The terms your buyers are using right now — agents, context windows, model weights, MCP — are not marketing language. They're technical vocabulary for specific things. Buyers who use them with precision have either deployed these systems or are deep in evaluation, and they've read enough documentation to use the terms correctly. When a term lands as a black box in a discovery call, the conversation doesn't stop. It continues, and you're navigating it blind.
That's the gap this section closes. The terms showing up in public sector AI conversations, what they actually refer to, and where your existing mental models help versus where they'll get you into trouble.
Why These Terms Are in the Room Now
Federal civilian agencies have been running AI pilots since at least 2024, and a meaningful number of those pilots have moved into production or active procurement. The vocabulary shift in buyer conversations tracks that transition. When a technology is in experimentation, buyers describe what they want to accomplish. When it's in procurement, they describe what they're buying — and they use the vendor's and the spec's language to do it.
The terms in this section aren't aspirational. They describe systems that exist, that your buyers are evaluating or have already deployed, and that create real identity and access questions that nobody has fully answered yet. That last part matters: some of what you'll encounter in these conversations is genuinely unsettled. Knowing which parts are settled and which aren't is itself a useful thing to bring to a call.
What "AI" Actually Refers To in These Conversations
When buyers say "AI" in an enterprise architecture conversation, they almost always mean one of three things, and they're usually not interchangeable.
A language model is a system trained on large amounts of text to predict and generate language. The model itself is a static artifact: a very large set of numerical parameters, fixed after training, that determines how the system responds to input. It doesn't learn during a conversation. It doesn't remember previous conversations by default. It processes what you give it and produces output.
A deployment of that model is the infrastructure that makes the model accessible, adds memory or retrieval capabilities, connects it to tools and data sources, and wraps it in whatever guardrails the organization has configured. This is where most of the enterprise architecture decisions live, and it's where identity questions start to get interesting.
An agent is a system that uses a language model to reason about a goal and take actions toward it, typically by calling external tools, querying data sources, or invoking APIs. Agents are where the identity surface area expands significantly, because an agent acting on behalf of a user is doing things in systems — with credentials, with access, with audit trails — and the governance questions around that are not yet resolved in any clean way.
Most buyer conversations that touch identity are really about the third category, even when they start with the first.
What This Section Covers
The pieces in this section build a working vocabulary for these conversations, in order of dependency. You need to understand what a language model is before MCP makes sense. You need to understand what an agent is before non-human identity governance makes sense. Fuzziness on the early concepts compounds forward into confusion on the later ones.
The section covers: how language models process input and why that shapes what they can and can't do; what agents are and how they differ from the automated workflows your buyers already have; what MCP is and why it's becoming the standard for how agents connect to external systems; and what the identity surface area looks like when agents are operating in production environments.
The goal is specific. No term in a customer conversation should land as a black box.
Note: In identity, a token is a credential — a signed, time-limited artifact that asserts authorization and carries claims about the subject. In AI, a token is the basic unit of text that a language model processes: roughly a word or word-fragment. A "token limit" describes how much text a model can handle in a single interaction, not a security boundary. The word is identical. The concept is completely different. If a buyer mentions "token limits" in the context of an AI architecture conversation, they are talking about model capacity, not credential scope. Conflating them in a response will be noticed.
Note: In identity, an agent typically refers to a software component that enforces policy on behalf of an identity system — something with defined, auditable behavior that operates within explicit rules. In AI, an agent is a system that uses a language model to reason about a goal and decide what actions to take, often dynamically and in ways that weren't explicitly programmed. The overlap is real: both involve non-human actors operating with delegated authority in systems that contain sensitive resources. The divergence is significant: an AI agent's behavior is not fully deterministic. You cannot audit what it will do the way you can audit a policy enforcement component. That gap is where most of the hard identity questions in this space currently live.
The Honest Starting Point
Some of what your buyers are asking about has clean answers. Some of it is genuinely unsettled — the standards are new, the implementations are inconsistent, and the governance frameworks are still forming. The goal of this section is to give you enough vocabulary to tell the difference in a live conversation: to know when you can answer with confidence, when you should ask a clarifying question, and when the honest answer is that the industry is still working it out.
Done well, that's the move that makes a buyer trust everything else you say.
The next piece starts with language models — what they actually are, mechanically, and why that shapes everything that comes after.