Enterprises are rushing to adopt so-called “agentic AI,” systems that promise to not just answer questions, but also take actions, automate tasks, and drive decisions. In theory, these agents can draft emails, update records, generate product specs, or flag anomalies without human involvement. In practice? Most deployments don’t make it past the demo.
We’ve seen this movie before. Pilots launch. Nothing goes live. The chatbot doesn’t work. And yet we keep calling it AI, when it’s actually giving us the illusion of intelligence.
We’ve talked to dozens of organizations across industries, and the pattern is depressingly familiar:
- A well-meaning team spins up a proof of concept using a large language model (LLM).
- They connect it to a content source, maybe add a prompt template.
- It seems to work ... until it doesn’t.
- Results are inconsistent. Business users don’t trust it. Adoption stalls.
The root cause isn’t model accuracy or even hallucination. It’s lack of context.
These systems may be capable of reasoning. But reasoning requires structured knowledge, and most enterprises simply haven’t done the work to provide it. Today’s language models are brilliant at generating language but not meaning.
They don’t know your business rules. They don’t understand your product catalog. They don’t know which version of the policy is current or which one is safe to use. That’s not intelligence. That’s autocomplete with attitude.
One knowledge management (KM) lead summed it up this way: “We didn’t build an agent. We built a guess engine.”
I’m not saying agentic AI is a dead end. Far from it. But until enterprises ground these systems in structured, governed, well-modeled knowledge, they’ll continue to overpromise and underdeliver.
DEFINING AGENTIC AI WITHOUT THE HYPE
The term “agentic AI” is already suffering from misuse. Like “digital transformation” and “machine learning” before it, it’s on a fast path to becoming a buzzword no one can define and everyone claims to offer.
Let’s be clear. Agentic AI refers to systems that can take autonomous action based on context and intent. That means they don’t just respond. They perform tasks, trigger workflows, retrieve information, make decisions, and learn from outcomes. They are meant to behave more like digital employees than tools. But there’s a catch. Agency requires judgment and judgment requires context.
Most of what’s being marketed as agentic AI today is really just a wrapper around a language model:
- A fancy chatbot that fills out a form
- A plugin that summarizes tickets
- A tool that emails a recommendation based on a prompt
These are helpful features, but they aren’t agents. They don’t understand your enterprise’s logic. They aren’t goal-directed. They don’t know when to stop, escalate, or check for exceptions.
As one engineer put it during a failed deployment: “Agents need to know when they are out of the correct context—they need to know the boundaries and only operate within them—just like the boundaries of what an employee can do. If that is not understood, it will not meet a higher threshold of capability. It will stay stupid.”
True agentic systems must operate within constraints. They need this type of knowledge:
- Business rules and policies
- Workflows and handoff points
- Task sequences and dependencies
- What constitutes a “good” or “safe” outcome
That is not something a foundation model brings out of the box. It happens through information architecture, contextual modeling, and structured content.
This is where many projects go sideways. Teams assume the model will figure it out. But the model is only as smart as the scaffolding around it. Autonomy is not magic. It’s architecture.
And until that architecture exists, as with consistent metadata, structured vocabularies, and well-formed content, the agent remains little more than a random content generator.
Why Agentic AI Fails Without Information Architecture
Most failed AI projects don’t fail because the model was bad. They fail because the organization handed the model a mess:
- A mess of content
- A mess of systems
- A mess of language and logic and context
This is the invisible root cause of what people call hallucination or drift. The model is not inventing answers because it needs to interpolate between outdated and inconsistent content and what the user is asking for. It is doing its best with fragmented, poorly tagged, and often conflicting information.
The problem is not the intelligence. The problem is the inputs. This is why information architecture (IA) matters. Taxonomies, metadata, content models, and governance processes are not optional. They are the foundation. Without them, agentic AI cannot function safely or reliably in an enterprise context. We have seen this up close.
- Agents were asked to summarize policies but could not tell which version was current.
- Chatbots pulled product descriptions from 5-year-old PDFs with outdated specs.
- A smart assistant offered contradictory advice because two similar documents were tagged inconsistently.
In each case, the problem looked like AI. But the real issue was a lack of structured data.
When teams invest in metadata alignment and content modeling, the quality and consistency of AI outputs improves dramatically, not just slightly. Dramatically.
There is no AI without IA. Structured knowledge is not just a best practice. It is a prerequisite. Without taxonomies, metadata, and governance in place, agents hallucinate, users lose trust, and compliance risks skyrocket.
IA is how you teach the system what things are, how they relate, and why it matters.
Without that scaffolding:
- Agents do not understand user intent.
- Results are incomplete or incorrect.
- Compliance and risk controls break down.
- Adoption falters.
This is not a side issue. It is the core of whether your agent adds value or adds liability. Many CIOs still approach generative AI as a technical exercise. What they need is a knowledge strategy.
The Static Content Trap
One of the most common reasons agentic AI fails is not that the AI is too advanced, it is that the content is not. We regularly encounter enterprise content environments that look modern on the surface but are functionally frozen in time.
- Pages that have not been updated in years
- SharePoint libraries full of duplicate or contradictory files
- Unstructured PDFs with no metadata
- Product information split across disconnected systems
- Taxonomies that are inconsistent, poorly structured, or not aligned with other systems
These are not technical problems. They are KM failures. Organizations assume that content is “there” because it lives somewhere in the system. But what matters is not whether it exists, but if it is structured, maintained, and accessible in context.
