Every term below is one you are likely to meet in a vendor deck, a proposal, or a contract. Definitions are written for the person signing, not the person building.
Software that performs tasks normally requiring human judgment, such as classifying documents, summarizing text, or forecasting demand. It is a category, not a product.
An approach where a system learns patterns from historical data instead of following rules a developer wrote by hand.
The trained artifact that turns an input into an output. Two vendors can wrap the same model and market it very differently.
The process of building a model from data. Training is expensive and infrequent; using the finished model is the part that runs every day.
Running a trained model to get an answer. Your ongoing AI cost is mostly inference cost.
Models that produce new content, such as text, images, code, or audio, rather than only scoring or classifying existing content.
A model trained on large volumes of text that predicts likely continuations. It is very good at language shape and cannot be assumed to be correct about facts.
The instruction and context you give a language model. Prompt quality often changes output quality more than switching models does.
The unit models read and bill in, roughly three quarters of a word in English. Pricing and length limits are expressed in tokens.
The maximum amount of text a model can consider at once. Long documents must be split or retrieved in pieces to fit.
Fluent output that is factually wrong. It is a property of how these models work, so any workflow relying on accuracy needs verification built in.
A setting that controls output variability. Lower values give more repeatable answers, which is usually what business workflows want.
Fetching your own documents and passing them to a model so answers are grounded in your content instead of general training data.
A numeric representation of text that lets a system find passages by meaning rather than exact keywords.
Storage designed to search embeddings quickly. It is the index behind most document question-and-answer systems.
Splitting documents into passages before indexing. Poor chunking is a common reason a retrieval system returns almost-right answers.
Further training a model on your examples to shift its style or format. It changes behavior, not knowledge, and rarely fixes accuracy problems on its own.
Change in real-world data over time that quietly degrades a model that was accurate at launch.
A system that plans multiple steps and calls tools to complete a task. Every added step multiplies the ways it can fail, so scope matters.
Letting a model trigger a defined function, such as looking up an order or creating a ticket, instead of only writing text.
Rule-based automation that clicks through existing interfaces. Deterministic and brittle, and often confused with AI in proposals.
A required review step before AI output takes effect. The standard control for workflows with financial, legal, or safety consequences.
The layer that decides which model, prompt, tool, and fallback a request uses, and where switching vendors becomes configuration rather than a rebuild.
The documented rules for approved uses, data handling, review requirements, and accountability when output is wrong.
Vendor documentation of a model's intended uses, limits, and evaluation results. Ask for one before an approval decision.
Where your data is stored and processed. Frequently the deciding constraint in regulated industries, before any accuracy discussion.
A fixed set of representative cases with known correct answers, used to measure whether a change actually helped.
Licenses plus integration, review labor, monitoring, retraining, and change management. Pilot pricing rarely reflects it.
Employees using unapproved AI tools with company data. Usually a sign that the sanctioned path is slower than the unsanctioned one.
Send us the deck. We will tell you which claims are standard, which are unusual, and which need evidence before you commit.