Reference

The AI and Revenue Dictionary

258 terms a director-and-above operator will hear this year, defined in plain language, without vendor spin. Every research figure, benchmark, and framework we publish uses this vocabulary. If a word in an issue is doing work you cannot check, it is defined here.

Search the dictionary

258 terms across 12 sections.

Part 1

AI systems and mechanics

The physical and mathematical machinery underneath everything else. Read this section first if the rest of the dictionary uses words you cannot ground.

Attention: The mechanism inside a transformer model that lets each token in the input look at every other token to decide what matters. Attention is what makes a language model able to keep track of a long instruction, connect a pronoun to its subject, or notice that a table column and a chart caption refer to the same thing. Multi-head attention runs several attention operations in parallel so the model can track different kinds of relationships at once.
Autoregressive generation: The way most language models produce output. The model generates one token, appends it to the input, and predicts the next token. Every word after the first is written with all the previous words as context. This is why the beginning of a response constrains everything that follows and why prompt structure matters as much as prompt content.
Base model: The raw pretrained network before any instruction tuning, safety tuning, or alignment. Base models produce fluent text but do not reliably follow instructions. Every model you use in production is a base model plus one or more layers of post-training. See instruction tuning, RLHF, and alignment.
Context window: The maximum number of tokens a model can process in a single request, including both the input and the output. A one-million-token context window sounds infinite until you try to fit six PDFs, a system prompt, tool schemas, and a running conversation into it. Context windows are a hard ceiling, and effective context is always smaller than the advertised maximum because attention degrades near the edges.
Diffusion model: A generation approach used mostly for images, video, and audio. The model learns to reverse a noise process: it starts with random noise and denoises step by step into a coherent output. Different math than a transformer language model, but the same core idea of learning a mapping from distribution to distribution.
Embedding: A vector representation of a piece of content. Embeddings turn text, images, or code into arrays of numbers where semantic similarity becomes geometric closeness. Embeddings are the substrate of every retrieval system, every semantic search, every vector database, and most recommendation engines.
Fine-tuning: Training a model further on a specific dataset to shift its behavior for a particular task, domain, or voice. Full fine-tuning updates every weight. LoRA and other parameter-efficient methods update a small subset. Fine-tuning is what you do when prompting cannot get you to reliable behavior and retrieval cannot supply the missing knowledge.
Foundation model: A large model trained on broad data that can be adapted to many tasks. The term originally described the class of models like GPT, Claude, and Gemini, but it has drifted to include large image, video, and multimodal systems. If a vendor calls their model a foundation model and it is only trained on their customer data, they are using the word loosely.
Grounding: Anchoring model output in verifiable source material, usually via retrieval. A grounded answer includes citations to the documents that support it. An ungrounded answer is a fluent guess. In enterprise settings, grounding is the difference between a system you can defend to compliance and one you cannot.
Hallucination: When a model generates content that is fluent, confident, and false. Hallucination is not a bug that can be fixed; it is a property of how autoregressive language models work. It can be reduced with retrieval, constrained decoding, verification loops, and evaluations, but never eliminated. If a vendor claims zero hallucinations, they are describing a system that is not a language model.
Inference: The act of running a trained model to produce output. Distinct from training. Inference cost is what you actually pay in production. Latency is inference speed. Throughput is inference volume per unit time.
Instruction tuning: A stage of post-training where a base model is trained on examples of instructions and desired responses. This is what makes a model follow directions rather than autocomplete text. Instruction tuning is typically followed by RLHF or a similar preference-learning stage.
Latent space: The internal representation a model uses to think in vectors before producing tokens. You cannot see the latent space directly, but everything the model does is a trajectory through it. The metaphor is useful when you are debugging why a model connects two ideas that should not be connected: something in the latent space is putting them near each other.
LLM (large language model): A neural network, almost always a transformer, trained on enormous amounts of text (and increasingly images, audio, and video) to predict the next token. Modern LLMs range from a few hundred million parameters to more than a trillion. In practice, most enterprise usage is calls to hosted frontier models via API.
LoRA (low-rank adaptation): A parameter-efficient fine-tuning method that inserts small trainable matrices into a frozen base model. LoRA lets you adapt a large model to a specific task without retraining the whole thing, which is both cheaper and safer. Most enterprise "custom model" offerings are LoRA-based fine-tunes.
Model: The trained artifact: weights plus architecture. A model without an inference server is a file. Most operational conversations about "the model" are really conversations about the model plus its serving stack plus the prompt harness around it.
Multimodal: A model that can process more than one input type. Modern frontier models handle text, images, and audio in the same call. Multimodal is what makes vision agents, screenshot-driven workflows, and voice interfaces possible.
Parameters: The individual weights inside a model. A 70B model has 70 billion parameters. Parameter count is a rough proxy for capability at a given training vintage, but not a reliable one across vintages. A well-trained smaller model can beat a poorly-trained larger one.
Post-training: Everything done to a base model after pretraining: instruction tuning, RLHF, constitutional AI, safety tuning, tool-use training, reasoning tuning. Post-training is where most modern capability gains come from, not from scaling parameters.
Pretraining: The initial training pass where a base model learns to predict tokens on trillions of tokens of text and other media. Pretraining is expensive, done at frontier labs, and largely a solved-but-expensive problem. The interesting engineering is downstream.
Quantization: Compressing a model by using lower-precision numbers for its weights. A model trained at 16-bit precision can be run at 8-bit, 4-bit, or lower with modest quality loss and dramatic memory savings. Quantization is what makes it possible to run large models on smaller hardware.
Reasoning model: A model that has been trained to spend more compute at inference time, typically by generating an internal chain of intermediate steps before producing a final answer. Reasoning models trade latency and cost for higher accuracy on complex tasks. o-series and Claude thinking modes are the canonical examples.
RLHF (reinforcement learning from human feedback): A post-training method where humans rank pairs of model outputs and the model is trained to produce outputs humans prefer. RLHF is what made instruction-tuned models feel usable. It is also where a lot of a model's personality, hedging, and refusal behavior comes from.
Speculative decoding: An inference-time trick where a small fast model drafts tokens and a large slow model verifies them. Speeds up inference without changing final quality. Invisible to end users but material to inference economics.
Temperature: A parameter that controls output randomness. Low temperature produces more predictable output. High temperature produces more variety. Zero temperature is deterministic in principle, but in practice hardware and floating-point quirks introduce small variations. Production systems typically run at low temperature for consistency.
Token: The unit a language model reads and writes. A token is roughly three-quarters of a word in English. Tokenization varies by model. Pricing, context windows, and rate limits are all denominated in tokens.
Transformer: The neural network architecture underneath almost every modern LLM. Introduced in 2017. Uses attention rather than recurrence, which is what made it scale.
Weights: The learned numerical values inside a model. Weights are the actual "knowledge" of a model in the technical sense. Model files are files of weights.