A large enterprise recently deployed an AI assistant trained on internal documentation. It failed immediately. The issue? Half the content was outdated. The other half had no consistent tagging or categorization. The assistant simply could not tell what was accurate or relevant. There was also difficulty identifying the correct level of granularity. An answer sufficient for me may not be to you. The background and metadata of the individual are important signals in agent response. Does it need to return technically crisp answers or more of a step by step? My background and characteristics provide clues in the form of metadata. Those are additional signals to tell the model LLM what to return and to personalize the response.
Static content is a trap. The velocity of new content has to be processed in addition to the backlog of legacy. That information requires cleanup and curation with elimination of the redundant outdated and trivial content. There is no readiness without content structure.
The agent needs fewer, more-focused documents. It needs well-modeled knowledge:
- Defined content types
- Maintained metadata
- Role-based access and governance
- Archival policies for outdated material
- Clear ownership for each knowledge domain
When these foundations are missing, prompt engineering will not make up for your content management sins.
Many organizations already have KM teams in place (who are hopefully reading this piece and printing it out to leave on their boss’ desks). What they lack is executive recognition that KM is essential to AI success. Without investment in curation, tagging, taxonomy, and governance, the system fails before it begins. The expertise is there in many instances. Mature organizations have at least some degree of knowledge function (training and development, documentation of all sorts, engineering archives, etc.). The question is whether they are fully funding and leveraging that capability.
Agentic AI cannot operate on autopilot in a junkyard. It needs a runway, and that runway is structured, living content. You may have excellent, costly software that is up-todate and capable. You have a Ferrari—or at least a nice BMW or Mercedes. But the data and content are like rutted roads. You cannot open up a performance car on dirt backroads in poor condition.
A Framework for Agentic Readiness
Enterprise leaders want agents that work out of the box. What they forget is that humans don’t work that way either.
You don’t hire someone and expect them to navigate your systems, understand your policies, and make decisions on Day 1. You train them. You give them guardrails. You limit their scope until they build trust. Agentic AI needs the same foundation.
Agentic readiness has four dimensions.
- Structured Knowledge
Agents need to know what content exists, what it means, and how it is organized. This includes:
- Taxonomies and controlled vocabularies
- Content models that define attributes, types, and relationships
- Metadata standards across systems
Without this structure, the agent can’t reason or retrieve.
- Contextual Integration
Agentic systems must operate within a business context, not a vacuum. This means:
- Tying content to workflows, processes, and task triggers
- Understanding user roles and permissions
- Mapping intent to steps to actions
This is where IA meets business process modeling. It is the connective tissue between knowledge and behavior.
- Interaction Design
Most AI projects fail not in the back end, but in the handoff to the user. Successful agents are designed around:
- Clear user expectations
- Intuitive escalation paths
- Transparent confidence indicators
Agents must not only act, but they must also know when to pause, ask for help, or hand off to a human. An agent must also enable real-time feedback and comments.
- Governance and Guardrails
As with human employees, agents require oversight:
- Who decides what an agent is allowed to do?
- At what point does it require human approval?
- How is feedback collected and used to improve the model?
Why, you may ask, do we need guardrails for agents if we don’t have them for people? The answer is that organizations absolutely do have guardrails for people. Not every person in your company can approve a budget or make a legal decision. Agents need the same role-based constraints. This framework does not eliminate complexity. But it provides a structure for evaluating risk, designing responsibly, and aligning your AI efforts with your business goals.
Just as no serious organization would deploy a new system without a security or compliance review, no agent should be deployed without passing a readiness check against these four dimensions.
Lessons From the Field: What Works
After 3 decades, we’ve seen what separates the AI winners from the noise. It is not the technology or the talent behind deployment It is application of a structured process to tell the agent what content is relevant to a person based on what they ask, when they ask it and who they are.
The difference is in the groundwork that leads to the architecture of personalization. Organizations that succeed with agentic AI do not start by building agents. They start by building the knowledge foundation those agents need to operate.
We have seen this play out in companies large and small and across industries. The successful ones share a few key traits:
- They treat taxonomy as infrastructure.
One retail organization reduced support costs by 30%. Not by building a better chatbot, but by restructuring its product taxonomy. The AI agent was the front end. The taxonomy was what made it smart. Teams that succeed invest in taxonomy early, maintain it consistently, and tie it directly to business outcomes. It is not a side project. It is the platform.
- They design for human logic, then translate to machine logic.
A B2B manufacturer approached agent design the same way it onboards new employees. It mapped out what a person would need to complete a task—what the inputs, decisions, validations, and handoffs were. It then translated that flow into an agentic model. By mirroring human cognitive paths, it created systems that were not only more accurate, but easier to govern and explain.
- They supervise agents like junior team members.
The most pragmatic teams view LLMs as interns. They do not expect them to know everything. They do not let them operate unsupervised. Instead, they use “confidence gates,” checkpoints where the agent must flag uncertainty and escalate to a human reviewer. This model does not slow things down. It builds trust and accelerates adoption.
- They build content pipelines before building agents.
One global distributor realized, mid-project, that 80% of its content was outdated, misclassified, or duplicative. It hit pause, built a content enrichment pipeline, and restructured its knowledgebase. Only after that cleanup did it deploy the agent and saw double the accuracy with half the post-processing.
None of these are flashy. But they work. They reflect a maturity that goes beyond experimentation and into operationalization.
Agentic AI is not about what the system can generate, it is about what it can understand. And understanding begins with structure, context, and accountability.
Neither is agentic AI about replacing people. It is about extending knowledge. And knowledge must be curated, structured, and stewarded.
Read the full article by Seth Earley on Enterprise AI World.