Is your Enterprise Architecture Data Agent-Ready?

As organizations mobilize to integrate AI agents into enterprise workflows a new battleground shapes up: The data layer itself.

Agents may represent the next frontier of business automation, but the question is how to make them experts in your business. There’s an emerging consensus that the answer lies in how you engineer your data.  

It’s this consensus that is driving both Orbus’ current development and our active research.

Now, everyone knows data matters: A cliché during the first great wave of enterprise digitization was that ‘data is the new oil’. Just as oil’s value is linked to its purity, abundance and flow, data’s is linked to its hygiene, access and lineage.

But these simple analogies we use to communicate abstractions can easily become thinking traps. If the extent of your thinking on the data’s role in agentic business is that it ‘should be accurate’, then you’re already behind.

Because while raw data may resemble an amorphous liquid made of ones and zeros flowing along your organizational pipework, information - which is what agents need - is not liquid but solid. It has structure, and it has behavior.

It has these qualities because it is a model of the world and mirrors its qualities. And when the world in question is your organization, information’s structure is what enables AI agents to comprehend, navigate and reason about that organization.

Nor is it just a static input, but a learning resource where decisions and outcomes can be written back to it to refine the next round of autonomous decision-making.

Don’t despair if you’re still at the beginning of your agentic information journey - we all are. The gate hasn’t closed yet, but now is the time to position yourself for success.

A good way to start is by familiarizing yourself with the information capabilities that will platform the next wave of business transformation. Because while the intelligence of the AI foundation models is the preserve of the AI Labs and the hyper-scalers, the intelligence of your organizational knowledge platform is something you have to engineer.

To get you started, we’ll look at three data capabilities we believe will dominate agentic deployments in coming years, and the six qualities they bring.

• Enterprise knowledge graphs - for truth and situational awareness

• Enterprise digital twins - for performance and action

• Agentic context graphs - for trust and learning

Individually each is important; but collectively they make up a profound organizational capability, and one you can’t easily buy: A learning model of your organization.

Knowledge Graphs

First, some disambiguation: A graph in data terms is not a chart, but a data structure where real-world entities and concepts – like a person, or a contract – are each represented by nodes, and the relationships between nodes – like a person agreeing a contract - by edges.  

Graph databases became prominent with the rise of social media, where their network structure mapped far more closely to those webs of social interactions than traditional table-based data stores. And as it turns out, graphs are also ideally suited to modeling the webs of dependency between processes, applications, data and teams that make up the enterprise’s operating model.

It’s no accident then, that graphs have seen a steady acceleration with the rise of generative AI particularly with techniques known as GraphRAG or KG-RAG (Knowledge-Graph Retrieval-Augmented Generation). To understand why, we need to look at the qualities that make them ideal platforms for AI agents; and particularly to understand the significance of the enterprise knowledge graph, or EKG.

Knowledge Graphs for Ground Truth

For all its astonishing capability, one the biggest threats of generative AI isn’t that it runs rogue, but that it steadily erodes our notions of truth. Businesses need trusted facts, both to run their own operations and to demonstrate compliance to external regulations.

But LLMs are probabilistic. They have no ‘world model’, just a series of weights representing more or less probable associations. The word ‘hallucination’ is itself a hallucination – it suggests AI models sometimes get it wrong, when in reality they have no concept of ‘right’ or ‘wrong’ to start with. Everything is a guess, but we only see it when their guesses don’t match reality.

Graph data structures can represent objective facts, not probabilities, as trusted inputs into AI agents. How they do it is simple but powerful.

Philosophers from Leibnitz to Wittgenstein who asked ‘how can me make statements that are true?’ built their foundations on propositions with a subject-predicate-object structure: Mary owns the cat; the cat is in the house. In the same way, nodes in a graph make statements about things: – Mary; the cat; the house - and edges, the relationships between them.

These relationships are represented by deterministic key values – very different to an LLM or vector database where they’re probabilistic weightings.