Part 2

Agents, harnesses, and orchestration

The layer that turns a model into a system that acts. This is where most enterprise AI value is being built and most enterprise AI value is being lost.

Agent: A system where a model chooses actions, executes them, observes results, and iterates toward a goal. The distinguishing feature is that the model is in a loop, not answering a single question. Every real agent has three things: a model, a set of tools it can call, and a controller that decides when to stop.
Agentic workflow: A workflow that includes at least one agent step. Distinguished from a pipeline, which is a fixed sequence of deterministic operations. Agentic workflows are more powerful and less predictable. Both properties matter.
Autonomy level: How much freedom an agent has to act without confirmation. On a spectrum from suggestions only (human approves each step), to per-action approval, to full autonomy (agent acts and reports). Higher autonomy is where most rollbacks come from. See the Reversal Ledger.
Chain-of-thought: A prompting or training pattern where the model generates intermediate reasoning steps before its final answer. Improves accuracy on tasks that require multi-step logic. Most modern reasoning models internalize this pattern during training.
Context engineering: The discipline of shaping what information the model sees, in what order, at what point in a task. Context engineering is the successor to prompt engineering. It includes retrieval, tool schemas, system prompts, few-shot examples, memory management, and the sequencing of all of them. In production AI systems, context engineering is where most quality lives.
Controller: The component that decides when an agent should call a tool, when to reflect, when to stop, and when to escalate. Controllers can be model-driven (the model decides), heuristic (a fixed policy decides), or hybrid. Controller design is the difference between an agent that finishes a task and one that spins forever.
Deterministic wrapper: A layer of ordinary code wrapped around a non-deterministic model call. Handles validation, retries, formatting, and guardrails. Most production AI systems are 90% deterministic wrapper and 10% model, by both code volume and reliability contribution.
Environment: The world an agent acts on: a codebase, a CRM, a browser, a filesystem, an operating system. Environments differ in how much they punish failure. A read-only environment is safer than one where an agent can send emails, delete records, or run trades.
Evaluations (evals): Automated tests that measure whether a model or agent behaves correctly on a defined task. Evals are the single most underrated component of production AI. Without evals, you cannot tell if your last prompt change made things better or worse. With evals, you can ship faster and safer than teams that do not have them.
Guardrail: A rule that constrains what an agent can do or produce. Can be a prompt-level instruction, a wrapper-level filter, a tool-level restriction, or a downstream reviewer. Real guardrails are enforced in code, not in prose.
Harness: The full runtime around a model: the prompt template, tool interfaces, retrieval, memory, controller, evals, and observability. Harnesses are what make one team's frontier-model agent work and another team's fail on the same model. The word comes from the machine learning research community and has spread as agent engineering matured.
Hierarchical agent: An agent that spawns subagents to handle parts of its task. Each subagent has its own model, tools, and context. Coordination happens through the parent agent. Hierarchical designs are how complex, long-horizon tasks get decomposed into manageable subtasks.
Human-in-the-loop: A design where a human approves, edits, or corrects agent actions before they take effect. HITL is the safest deployment pattern and the most expensive. Most production agents are HITL for high-stakes actions and autonomous for reversible ones.
Kill switch: A mechanism to stop an agent immediately, from outside the agent's own control. Every production agent needs one. Every production agent that has caused an incident either lacked one or had one nobody knew how to use.
Loop: The core cycle of an agent: observe, decide, act, observe. Also, colloquially, the pathological state where an agent repeats the same action indefinitely because it cannot recognize that it is stuck.
Loop engineering: The discipline of designing agent loops that make progress, know when to stop, and recover from stuck states. Includes turn budgeting, progress checks, reflection steps, and escape hatches. A well-engineered loop is the difference between an agent that ships and one that burns tokens.
Memory: Information an agent retains across turns or across sessions. Short-term memory is the current context window. Long-term memory is what gets written back to storage and retrieved later. Memory design determines whether an agent feels intelligent over time or amnesiac.
Multi-agent system: Two or more agents that interact to accomplish a task. Can be hierarchical (one agent directs others) or peer-based (agents negotiate). Multi-agent systems are more capable than single agents on some tasks and more expensive and less predictable on almost all of them.
Observability: The ability to see what an agent did, why, and with what inputs. Includes logging, tracing, replay, and metrics. If you cannot answer "what did the agent do at 2:47 pm on Tuesday and why," you do not have observability.
Orchestration: The layer that coordinates multiple agents, tools, and steps toward a business outcome. Overlaps with workflow engines from the pre-AI era, but adds the ability to route based on model output. Tools like LangGraph, Temporal, and Airflow are common substrates.
pass^k: A reliability measure from the tau2-bench line of research. Pass^k is the fraction of k independent runs of the same task that succeed. Pass^1 is what a demo shows. Pass^4 or higher is what a deployment looks like. Most agent quality gaps hide between pass^1 and pass^4. See The Revenue AI Report's research on agent reliability.
Planner: A component that decomposes a goal into steps before execution. Can be the same model that executes the steps or a different one. Planning quality determines whether the rest of the agent has a chance.
Reasoning trace: The visible chain of thought a reasoning model produces before its final answer. Can be exposed to users, used internally, or discarded. Traces are useful for debugging and dangerous when they leak proprietary information.
Reflection: A step where an agent reviews its own output before continuing. Can catch errors, improve quality, and detect stuck loops. Reflection is a real capability improvement and a real cost multiplier.
Retrieval: Fetching relevant documents, records, or context to include in a model's input. Almost always done via embedding similarity, sometimes combined with keyword search. Retrieval is the difference between a model that guesses and a model that cites.
Retry: Running an action again after a failure, often with modifications. Retry policies matter enormously in agent systems because the cost of a retry (in tokens and time) is real and the failure modes it papers over are real.
Skill: A reusable capability an agent can invoke: a prompt template plus a tool set plus expected inputs and outputs, wrapped in something the agent knows when and how to use. The term is used by Anthropic's Claude, some agent frameworks, and various vendor products. Not standardized, but converging.
Subagent: An agent invoked by another agent to handle a subtask. Subagents have their own context and can be delegated to in parallel. Used for isolation (their context does not pollute the parent's) and for parallelism.
System prompt: The instructions given to a model at the start of a conversation, usually invisible to the end user. Defines the model's role, constraints, and behavior. System prompts are where most of an application's personality and safety lives.
Tool: An external function an agent can call: a database query, an API request, a shell command, a browser action. Tools are how agents affect the world. Tool design is where most agent quality lives.
Tool-use: The capability of a model to select an appropriate tool and generate correct arguments for it. Modern frontier models are trained specifically for tool-use. Reliability varies widely by tool complexity.
Turn: One cycle of the agent loop. Also a unit of budgeting: an agent might have a maximum of 20 turns to complete a task before it must return or escalate.
Verifier: A component that checks agent output before it commits an irreversible action. Verifiers can be another model, a heuristic, a schema, or a human. The presence of a cheap automatic verifier before an irreversible action is what separates agents that work from agents that do not. See research/agent-reliability.

