Reference · Updated August 2026
The AI glossary for people signing the checks
Technical glossaries define what the terms mean. This one adds what they mean for your budget — because every term below will eventually appear in a proposal, and each one carries an assumption worth catching before you sign.
How to use this page: each entry has a plain-English definition and a buying implication — the question to ask or the trap to avoid when the term shows up in a pitch. Twenty-eight terms, four groups, ordered from the words in every deck to the ones that decide whether the system still works in month six.
Core concepts
The words every pitch deck opens with — and what they commit the vendor to.
Large language model (LLM)
A model trained on vast text data that generates language — the engine behind ChatGPT, Claude, and most of what gets sold as "AI" in 2026.
Buying implicationAsk which model, at what cost per unit of work, and what happens to your pricing when the provider changes theirs.
Foundation model
A large pre-trained model (from OpenAI, Anthropic, Google, Meta) that applications build on rather than training from scratch.
Buying implicationAlmost no business should train its own model. A vendor proposing to should face one question: why isn’t building on a foundation model enough?
Generative AI
AI that produces content — text, images, code — rather than only classifying or predicting from existing data.
Buying implicationGenerative and predictive problems price differently. Forecasting demand is classic machine learning; drafting responses is generative. Vendors blur this; budgets shouldn’t.
Inference
Running a trained model to get an output. Every user query costs inference compute — this is the meter that runs forever.
Buying implicationBuild cost is one-time; inference cost scales with usage. Any proposal without a monthly inference estimate at your real volume is missing half the price.
Token / context window
Tokens are the chunks models read and write (roughly ¾ of a word); the context window is how much a model can consider at once.
Buying implicationAPI pricing is per token, so verbose systems cost real money at scale. "We’ll just put everything in the context window" is a cost decision disguised as an architecture.
Hallucination
A model producing fluent, confident, wrong output. Not a bug to be patched — an inherent property to be engineered around.
Buying implicationAny vendor who says "our system doesn’t hallucinate" just failed vetting. The right answer describes containment: retrieval grounding, evaluation, human review where stakes are high.
Multimodal
Models that handle more than text — images, audio, documents, video — in one system.
Buying implicationUseful when your data is genuinely mixed (invoices, photos, call recordings). As a selling point on a text-only problem, it’s decoration.
Building blocks
The architecture words in every proposal — where most of the price variance hides.
RAG (retrieval-augmented generation)
Connecting an LLM to your own documents and data so answers come grounded in your content rather than the model’s memory.
Buying implicationThe default first architecture for "AI that knows our business" — typically $25k–75k built well. If a vendor proposes fine-tuning before trying RAG, ask why.
Embeddings / vector database
Embeddings turn text into numbers that capture meaning; a vector database searches them by similarity. Together they power retrieval.
Buying implicationPlumbing, not magic. Treat a proposal that itemizes "vector database implementation" as a major line item the way you’d treat a plumber itemizing "pipe".
Fine-tuning
Additional training that adjusts a foundation model’s behavior on your examples. Changes how a model writes more reliably than what it knows.
Buying implicationOccasionally right, frequently oversold. It adds training cost, evaluation burden, and maintenance forever. Legitimate sequence: prompting first, then RAG, then fine-tuning with a measured reason.
Agent / agentic AI
AI that takes multi-step actions — querying systems, calling tools, executing workflows — rather than answering a single prompt. 2026’s most-used sales word.
Buying implicationAutonomy multiplies both value and blast radius. Ask what the agent is permitted to do without a human, and what happened in testing when it was wrong.
Tool use / function calling
The mechanism letting a model trigger real actions — search a database, draft an email, update a record. The load-bearing part of any "agent".
Buying implicationThe demo question that separates builders from decks: "show me the list of tools and what each is allowed to touch."
Structured outputs
Forcing model responses into exact formats (JSON, defined fields) so software — not just people — can consume them reliably.
Buying implicationThe unglamorous feature that makes AI integrate with your CRM or ERP. Its absence in a proposal that promises integration is a red flag.
Guardrails
Controls that constrain what an AI system will say or do: input filtering, output checking, permission boundaries, escalation to humans.
Buying implicationAsk to see them demonstrated — what happens when a user asks for a refund policy that doesn’t exist? "The model is well-behaved" is not a guardrail.
Prompt engineering
Crafting the instructions that steer model behavior. Now a baseline skill inside AI engineering, not a standalone discipline.
Buying implicationIf prompt work is a major billed line item in 2026, the price does not reflect the market.
Operations and quality
The terms that predict whether the system still works in month six.
Evals (evaluation)
Systematic testing of AI output quality against defined criteria — before launch and continuously after. The discipline that separates production systems from demos.
