AI-ready knowledge requires more than a knowledge graph

To deliver the outcomes you want it to deliver, AI needs more than data. It also needs meaning, context, relationships, business rules and trusted knowledge.

This is driving interest in ontologies, knowledge graphs and semantic layers. So far, so good — but the technology works best when it is grounded in a clear understanding of what your business actually needs to know. To build an ontology, you need to understand what knowledge matters most. And a knowledge graph is most valuable when you know which concepts, decisions, rules and exceptions it needs to represent.

In other words, you have to get the basics right first. And to discover, capture, prioritize and govern knowledge in a consistent, repeatable way, you need a knowledge operating model.

Give AI the right knowledge, not all the knowledge

Enterprise knowledge is scattered across databases, operational systems, applications, dashboards, documents, policies, procedures, videos, meeting notes, collaboration channels and — most importantly — your employees’ know-how. It might be accurate and valuable, but it might also be outdated or contradictory.

Getting AI to make sense of it involves a lot more than collecting documents, interviewing experts, connecting systems, then loading everything into a repository, however. AI needs to know which sources to trust, which rules apply and how concepts relate to one another.

A more useful approach, then, is to create a knowledge criticality model that assesses each source based on value, access cost, risk and volatility. A policy document that determines regulatory compliance, for example, is likely to be more critical than a generic process description. Likewise, a dashboard metric is useful only if its definition and lineage are clear.

Start by deciding what matters most:

• Which business domain or use case should you focus on?
• Which decisions, workflows or outcomes are you trying to improve?
• Which knowledge sources are most important to those outcomes? Are they reliable, accessible and properly governed?

The goal is to identify the knowledge that is most relevant, trustworthy and reusable for a specific business outcome. From there, learn quickly and design with scale in mind.

• ALSO READ: 8 ways to improve your enterprise knowledge strategy for AI

Connect workflows, data and decisions

Many AI initiatives assume that the underlying workflow is already understood. In reality, that’s often not the case. Processes may be only partially documented, and critical knowledge often exists only in employees’ heads. Different teams may even use different names for the same business concept.

Making knowledge reusable and governable means connecting three complementary views:

1. The workflow, data and system view: The activities, roles, systems, data assets, documents, reports, handoffs and dependencies that show how work actually gets done. It answers questions such as: What work is performed? Who performs it? Which systems are used? Which data is consumed and produced? Where are the handoffs?
2. The operational knowledge view captures rules, decisions, tacit know-how, exceptions, workarounds, quality checks and golden cases that explain how people apply judgment in practice. It answers questions such as: How do people know what to do? Which rules are applied? Where do exceptions occur? What judgment is required? What knowledge separates an experienced performer from a novice?

The governance view: Ownership, validation, traceability, versioning, permitted AI usage and controls that keep knowledge up to date and therefore and safe to use. It answers questions such as: Who owns this knowledge? Who validated it? What is the source of evidence? When should it be reviewed? Can AI use it directly, or does it require human-in-the-loop approval?

Bringing these three views together transforms knowledge capture from a one-off consulting exercise into a repeatable, scalable practice.

Choose standard objects to make knowledge reusable

If every team documents knowledge differently, it’s difficult to reuse it.

A common object model, as its name implies, gives you a common way to describe and connect the knowledge your organization needs to reuse. Structure workflow knowledge around reusable objects such as workflows, activities, roles, systems, data assets, decisions, rules, exceptions and controls. Then define how these objects relate to one another: A workflow contains activities; an activity is performed by a role; a decision depends on a rule; a system produces data; and a control governs a process.

This structure makes knowledge easier to discover from day one while creating a bridge between business and technology. Business teams can recognize workflows, rules and decisions, and data teams can connect those objects to data assets, lineage and quality rules. AI teams can use them to build context packages, retrieval patterns, agent instructions and evaluation criteria.

The key is not to overengineer the model. Create enough structure so knowledge captured in one use case can be reused in the next.

Use workflow knowledge cards for consistency and governance

One practical way to build a common object model is through workflow knowledge cards. Each card captures a single activity, decision, handoff or exception in a standard format.