Part 3

Prompting and context

The interface between humans and models. This is where non-engineers do the most damage and the most good.

Chain-of-thought prompting: Instructing a model to show its work: "think step by step." Improves reasoning on complex tasks. Largely subsumed by reasoning models trained to do this internally, but still useful for non-reasoning models.
Constitutional AI: An alignment technique where a model is trained against a set of written principles rather than only against human preferences. Introduced by Anthropic. Produces models that can articulate why they refuse a request rather than just refusing.
Few-shot prompting: Providing several worked examples of the task in the prompt before asking the model to do it. Few-shot is often the fastest way to improve model behavior without fine-tuning.
Function calling: The API-level mechanism by which a model produces structured output that a program can execute as a tool call. Sometimes called tool use. All major model APIs support this.
JSON mode: An API setting that forces the model to produce valid JSON output. Reduces parsing errors. Not a substitute for evals; a model can produce valid JSON with wrong content.
Prompt engineering: The practice of designing prompts to produce desired behavior. Once a discipline unto itself, now a subset of context engineering. Still useful, but rarely sufficient for production systems.
Prompt injection: An attack where malicious input embedded in retrieved documents, user messages, or tool outputs overrides the system prompt. Prompt injection is unsolved and probably unsolvable in the general case. Real defenses require containment, not clever prompting.
RAG (retrieval-augmented generation): The pattern of retrieving relevant documents and including them in the prompt before generation. RAG is the workhorse of enterprise AI. Also the source of most enterprise AI failures, because retrieval quality is usually worse than teams assume.
Retrieval quality: How often the retrieval step finds the right documents. A model can only be as good as the context it sees. Retrieval quality is the single most underrated component of RAG systems.
Structured output: Model output constrained to a specific schema: JSON, XML, a Pydantic model, a SQL query. Structured output is what makes model outputs safe to pass to downstream code.
System prompt injection: When a user or upstream input tries to override or bypass the system prompt. A subset of prompt injection.
Zero-shot prompting: Asking a model to perform a task with no examples, only instructions. Modern frontier models are strong zero-shot on most tasks that were in their training data.

Part 4

Retrieval, memory, and knowledge

Where the information the model reasons over actually comes from.