For organizations more concerned with operations, investment and governance than cats and houses we have to move beyond these simple graphs to the Enterprise Knowledge graph.  

Here, every node and edge is given a type - a business process, an application, a team – so we can use the same principle to make true statements about the enterprise operating model itself: The Accounts Receivable team uses the SAP ERP application; or the SAP ERP application processes billing data.  

These are the true facts about business and IT operations AI agents need to make informed actions about the enterprise, instead of trying to infer those facts from probabilistic associations in documents.

Knowledge Graphs for Situational Awareness

Graphs matter, not only to represent ground truth but also organizational context.  

The enterprise is a machine; its parts interact daily to deliver business and customer outcomes. No part of it exists in isolation - a change to one part can have knock-on effects on others.

Now, we understand this principle instinctively because we inhabit that context. But how can agents with no understanding of that operating model make decisions that work not just inside their own their own context but within their ecosystem?

The answer is, extend their context.  

The relationships in the enterprise knowledge graph do that – each agentic operation, whether it’s against an application or a process or a project, can be contextualized within the ‘halo’ of its surrounding operating model context, taking account of the service, risk, cost or time-based dependencies that might otherwise cause critical breakages from siloed decisions.  

The net result is situational awareness.

Digital Twin

Enterprise knowledge graphs are a huge enabler for agentic processes that are both grounded in fact and aware of their surroundings.  

But that graph is also a fairly static structure. It describes the enterprise’s composition, not so much its day-to-day operations. It tells us what the enterprise is, but not how it acts.

Digital Twin is an emerging pattern that adds another layer on top of that description: The enterprise’s minute-by-minute behavior: Business transactions; system outages; team resource profiles – and most critically, key performance indicators (KPIs).

it isn’t a separate graph as much as an overlay. It adds a heartbeat to each of the connected nodes that make up the enterprise knowledge graph.

And by doing so, it adds a whole new level of organizational context into agentic operations: If the enterprise knowledge graph answers, the what, digital twin adds the why and the when?

Digital Twin for Performance

Let’s start with the ‘why?’

Because while the enterprise knowledge graph provides the foundational data context for agent operations, at one level it’s a little too blandly descriptive: It can answer the what but not the why.  

How does an agent know whether it’s performed its task well or badly?  Well, how do we?

Most of us have objectives decomposed from organizational objectives. Ideally, those objectives are SMART and can be linked directly back to actual measures of performance.

With a digital twin, those performance measures can be aggregated or derived directly from those event streams: Counts of customers onboarded; average processing and wait times; error rates; IT and business resource costs. And we can track changes to those measures over time to monitor alignment to or drift from those objectives.

Whether your role is running agents or governing them, this is really a critical capability.  

For the agent with a repeating task, it needs to know more than just ‘what’ the outcome is – it needs to understand how well it has done it.

For people - or agents - charged with monitoring agentic processes, they need to be able to assess the health of those agents within the context of process performance.

Digital twin gives you a framework for assessing not just the ‘why’ (via objectives) but also the ‘how well’ of agentic business. But insight isn’t worth much unless it’s followed by action.

Digital Twin for Action

Digital twin also adds a second critical dimension to the enterprise knowledge graph: When.

Agents may be autonomous processes - but when should they act?

Digital twins integrate the knowledge graph’s static representation of the operating model with a continuous flow of events representing its changes of state. Each state change carries the potential for action: A condition may be met; a performance threshold exceeded. Either way, it’s the starting gun on our agents.

For enterprise architects those state changes may be simple events – an individual technology goes out of support; a project passes a governance gate – but for the most part our job is to take the long view, not govern individual transactions. We must look at how the enterprise machinery itself is performing.

For this we need to look at collections of events: Trends, correlations and thresholds. Supplier risk passes an acceptable threshold; AI token costs rise; core system availability is trending the wrong way.

Whatever the case, something is outside of normal operating parameters and we need to take action – not to intervene in individual cases but to plan maintenance work to the machinery of the business and IT operating model itself, whether it's maintenance (e.g. addressing technical debt to improve reliability), an upgrade (e.g. upgrading your CRM to improve productivity) or turning something off altogether (e.g. dDiscontinuing use of an AI tool to counteract rising token costs).

