We are entering a new phase of AI development. This is not just about generating content or answering questions. It’s about building systems that can act on our behalf. It’s an exciting idea, agents that do things for us. But here’s the uncomfortable truth: most organizations aren’t ready for it. They don’t understand what “agentic AI” actually means, and they’re not building the foundations needed to make it work.
The Problem with Incomplete Context
Let’s start with the basics. If you’re relying on a large language model (LLM) out of the box, you’re not necessarily getting a competitive advantage. You’re getting efficiency, yes. But so is everyone else. That’s because the LLM is built on a general model of the world. It doesn’t know your world. It doesn’t know your products, customers, or intellectual property. And it’s certainly not drawing from your proprietary data unless you give it access, securely, intentionally, and with structure.
So how do you fix that?
This is where Retrieval-Augmented Generation (RAG) comes in. The LLM isn’t the oracle. It’s the processor. You don’t ask it to know everything. You give it trusted sources, curated content, and rich, contextual signals. Then you ask it to work with that. That’s the right architecture. That’s how you start getting real results.
The Difference Between Efficiency and Differentiation
If you’re doing what everyone else is doing, you’re not going to win. Organizations compete on their knowledge. Knowledge about the market, about customer needs, about technical possibilities, about what’s been tried and what hasn’t. Standardization gives you scale and efficiency. But differentiation comes from your unique knowledge, your proprietary data, and your internal language.
Agentic AI needs to be grounded in that. It needs a reference architecture, structured, curated, and aligned to your business goals.
Why 40% of Agentic AI Projects Will Fail
I’ve shared this stat before. By 2027, 40% of agentic AI projects will be canceled[1]. Why? Because the term “agent” is being stretched beyond usefulness. People are building agents to call calculators and tell you what 2 plus 2 is (that was an example from a webinar I recently attended!). That’s not intelligence. That’s a demo (and a trivial one at that).
Agentic AI is complex. It requires orchestration. Multiple systems. Multiple models. Clean handoffs. Preserved context. It needs engineering. And more importantly, it needs understanding.
You cannot automate what you do not understand.
So start with the process. What does a human do in that workflow? What are the steps? What decisions need to be made? Once you understand the process, you can ask how an agent might augment or automate parts of it. But jumping straight to automation without that process understanding is a great way to waste money and time.
What Context Really Means
When I ask an LLM to write in my voice, I give it my voice. I give it writing samples, style, tone, and structure. Context isn’t just what the question is. It’s who’s asking, what the audience needs, and what the intended outcome is.
If your agent doesn’t know who it’s talking to, or what problem it’s solving, it’s not going to give you a good answer. That context needs to be embedded (standards such as MCP – Model Context Protocol – seek to capture and communicate context). That’s true whether you’re building a shopping bot or supporting a field technician with an 800-page manual.
And by the way, no single LLM can do all of that. We use multiple models. Some for document intelligence. Some for table extraction. Some for classification and tagging. This is what agentic orchestration really looks like. Multiple tools working together in a coordinated way, each with its own role, permissions, and guardrails.
You Can’t Automate a Mess
Many organizations are using AI to make up for years of poor content hygiene. Unstructured data is where knowledge is captured, but it is not usable if it lacks context. You need to curate your content, build lifecycle models, and apply metadata in the form of “is-ness” (the nature of the content) and “about-ness” (the descriptors that tell one instance from another).
What is this document? What type is it? What’s it about? How do we tell similar documents apart?
This is reference architecture. It’s not sexy, but it’s essential. And yes, LLMs can help. But only if you give them the right instructions. We’ve built an LLM-powered virtual information architect that applies 30 years of methodology to guide that process. But it all starts with structure.
There’s Always a Human in the Loop
With agentic automation, governance and human oversight aren’t optional. Agents need to operate within clear boundaries, with metrics, thresholds, and alerts that trigger human intervention when something goes sideways.
You don’t give a junior employee full access to every system or significant authority without oversight. You wouldn’t give that to an agent either.
Start small. Scope your use cases. Define your baselines and your desired outcomes. What are you trying to change? What does success look like? Have an answer to that before you deploy anything.
What Excites (and Frightens) Me
The most powerful use cases today? Product data remediation. Hyper-personalization. Recommendation engines. Knowledge curation and access. These are areas where agentic AI shines because they require the orchestration of multiple models, APIs, and systems to drive intelligent decisions.
And the future? Think personal agents that know your preferences, values, and your decision-making style. Agents that can talk to other agents and advocate on your behalf. Not just search, but anticipate. Not just summarize, but prioritize.
That’s personalization at a level we haven’t seen before. That’s where we’re headed.
LLMs can describe context in tens of thousands of dimensions that defy description in human-understandable terms. The mathematics of the model are not decipherable by mere mortals. A model can understand latent attributes that you don’t even understand about yourself. Imagine the power of personalization at that level. Imagine the peril of personalization at that level. Models that understand human motivations in ways that are not understandable by humans. That is exciting. And scary.
Notes
[1]https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
Seth, a great and very insightful article. You highlight something organisations urgently need to hear: agentic AI is not a shortcut, it’s the next level of disciplined design, structure, and contextual intelligence. I especially agree with your point that differentiation comes from proprietary knowledge and alignment, not from generic LLM capabilities.
From a CX and service-delivery perspective, the idea of orchestrated models with clean handovers, context layers, and human oversight is exactly where things break — or start to create real value. Your article brings clarity to a complex space, and it was a pleasure reading it. Thank you –R
Thanks for the validation. It seems that organizations are beginning to realize this. We have an upcoming webinar on the topic of addressing these issues in order to scale GenAI. Governance and content operations along with a reference architecture are pieces to the puzzle. Here is an upcoming webinar on the topic: https://www.earley.com/why-genai-pilots-fail-and-how-to-scale-successfully-webinar