Chunk: A slice of a document used as a unit of retrieval. Chunk size and overlap are quiet but consequential design choices. Too small and you lose context; too large and you dilute relevance.
Chunking strategy: How documents are divided into retrievable units. Options include fixed-size, sentence-boundary, semantic (grouping related content), and structural (respecting document sections). Chunking strategy has a bigger effect on retrieval quality than most teams realize.
Embedding model: The model that converts text (or images, or code) into vectors. Different from the generation model. Embedding models are typically smaller and cheaper. Choice of embedding model is a foundational RAG decision.
Graph engineering: Designing the entity and relationship graph that underlies a knowledge system. Distinct from vector-based retrieval. Graph engineering answers "how are these things connected" where embeddings answer "how are these things similar." Both matter, and modern systems increasingly use both together.
Graph RAG: A retrieval pattern that traverses a knowledge graph in addition to (or instead of) vector similarity. Better for queries that require multi-hop reasoning across entities. More expensive to build and maintain.
Hybrid search: Combining keyword search and vector search. Almost always outperforms either alone. Most production retrieval systems are hybrid.
Index: The data structure that supports fast retrieval. For vector search, typically a specialized structure like HNSW or IVF. For keyword search, an inverted index. For graph search, a graph database.
Knowledge graph: A structured representation of entities and the relationships between them. Wikipedia is a knowledge graph. Your CRM is a knowledge graph, mostly poorly maintained. Enterprise knowledge graphs are having a resurgence because agents work better on them than on unstructured documents.
Long-term memory: Information an agent stores and can retrieve across sessions. Distinct from context window (short-term). Long-term memory design is where most agent-continuity failures happen.
Reranking: A second retrieval pass where a smaller, faster model or heuristic reorders the candidate results from the first pass. Reranking is almost always worth it if you care about retrieval quality.
Semantic search: Retrieval based on meaning, not keywords. Enabled by embeddings.
Vector database: A database optimized for storing and querying embeddings. Examples: Pinecone, Weaviate, Qdrant, Chroma, and increasingly the vector features inside general-purpose databases like Postgres (via pgvector).
Vector search: Retrieval by finding vectors closest to a query vector in embedding space. The core of most modern retrieval systems.

Part 5

Evaluation, safety, and reliability

How to know whether the thing you built works, and how to stop it when it does not.

Alignment: Making a model behave according to intended values, safety constraints, and instructions. Alignment is both a technical field and an operational responsibility. In production, alignment is what you actually monitor for.
Benchmark: A standardized test used to compare models on a task. Public benchmarks (MMLU, HumanEval, tau2-bench) are useful for model selection but frequently gamed. Private benchmarks that reflect your specific workload are what you actually need.
Confidence tag: A visible marker on a published claim indicating how strong the underlying evidence is. Used by The Revenue AI Report to sort research by evidence strength rather than by headline. See glossary of publication conventions.
Contamination: When benchmark data leaks into training data, causing a model to look better on the benchmark than on real-world tasks that resemble it. A perennial problem with public benchmarks.
Distillation: Training a smaller model to imitate a larger one. Cheaper to serve, sometimes only slightly less capable. Most fast models are distilled from larger ones.
Drift: When a deployed model or agent's behavior changes over time without any code change, usually because of upstream model updates, retrieval index changes, or shifting inputs. Drift detection requires ongoing evaluation.
Eval harness: The infrastructure for running evaluations: dataset, model calls, scoring, and reporting. Every serious AI team has one. Most enterprise teams do not.
Golden set: A curated dataset of high-quality examples with known correct outputs, used for evaluation. Building a golden set is the highest-leverage thing an AI team can do in the first month of a project.
LLM-as-judge: Using one model to evaluate the output of another. Cheap, scalable, and biased. Useful for triage and rough scoring. Not sufficient for decisions that matter.
Red teaming: Systematically trying to break a model or agent to find vulnerabilities before adversaries do. Enterprise red teaming is still a young discipline.
Regression: When a model or agent gets worse at a task it previously handled correctly. Regression testing is what evals prevent. Without evals, regressions surface as customer complaints.
Rollback: Pulling an AI deployment out of production and reverting to the previous system, human or otherwise. Rollback is measured, common, and disclosed only when required. The Reversal Ledger tracks documented cases.
Rollback rate: The share of deployed AI agents that have been pulled back over a governance failure. Published survey data puts it at 74 percent of enterprises. See research/rollback.
Speakable content: Content marked in schema (via `SpeakableSpecification`) as suitable for voice interfaces and AI answer surfaces to read aloud. Affects which parts of a page LLMs preferentially quote.

Part 6

Model economics and infrastructure

What all of this actually costs and where the money goes.

Batch inference: Running many inference requests together to improve throughput and reduce cost. Not always available; not always applicable. When it is, it can cut costs meaningfully.
Compute: The GPU, TPU, or accelerator time consumed by training or inference. The dominant cost driver of every AI system. Compute pricing is opaque, tiered, and frequently negotiated.
Context caching: Reusing computed key-value caches across similar requests to avoid recomputing them. Reduces cost and latency for repeated context patterns. Most major model APIs now support it.
Cost per query: The average cost of a single end-user interaction with your AI system. Rarely a single model call. Almost always includes retrieval, tool calls, verification, and sometimes multiple model turns. Real cost per query is often 3-10x the naive model call estimate.
Deployment: The full stack that serves a model in production: model, inference server, load balancing, monitoring, and integration into your product. Deployment is where research quality becomes user experience.
Fine-tuning cost: The cost of training a fine-tuned model. Distinct from inference cost. Amortized across all inference calls that use the fine-tune. Rarely worth it for narrow use cases; often worth it for wide ones.
GPU: Graphics processing unit. The hardware most AI training and inference runs on. Nvidia dominates. AMD and custom silicon (Google TPU, AWS Trainium, various startups) are gaining share.
Hosted inference: Running inference on a cloud provider's infrastructure via API. What most enterprises do. Cheaper than self-hosting for most volumes; more expensive at very high volumes.
Latency: The time from request to response. Matters more than teams initially assume. A 300ms difference in AI agent response time can meaningfully change user behavior.
Rate limit: The maximum number of requests, tokens, or compute units a customer can consume per time window. Rate limits are how model providers manage capacity. Enterprise contracts negotiate higher limits.
Serving: Running a model in production to handle real requests. Distinct from training. Serving optimization (batching, caching, quantization, speculative decoding) is where much of the applied AI infrastructure work happens.
Total cost of ownership (TCO): The full cost of running an AI system: inference, retrieval, storage, engineering, monitoring, governance, and remediation. Vendor pricing shows a fraction of TCO. Real TCO surprises finance teams every quarter.