A card might include the activity name, parent workflow, responsible role, inputs and outputs, systems used, data elements, documents, business rules, tacit knowledge, common exceptions, controls, owner, pain points, AI opportunities, validation status, source traceability and review date.

Using a standard structure creates consistency across workflows and makes knowledge easier to review, validate and govern. It also supports prioritization by allowing each card to be assessed according to business value, risk, AI relevance and governance readiness.

This approach also encourages reuse. For instance, rules, decisions, exceptions and data dependencies can be shared across workflows, while ownership, validation, traceability and review cycles become part of the knowledge asset itself.

Use the same language to turn documentation into data

Traditional documentation is difficult to compare because every team describes its work differently. A controlled vocabulary solves this problem by giving you a standard list of terms to use when classifying workflows, information and AI opportunities, such as:

• Activities can be classified as collect, validate, approve, reconcile or report.
• Knowledge can be classified as business rule, heuristic, workaround, policy or quality check.
• Data usage can be classified as input, reference, calculation or audit evidence.
• Exceptions can be classified as missing data, conflict, manual workaround or compliance issue.
• AI opportunities can be classified as retrieve, summarize, validate, explain or recommend.

A controlled vocabulary also lets you analyze knowledge across your organization. You can identify where manual decisions are concentrated, where exceptions occur repeatedly, where data quality problems exist, where controls are weak and where AI can safely add value.

This is how documentation becomes structured, measurable and reusable knowledge.

Do your knowledge due diligence

Not every knowledge source is ready for AI. Each source should first undergo knowledge due diligence to assess its quality, ownership, freshness, completeness, consistency, source-of-truth status and risk:

• Data sources should be assessed for ownership, stewardship, lineage, quality rules and semantic mapping.
• Documents should be assessed for structure, version control, approval status and alignment with current practice.
• Tacit knowledge should be validated by multiple experts and linked to real-world examples.
• Rules and decisions should be assessed for policy alignment, exception handling and auditability.

This exercise will help you decide whether or not a source can be used, and where there are issues that need to be resolved first. For example, some sources may be valuable but need better ownership, updated documentation, clearer definitions or stronger quality controls before they can support AI safely.

Embed knowledge governance by design

If you capture knowledge first and think about governance later, keeping it trustworthy and up to date becomes much harder. Governance should be built into every knowledge asset from the start.

Each asset should have a clearly accountable owner, a validation status, source traceability, review dates, version history and defined usage permissions. It should also specify whether AI can use the knowledge directly for retrieval, recommendations or other purposes, and when human review is required.

Governance should include ongoing controls such as freshness checks, completeness checks, contradiction detection, source-of-truth mapping, drift monitoring and approval workflows.

This is where knowledge governance complements data governance, but the two are not the same. Data governance focuses on data assets, ownership, quality, lineage and controls. Knowledge governance extends those principles to business meaning, rules, processes and AI context. You need both in an AI-enabled organization.

Start with one use case to build scalable knowledge architecture

A practical way to begin is by choosing one high-priority business domain and one use case. Use it to define your knowledge architecture and governance model.

Because the result includes reusable structures and components, your next use case won’t start from scratch. You can build on existing governed knowledge assets, reuse common business entities and rules, extend existing mappings, adapt context packages and add only what’s new.

That’s how knowledge initiatives scale — not by trying to capture everything at once, but by steadily building reusable knowledge assets.

The future is governed knowledge production

New technologies can help extract concepts, recommend ontologies, classify documents, assess quality, detect contradictions and automate stewardship. These are valuable capabilities, but they don’t replace the need for a disciplined operating model.

When you standardize how knowledge is discovered, validated, governed, reused and applied, you move beyond isolated AI experiments and toward AI systems that better understand your business and keep improving over time.

• ALSO READ: You can’t build an autonomous factory on fragmented data

WHAT TO DO NEXT

Read more about NTT DATA’s AI Services to learn how we can help you operationalize AI throughout your organization.