Buying implicationThe single best vetting question you have: "show me your evaluation harness." Confident specifics mean engineering; hand-waving means you are the evaluation.
Drift
Performance quietly degrading over time as data, usage, or underlying models change. Every AI system drifts; the question is who notices first.
Buying implicationAsk what gets monitored after launch and who is on the hook when quality drops. The answer defines whether you’re buying a system or a snapshot.
MLOps
The operational machinery around models: versioned data, automated retraining, monitoring, rollback. The "boring" work that keeps ML honest.
Buying implicationFor classic machine learning projects, this is most of the real cost — and the part cut first by vendors competing on price.
Human-in-the-loop
Designing a person into the workflow — reviewing, approving, or handling escalations — rather than full automation.
Buying implicationUsually the right launch design, and it changes the ROI math: you’re buying speed with oversight, not headcount replacement. Distrust proposals that skip straight to full autonomy.
Latency
How long the system takes to respond. Model choice, architecture, and cost all trade against it.
Buying implicationSet the requirement in the contract in seconds, per use case. "Fast" is not a number.
Observability
Being able to see what the AI system actually did — inputs, outputs, costs, failures — in production.
Buying implicationWhen something goes wrong (it will), this determines whether the fix takes an afternoon or a forensic engagement. Ask what you’ll be able to see.
Buying and governance
The engagement-structure vocabulary — where deals are won and budgets are lost.
PoC vs. pilot
A proof of concept shows something can work, in weeks, on sample data. A pilot runs it with real users and real data against success metrics.
Buying implicationThe classic failure is a PoC celebrated as if it were a pilot. Fund PoCs small ($15k–50k), and define pilot success metrics before writing the larger check.
Data readiness
Whether your data is accessible, clean, and permissioned enough for the AI use case — the most common reason projects stall.
Buying implicationA vendor who quotes a price without asking about your data is quoting fiction. Expect a readiness assessment first; budget for remediation if the honest answer is "not yet".
Build vs. buy
Custom development against off-the-shelf AI products. In 2026 the honest answer is usually a mix: buy the commodity, build the differentiator.
Buying implicationAsk any builder which parts of your problem an existing product already solves. The quality of that answer is a vetting signal in itself.
Vendor lock-in
How hard it is to leave — proprietary platforms, undocumented systems, or code you don’t own.
Buying implicationContract for source code ownership, documentation, and handoff support. The time to negotiate exit is before entry.
Shadow AI
Employees using AI tools without approval or oversight — pasting customer data into free chatbots being the canonical case.
Buying implicationAlready happening at your company. The fix is a clear acceptable use policy plus sanctioned tools, not a ban that drives usage underground.
AI governance
The policies and controls for how your organization uses AI: what’s permitted, who approves, how outputs are checked, what’s off-limits.
Buying implicationCheap before an incident, expensive after. A one-page policy and a training hour beat both blanket bans and blind adoption.
Responsible AI / AI safety
Practices for deploying AI that behaves within intended bounds: bias testing, misuse prevention, transparency about AI involvement.
Buying implicationIn vendor pitches this is often a slide, not a practice. Ask for the concrete artifact behind the phrase — a bias test result, a red-team report, an audit trail.
Fluent enough to be dangerous?
The vocabulary gets you through the pitch. Vetting the people behind it is the part we do: brief us on your project and we’ll match you with talent whose production record has already answered the hard questions above.
Frequently asked questions
What is the difference between AI, machine learning, and generative AI?+
AI is the umbrella term for software that performs tasks associated with human intelligence. Machine learning is the dominant approach — models that learn patterns from data to predict or classify. Generative AI is the newer branch that produces content: text, images, code. Practically: forecasting next quarter’s demand is machine learning; drafting the email about it is generative AI.
What is the most misused term in AI vendor pitches?+
"Agent" currently leads. It correctly means AI that takes multi-step autonomous actions, but in 2026 sales usage it is applied to everything from a chatbot to a scheduled script. The test: ask what the system does without a human approving each step. If the answer is "nothing", it is not an agent — which may be exactly what you want, just priced accordingly.
Do I need to understand these terms to buy AI services?+
You need roughly this page’s worth — not to build anything, but because the terms are where proposals hide their assumptions. Knowing that fine-tuning usually follows RAG rather than replacing it, or that inference costs recur while builds don’t, changes what you ask for and what you pay. It is the same reason you learn a little construction vocabulary before a renovation.
What questions should these terms lead me to ask a vendor?+
Four cover most of it: Which parts of this are commodity versus custom? What does inference cost at our real volume? Show me the evaluation harness. What do we own and see when the engagement ends? Vendors comfortable with all four are usually worth shortlisting — it is the same bar we vet against in our network.