Part 7

AI in revenue: applied categories

The specific product categories AI is showing up in inside revenue teams.

Agentic CRM: A CRM where AI agents update records, execute follow-ups, and orchestrate workflows across the pipeline. Distinguished from CRMs that add AI features to a fundamentally human-driven system. Agentforce, Claudeforce, and several startups are competing here. See research/rollback for outcomes.
AI-generated content: Content produced primarily by a model, with varying degrees of human editing. Includes emails, ads, blog posts, product descriptions, and outbound sequences. Volume is not the problem; discernment is.
AI SDR: Software that automates outbound prospecting: list building, personalization, email generation, sending, and follow-up. The category has produced both the largest ARR growth and the largest documented reversals in AI-era revenue. See research/named-reversals.
AI voice agent: An agent that handles phone calls: qualifying, booking, answering questions, taking payments. Improving fast on structured tasks. Still uneven on judgment calls.
Autonomous prospecting: Fully agent-driven outbound: research, list building, sequencing, and sending with no human in the loop. Distinct from AI-assisted prospecting, where a human reviews before sending. Autonomous prospecting is where most AI SDR reversals have happened.
Content marketing automation: AI systems that produce, distribute, and optimize marketing content at scale. Has flooded the internet with low-quality content and materially reduced the return on generic content marketing.
Conversation intelligence: Software that records, transcribes, and analyzes sales calls. Gong and Chorus were the first generation. Modern systems add coaching, forecast signals, and deal risk scoring.
Copilot: An AI feature embedded in an existing product that assists a human user, rather than replacing them. Microsoft Copilot, Salesforce Einstein Copilot, and GitHub Copilot are canonical examples. Copilots are safer to deploy than fully autonomous agents and produce smaller productivity gains.
Deal intelligence: AI systems that score deals, predict close probability, and surface risks. Data quality is the ceiling on all of these systems.
Enrichment: Adding attributes to a record: firmographics, technographics, intent signals, contact information. AI has made enrichment cheaper and less reliable at the same time.
Forecasting: Predicting revenue outcomes: pipeline, bookings, retention, churn. AI has changed how forecasts are computed but not why they are wrong.
Guided selling: AI that recommends next steps to reps in the flow of their work. Success depends on whether the recommendations are trusted. Trusted recommendations happen when the underlying signals are visible.
Lead scoring: Ranking leads by predicted fit or intent. Most modern lead scoring is model-driven. Most modern lead scoring is not evaluated on downstream conversion.
Meeting automation: AI systems that schedule, prep for, run notes on, and follow up after meetings. Fastest-growing category by user count. Slowest-growing category by measurable revenue impact.
Personalization: Tailoring outreach or content to a specific recipient. AI has made personalization cheap. Cheap personalization is not the same as effective personalization; specificity that references the wrong thing worse than a generic message.
Pipeline generation: The process of creating qualified opportunities. Historically SDR-driven; increasingly AI-assisted or AI-driven. Pipeline generation quality is where most AI SDR reversals reveal themselves.
Revenue intelligence: A category that combines conversation intelligence, deal intelligence, and forecasting. Gong, Clari, and Salesforce Revenue Intelligence are the incumbents. Every AI-native competitor is threatening to disrupt them.
Sales enablement automation: AI that produces training content, coaching feedback, and skill assessments. Effective when tied to specific competencies. Less effective as a general "AI coach."
Signal capture: Detecting and acting on buyer behavior: pricing page visits, competitor comparisons, job changes, funding events. Signal capture has been the fastest AI-adoption area in RevOps because the ROI is legible.
Zone 1 work: Work that could not previously be economically staffed with humans: 24/7 monitoring, real-time enrichment at every touchpoint, simultaneous parallel outreach to thousands of accounts. Zone 1 is the largest AI opportunity because it creates new work rather than automating existing work.

Part 8

Revenue and GTM: the base vocabulary

The economic and operational terms every revenue conversation eventually returns to. Defined here so the AI-adjacent terms above have anchor points.