By linking together an understanding of business and IT operations, a live view of business performance, and clear triggers for action, the digital twin provides the foundation for a self-correcting agentic operating model.

Context Graph

Sadly, for all this talk of brilliant automation, there’s one thing we can’t as easily automate: Accountability.  

Even in the era of agentic, Murphy’s Law (‘anything that can go wrong will go wrong’) still holds true. When it does, pointing the finger and saying ‘the agent did it’ won’t get you very far.

We can frame this as an issue of AI alignment, policies and guardrails. But it’s a data problem too.

Through the enterprise knowledge graph and the digital twin, we’ve fed our agents with comprehensive view of business and IT operations. What’s missing is a comprehensive view of agent operations.

Context graphs fill this gap. Like the enterprise knowledge graph, they are graphs, made up of nodes and edges; unlike it, these don’t represent the macroscopic structure of business and IT operations, but the microscopic structure of agent reasoning.

As it turns out, this is something more than a safety net – it’s a huge AI accelerator.

Context Graph for Trust

Let’s start with trust.  

Automation is a double-edged sword. Being able to delegate increasingly complex and multi-variant tasks and decisions to a machine represents a huge productivity opportunity.  

But how do you trust that process? Machines carry neither an understanding of consequences not any accountability for them. For businesses operating in highly regulated environments, this is a significant problem.

The context graph serves as a decision trace for AI agents. It represents as nodes and relationships the variables, processes and choices that went into an individual decision as well as the decision itself. The result is something like a blockchain of agent reasoning which can be traversed forwards and backwards to reconstruct a chain of events for audit and compliance purposes.

Context Graphs for Learning

There is a secondary and even more profound implication of context graphs. It is a data structure that enables agents to learn.

Think about the list of parts we’ve assembled and how they interact:

• From the enterprise knowledge graph we have a true representation of connected business and IT operations.

• From the digital twin we have event-based triggers to spark our agents into action.

• From the context graph we have a record of that action, including the input parameters and the output decision

• And back to the digital twin, we have the means to track the outcome of that agentic decision.

And all this data exists in a persistent and query-able data layer outside of any agent.

What we have is not just an audit record of automated decisions but a living knowledge base of successful decision-making: Were those decisions good decisions or bad decisions? Did they realize the anticipated outcomes; or maybe have unintended consequences?  

By linking decisions to outcomes, we weigh those decisions for alignment to our own business’s mission and strategy. As organizational knowledge assets go, it's hard to think of a more consequential one.

Conclusion

It’s not an original point, but there’s a huge difference between technology innovation and technology integration.

Gartner’s famous hype cycle and Amara’s Law (‘we tend to overestimate the short-term impact of new technologies while underestimating their long-term effects’) equally make the point that when the shiny new tech rolls out of the lab, that’s when the work really begins.

The performance of LLMs will continue to improve at an accelerated rate. The idea that agents will augment - and in some cases supplant – the human workforce will continue to drive eye-watering Silicon Valley valuations in the near-term.

But organizations striving for AI-driven market advantage and fretting over ‘tokenomics’ and ROI face a very different challenge: How to integrate that AI into enterprise workflows in a way that aligns operations with each other and with organizational objectives.

That alignment can’t be implicit and sit within those agents; it has to be explicit and sit across them. Without it, our agents won’t resemble a free-wheeling flock of birds as much as a fleet of fairground dodgems, continually colliding and reorientating before colliding again.

That alignment framework is made of more than just data, but data is its foundation, and those key assets of the enterprise knowledge graph, the digital twin, and the agentic context graph make explicit what the business model is, how it acts, and why it changed.  

It’s a potent platform for organizational recursive self-improvement. And the time to start investing in it is right now.

Learn more about how OrbusInfinity enables enterprise AI adoption and governance.

‍Book a demo today.

‍

‍

‍

‍