The AI your customers are currently buying is, at its mechanical core, a prediction system. It takes an input and generates an output by predicting what a plausible response looks like, based on patterns learned from an enormous amount of text. The chat interfaces, the autonomous agents, the "reasoning" capabilities — all of it is built on top of that prediction loop.
The marketing layer on top of that mechanism is thick. Vendors describe their AI products using words that sound precise but aren't: intelligent, reasoning, understanding, knowing. These are borrowed from cognitive science and applied loosely to systems that don't work the way cognition works. The gap between the marketing description and the mechanical reality is where buyer confusion lives, and it's where your conversations get complicated.
The vocabulary problem runs deeper than unfamiliar terms.
You've spent a decade or more in rooms where precision matters, where "federation" means a specific trust relationship, where "assertion" means a specific data structure, where "delegation" means a specific grant of authority. You know what it costs to use a word loosely in a technical conversation with a buyer who knows the spec. That precision is an asset in 95% of your accounts.
In AI conversations, that same precision creates a specific trap. The words on the table — model, agent, context, token, session — exist in your vocabulary already. They mean things. The problem is they don't mean the same things here, and the false familiarity is harder to catch than pure ignorance would be. A seller who has never heard "inference" knows they need to ask. A seller who has heard "inference" in a data context thinks they already know.
Federal RFPs are already using terms like "agentic workflow," "model context," and "retrieval-augmented" in ways that assume shared vocabulary. When a buyer's technical lead uses one of these terms and you respond with a definition that's 80% right, you've done something worse than admitting uncertainty: you've demonstrated that you're working from a different map.
Why the Vocabulary Is Borrowed Twice
The current wave of AI — large language models, generative systems, the technology underneath the products your customers are evaluating — emerged from academic research in machine learning and natural language processing. The researchers borrowed vocabulary from linguistics, statistics, and cognitive science. The vendors then borrowed from the researchers, sometimes accurately and sometimes not. By the time these terms reach a procurement conversation, they've been through three rounds of translation, and the original technical meaning is still present but buried under layers of repurposing.
"Model" is a good example. In data architecture, a model is a schema: a structured representation of how data is organized. In AI, a model is a trained artifact, a system that has processed vast amounts of data and encoded statistical patterns in a form that can generate outputs. These are not the same thing, and the differences matter when a buyer asks about model governance, model versioning, or model risk. Same word, different object.
"Agent" is another. The word has a specific meaning in IDAM infrastructure: a software component that runs on a system, communicates with a central authority, and performs defined operations. In AI, an agent is a system that uses a model to take actions in pursuit of a goal, often by calling external tools or services. The surface similarity is real. The architectural implications diverge quickly, and the identity implications are in a different category entirely.
You'll find this pattern across the AI vocabulary: terms that were precise in their original context, repurposed into a new context where they're approximately right but specifically wrong in ways that matter. The buyer's CTO who studied machine learning in graduate school and the vendor's sales engineer who learned the vocabulary from a product brief are using the same words to mean different things. You need to know which version is on the table before you respond.
Note: In identity, a model typically refers to a data schema or object structure — something you can inspect, version, and validate against a spec. The closest AI equivalent is a trained model artifact: a system that has encoded patterns from training data into billions of numerical parameters. It diverges here: you cannot inspect what an LLM "knows" the way you can read a schema. Model governance in AI is an open problem, not a solved one.
Note: In identity, an agent is a deterministic software component with a defined scope of operation. It does what it's configured to do, and its actions are auditable. In AI, an agent is a system that uses a model to decide what actions to take, which tools to call, and in what sequence. It diverges here: the behavior of an AI agent is not fully predictable from its configuration. This is not a limitation that will be engineered away — it's a property of the underlying prediction mechanism.
What This Section Does
What follows is a working vocabulary. Precise, not comprehensive. Each piece takes a term that's already appearing in your buyer conversations, explains what it actually refers to mechanically, and marks the places where your existing mental models will help and the places where they'll steer you wrong.
The goal is narrow: when a buyer's CTO uses a term, you know what they mean, you know what you mean, and you know whether those are the same thing. Sellers who try to become AI generalists before their next call end up with a lot of surface knowledge and no depth. Sellers who close the vocabulary gap precisely — who know exactly which five terms are live in their accounts right now and what those terms actually mean — have what they need.
Vocabulary, not expertise. That's the gap this section closes.