AOV (average order value): The average revenue per transaction. A retail and ecommerce metric that occasionally applies in B2B for transactional products.
ARR (annual recurring revenue): The annualized value of subscription revenue at a point in time. The dominant metric in B2B SaaS. Does not include one-time fees, services revenue, or usage overage.
Attribution: Assigning credit for a revenue outcome to the marketing, sales, or partner activities that contributed. Attribution is unsolved in principle and approximated in practice. Multi-touch, first-touch, last-touch, and data-driven models each capture different things and none capture everything.
Blended CAC: Customer acquisition cost calculated across all channels, paid and organic. Distinguished from paid CAC, which only counts paid acquisition spend. Blended CAC is closer to economic truth.
Book of business: The set of accounts owned by a specific rep or team.
Bookings: The dollar value of contracts signed in a period. Distinct from revenue (which recognizes over time) and billings (which invoice on a schedule).
Bottom-up: A GTM motion where end users adopt a product first and drive expansion through the organization. Slack, Zoom, and Notion are canonical examples. Distinguished from top-down enterprise sales.
Buying committee: The group of people involved in a B2B purchase decision. In enterprise deals, typically 6-10 people spanning economic buyer, technical evaluator, end user, procurement, security, and legal.
CAC (customer acquisition cost): The total sales and marketing cost divided by new customers acquired in a period. The denominator of most GTM efficiency math. Fully-loaded CAC includes people costs; ad-hoc CAC often does not.
CAC payback: The number of months required for gross profit from a new customer to equal the cost of acquiring them. Sub-12-month payback is aspirational. 18-24 months is normal in enterprise SaaS. Over 30 months is a problem.
Churn: The rate at which customers stop paying. Logo churn is customer count. Revenue churn is dollars. Gross churn is total lost. Net churn subtracts expansion.
CLV (customer lifetime value): See LTV.
Cost of goods sold (COGS): The direct cost of delivering a product. In SaaS, includes hosting, third-party APIs, customer support, and payment processing. Distinct from sales and marketing spend.
CRM (customer relationship management): The system of record for customer and prospect data. Salesforce dominates enterprise. HubSpot dominates mid-market. Every AI system in revenue is either connected to a CRM or trying to replace one.
Cross-sell: Selling additional products to an existing customer.
CS (customer success): The team responsible for post-sale retention, expansion, and outcomes. Distinguished from support (reactive) and account management (commercial).
CSAT (customer satisfaction score): A survey-based measure of customer satisfaction with a specific interaction or product.
Deal size: The dollar value of a single deal. ACV (annual contract value) is the annualized version. TCV (total contract value) includes multi-year commitments.
Demand generation: The marketing function responsible for creating pipeline. Historically distinct from brand marketing. Modern demand gen is increasingly indistinguishable from AI-assisted outbound.
Discount: Reducing price below list. Discounting is a strategic choice that becomes a habit that becomes a problem.
Downsell: A customer moving to a lower-priced plan or fewer seats. Counts against expansion revenue.
Enterprise: Deals typically above $100K ACV, sold to organizations of thousands of employees, with committee-based buying and multi-quarter cycles.
Expansion revenue: Revenue growth from existing customers: upsells, cross-sells, seat additions, usage growth.
Fully-loaded cost: The total cost of an employee including salary, benefits, taxes, tools, and overhead. Typically 1.3-1.5x salary. Used in real CAC and productivity calculations.
GTM (go-to-market): The combined function of sales, marketing, and customer success. Also, the strategy for how a company reaches and serves its market.
Gross margin: Revenue minus cost of goods sold, divided by revenue. In SaaS, gross margins above 75% are healthy. Below 60% suggests either a commodity or a business model problem.
Gross retention rate (GRR): The share of revenue retained from existing customers, excluding expansion. GRR of 90%+ is enterprise-healthy.
ICP (ideal customer profile): The type of customer most likely to succeed with your product and stay. Defined by firmographics, technographics, and behavior. Every serious GTM function has one. Most have one that is out of date.
Impressions: A count of how many times an ad or piece of content was displayed. A leading indicator that is easy to inflate and easy to misread.
Inbound: Leads generated by prospect-initiated action: form fills, demo requests, self-serve signups. Distinguished from outbound (rep-initiated).
Inside sales: Sales conducted primarily over phone, email, and video, without in-person meetings.
Land and expand: A GTM motion where the initial sale is small and the customer grows revenue over time through expansion.
Lead: A person who has expressed some interest, usually by giving up an email address. Leads become MQLs become SQLs become opportunities become customers, or not.
LTV (lifetime value): The total gross profit a customer generates over their lifetime with the company. Notoriously easy to overestimate. Most published LTV numbers are calculated wrong.
LTV:CAC: The ratio of customer lifetime value to customer acquisition cost. 3:1 is a commonly cited benchmark. 3:1 is also frequently gamed by manipulating both sides.
Magic number: Net new ARR in a quarter divided by sales and marketing spend in the prior quarter. A shorthand measure of GTM efficiency. Above 1.0 suggests efficient growth; below 0.5 suggests inefficient growth.
Mid-market: Deals typically between $25K and $100K ACV, sold to companies of 100-2000 employees.
MQL (marketing qualified lead): A lead that meets marketing's criteria for being handed to sales. Definition varies enormously by company. MQL volume is a leading indicator; MQL-to-opportunity conversion is a lagging indicator that matters more.
Net dollar retention (NDR): Revenue retained from a cohort of customers, including expansion, divided by their revenue one year prior. NDR above 120% is best-in-class. NDR below 100% means the business is leaking.
Net promoter score (NPS): A survey score measuring customer likelihood to recommend. Widely used, widely criticized. Directionally useful; specifically misleading.
New logo: A customer that did not exist in the prior period. New logo revenue is distinct from expansion revenue.
Opportunity: A qualified deal in the pipeline. Distinct from a lead (earlier) and a customer (later).
Outbound: Rep-initiated prospecting: calls, emails, LinkedIn messages, and combinations of the above.
Payback period: See CAC payback.
PLG (product-led growth): A GTM motion where product usage is the primary acquisition and expansion driver. Distinguished from sales-led growth. Not mutually exclusive; most successful modern SaaS is hybrid.
Pipeline: The dollar value of open opportunities weighted by stage. Also, the process of generating those opportunities.
Pipeline coverage: The ratio of pipeline to quota. 3x coverage is a common target. Coverage is a lagging indicator that reveals problems late.
Product-market fit: The state where a product is pulled by its market rather than pushed to it. Notoriously hard to measure. When it exists, most GTM problems get easier. When it does not, most GTM investments get wasted.
Quota: The revenue or bookings target assigned to a rep or team over a period.
Quota attainment: The share of reps hitting quota in a period. Below 50% attainment is a signal that either targets or coverage or motion is broken.
Ramp: The time from when a rep starts until they are producing at full capacity. Typical ramp is 3-9 months in mid-market, 6-12 months in enterprise. Ramp time is one of the most under-measured levers in revenue.
Retention: The share of customers or revenue that stays. See gross retention and net retention.
Revenue operations (RevOps): The function that owns the systems, data, processes, and analytics powering sales, marketing, and customer success. RevOps has expanded from a back-office role to a strategic function over the last decade.
Rule of 40: The sum of annual growth rate plus profit margin. Above 40 is healthy for growth-stage SaaS. Below 20 is a problem.
Sales-led growth: A GTM motion where sales-driven relationships are the primary acquisition mechanism. Distinguished from PLG.
Sales cycle: The elapsed time from first contact to closed deal. Enterprise cycles range from 3-18 months. Cycles get longer during recessions and shorter during peaks.
SDR (sales development representative): An outbound-focused role that generates pipeline for closers. Under active reinvention by AI. See AI SDR and research/named-reversals.
Seat: A single user license. The dominant pricing unit in B2B SaaS. Under pressure from usage-based and outcome-based pricing.
Segment: A group of accounts with similar characteristics: enterprise vs mid-market, industry, region, product tier.
Self-serve: A product experience where users can sign up, pay, and expand without talking to a human. Distinct from PLG (a broader concept).
SMB (small business): Deals typically below $25K ACV, sold to companies of fewer than 100 employees. High volume, high churn, low sales-cost tolerance.
SPIFF: A short-term incentive paid to reps for hitting a specific action or product-attach target. Latin origin unclear; usage universal.
SQL (sales qualified lead): A lead that sales has accepted as worth pursuing. Higher bar than MQL.
Territory: A geographic, industry, or account-based subdivision assigned to a specific rep.
TAM, SAM, SOM: Total addressable market, serviceable addressable market, serviceable obtainable market. A cascade of shrinking numbers used to size opportunities. Frequently inflated at the top and ignored at the bottom.
Top-down: An enterprise sales motion driven by executive alignment before broad user adoption. Contrasted with bottom-up.
Upsell: Selling a higher-priced tier or additional product to an existing customer.
Usage-based pricing: Pricing tied to product consumption rather than seat count. Snowflake, Twilio, and AWS are canonical examples. Under experimentation across SaaS.
Velocity: Deals or dollars closed per unit of time. Sales velocity = (# opps × ACV × win rate) / cycle length.
Vertical: An industry-specific segment. Vertical SaaS targets one industry deeply; horizontal SaaS targets one function broadly.
Win rate: The share of qualified opportunities that close as won.

