in various enterprise databases listed in different ways (J. Bloggs, Jane Bloggs, Bloggs, etc.). Similarly, a single corporate entity might be flagged separately across its several subsidiaries. How can agents know that these seemingly different entities in fact relate to a single profile?
At present they often can’t. This is a result of the way agents are built. An agent is, at core, a system that takes inputs and returns outputs. But how much data, exactly, should we feed into agents? Research shows that overly long context windows cause agents to hallucinate. Conversely, provide too little data and the agent will make a confident decision and act on inaccurate information. Enterprises need to tread a fine line, offering no more nor no less context than the system requires.
Data Integration Is Key
Data volume is only part of the challenge. Enterprises also need to consider data quality and integration. Bad context comes from poorly connected data. Hallucination can occur when the data that an agent needs to come to a decision is fragmented across a range of siloed databases, apps, and business units. Approximately 80-90% of business context lies in siloed and unstructured files across disconnected systems. Without access to this information, AI outputs will be based on incomplete assumptions.
Despite all these issues, the deployment of AI agents continues at pace. If business leaders cannot trust their agents to reach accurate decisions and act correctly, they will have to invest in greater levels of human oversight than would otherwise be necessary. Rather than the autonomous, highly efficient agents they’d hoped for, enterprises will be left with untrustworthy tools they need to babysit. The alternative is to increase business risk and the workloads of compliance and legal teams.
Building a Context Data Foundation
So what needs to be done? The key is to arm agents with an accurate and verified data foundation. This foundation unifies fragmented data across internal and external sources and applies consistent governance across structured and unstructured data. From there, it adds context so that the data being fed into the AI reflects real-world relationships rather than isolated records. This allows agents to connect a person to their aliases, or a subsidiary to its parent.
With a context foundation in place, AI-driven operations become accurate, explainable, and capable of supporting confident decisions. Enterprises can trust their agents, rather than having to babysit them, enabling them to enjoy the full benefits of agentic automation without the need for excessive oversight.
From a technical perspective, context is achieved through two capabilities. The first is entity resolution. This connects records that relate to the same entity where these references are scattered across various systems and data types. The records are harmonized in a single and reliable view of people, organizations, and assets. The duplication and ambiguity that often erodes data quality and leads to erroneous AI outcomes is solved.
The second capability is knowledge graphs. These add context by modeling the relationships between entities: how customers connect to accounts, how transactions relate to behavior, and how risks propagate across networks. Combined with entity resolution, knowledge graphs build an accurate picture of data as it exists in the real world, providing AI agents with the foundation they need to operate successfully at scale.
Benefits Beyond Agentic AI
Of course, AI agents are not the only consumers of data in the enterprise. A context data foundation ensures that operational staff, compliance teams, business analysts, HR teams, and all other data users similarly benefit from a unified and accurate view of the world.
In turn, this can help to elevate the customer experience. For instance, if a business maintains multiple profiles of the same customer, that can lead to them being contacted on multiple occasions by different teams. Such experiences can leave customers feeling frustrated to say the least. Context-fuelled data boosts the customer experience as it enables organizations to understand their customers in a more complete way, and ensure that they’re engaged with correctly. This is ultimately a more human-centric approach, one that can save businesses time and money.
Keeping AI Honest
Gartner predicts that 40% of enterprise apps will feature task-specific AI agents this year, up from less than 5% in 2025. The agentic AI era is upon us.
However, the full benefits of agentic AI workflows will not be enjoyed equally by all who invest in the technology. Those organizations that prepare the ground by ensuring a unified, context-rich data foundation will be able to trust AI outcomes and move to scale at pace. The rest will waste time and effort babysitting their new systems. Success therefore comes down to embedding a living context layer into the agentic tech stack from day one.
英文原文如下；中文版正在随流水线自动整理，稍后刷新即可。