Part 9

Pricing, packaging, and unit economics

The specific vocabulary of how software companies capture value.

ACV (annual contract value): The annualized dollar value of a contract. Distinct from TCV.
Add-on: An optional product or module sold alongside the core.
Committed spend: A minimum dollar amount a customer commits to over a contract period, regardless of consumption. Common in usage-based pricing.
Contribution margin: Revenue minus variable costs of delivering that revenue. A finer measure than gross margin for understanding unit economics.
Discount curve: The pricing structure that reduces price at higher volumes or longer commitments.
Feature gating: Restricting features to specific pricing tiers. The mechanic that makes tiered pricing work.
Floor: The minimum revenue commitment on a usage-based contract.
Freemium: A pricing model with a free tier that converts a share of users to paid. Common in PLG. Requires large top-of-funnel volume to work.
List price: The published price. Distinct from realized price, which reflects discounts.
Minimum contract value (MCV): The smallest deal a company will sign. Setting MCV too low destroys unit economics; setting it too high shrinks the market.
Outcome-based pricing: Pricing tied to a business outcome the customer cares about: revenue delivered, meetings booked, tickets deflected. Growing but rare because outcomes are hard to measure and attribute.
Overage: Charges applied when a customer exceeds their contracted usage. Overage revenue is high-margin and unpredictable.
Platform fee: A recurring fee for access to the platform regardless of usage. Common in usage-based pricing to smooth revenue.
Price realization: The ratio of realized price to list price. A measure of how much a company actually captures versus what it publishes.
Ramp deal: A contract that starts at a lower price and increases over time. Used to bridge budget constraints or usage growth curves.
Renewal: The event where a customer extends their contract. Renewal rates and renewal ACV are core revenue metrics.
Seat licensing: Pricing per user with a fixed price per seat.
Term: The length of a contract commitment. Longer terms usually earn discounts.
Tier: A pricing level with a defined feature set. Most SaaS uses 3-5 tiers.
Tiered pricing: Pricing with multiple named packages at different prices and feature sets.
Volume discount: Price reduction for larger quantities.

Part 10

Metrics and measurement

The vocabulary of what gets counted.

Cohort: A group of customers acquired in the same period, tracked together. Cohort analysis is how retention and lifetime value actually get measured.
Funnel: The progression from top-of-funnel awareness through purchase and retention. Every stage has its own conversion rate and its own leak.
Funnel conversion rate: The share that moves from one stage to the next. Multiplied across stages, gives end-to-end conversion.
Growth rate: Revenue change over a period. YoY (year-over-year) and QoQ (quarter-over-quarter) are the most common.
KPI (key performance indicator): A metric selected as a headline measure of performance. Often over-indexed. Frequently gamed once it becomes the KPI.
Leading indicator: A metric that predicts future outcomes. Pipeline generation is a leading indicator of future bookings.
Lagging indicator: A metric that measures past outcomes. Bookings are a lagging indicator of prior GTM investment.
North star metric: A single metric that best represents the value delivered to customers. Contested concept; sometimes useful, sometimes distorting.
Time to value: How quickly a new customer realizes the promised benefit. Shorter time to value correlates strongly with retention.
Vanity metric: A metric that looks good in a deck but does not correlate with business outcomes. Impressions, followers, and hours saved are common examples. See The Proof Gap.

Part 11

Data and operations

The plumbing under everything.

Data enrichment: Adding attributes to records from external sources. See enrichment.
Data hygiene: The ongoing work of keeping data accurate, complete, and current. Systematically underfunded. See Pipeline Truth Test.
Data model: The structure of how records and their relationships are represented in a system. Data model quality determines what questions you can ask and what agents you can build.
Field: A single attribute on a record: name, title, company, stage, close date. Fields are where data quality lives or dies.
Firmographics: Company-level attributes: industry, size, revenue, geography. Used for segmentation and scoring.
Instrumentation: The code that captures behavioral data. Under-instrumented systems produce unreliable analytics.
Object: A record type in a CRM or database: account, contact, opportunity, lead, activity. Object model design is a foundational GTM decision.
Pipeline hygiene: The discipline of keeping opportunity data accurate: stages current, close dates realistic, amounts correct. Bad pipeline hygiene invalidates forecasts.
Record: A single row in a database: one account, one contact, one opportunity.
Signal: A behavior or event that indicates buyer intent: pricing page visit, competitor comparison, job change, hiring surge. See signal capture.
System of record: The authoritative source of truth for a given data type. Sales pipeline lives in CRM. Financial data lives in the ERP. Every AI agent needs to know which system is the system of record for what it touches.
Technographics: Technology stack attributes: what tools a company uses. A common enrichment layer for B2B targeting.

Part 12

Publication conventions (used by The Revenue AI Report)

Terms defined by this publication and referenced throughout its issues, research, and blog. Included here for readers who arrived via glossary rather than via context.

Eight Seats: The functions every issue is cut for: sales, marketing, RevOps and GTM engineering, enablement, customer success, partnerships and BD, exec and founders, revenue finance. One case, eight reads, a decision for each.
Hybrid 1.9x: Bridge Group data showing hybrid human-plus-AI outbound teams producing 1.9x qualified meetings per dollar versus AI-only, and 2.4x versus human-only. An efficiency result, not a volume result.
Kill criteria: The fail conditions written into an AI contract before signature: the metric, the floor, the date it is measured, and the consequence for a miss. Without them a failed pilot becomes a two-quarter argument.
OAR Matrix: Maps AI work across Optimize (make existing work faster or cheaper), Amplify (make existing work more effective), and Reinvent (do work that could not previously be done). A sequencing framework for where to invest.
Optimization Theater: A year of pilots, dashboards, and reported time savings presented upward as transformation. Activity is measured, adoption is celebrated, no revenue outcome changes. The visible behavior that produces the Proof Gap.
Pipeline Truth Test: A five-part check of whether CRM data can support an AI agent: field completeness, stage honesty, contact freshness, activity capture, and outcome labeling. Agents inherit the pipeline they are pointed at.
Proof Gap: Money spent on AI with nothing attributable behind it. Tools were bought, pilots ran, time savings were reported upward, and revenue still cannot be tied to any of it.
Reinvestment Gap: The distance between hours AI gives back and revenue those hours produce. Time is saved, nothing is redeployed to a named higher-value activity, and the saving evaporates before it reaches the number.
Reversal Ledger: A running count of AI decisions a human had to undo: agent outputs corrected, stages rolled back, sends pulled. Tracked over time, the reversal rate shows whether agents are earning trust or borrowing it. See /reversal-ledger.
Shadow AI Stack: The set of AI tools reps already use outside the sanctioned roadmap: personal accounts, pasted call notes, buyer data in consumer tools. Running your go-to-market whether or not it is governed.
Single-Player AI Problem: A real AI win that stays with one operator. The workflow was never written down, owned, or wired into the system of record, so the gain never becomes a team result.
The Bet: Shorthand for the position readers are asked to make: move first on AI and be able to defend it. See /about.

How this dictionary is maintained

Terms are added when they earn a place in operator conversations, not when they appear in a vendor deck. Definitions are revised when the underlying concept shifts. Every entry is written to survive a director-level reader who has heard the word used loosely and wants a version they can act on. Corrections and additions: /corrections.

Cite as: Kvarfordt, Jonathan. "The AI and Revenue Dictionary." The Revenue AI Report. https://www.therevenueaireport.com/dictionary