Glossary

The vocabulary, without the sales pitch

40 terms from AI agents and go-to-market, defined as the field uses them. Where we disagree with the common usage, we say so rather than quietly redefining the word to suit ourselves.

40 of 40 terms

Account-based marketing (ABM): Treating a named list of accounts as the unit of marketing rather than a segment or a persona, and personalising to each. It trades reach for depth, so it only pays where deal sizes justify the research.; ABM campaign guide
Agent: A language model given a specific job, a set of instructions about how to do it well, and usually some tools it can call. The difference from a chatbot is scope: an agent is built to produce one kind of output to a standard, rather than to answer anything.; Used loosely enough in marketing that it now covers everything from a saved prompt to an autonomous system. Ask what it can do without a human in the loop; the answer tells you which meaning is in play.; The Fleet
Agent fleet: A set of specialist agents that share context and can hand work between each other, rather than one general assistant asked to do everything. The argument for a fleet is that domain instructions conflict: what makes a good cold email makes a bad case study.; The Fleet
Autopilot: A weekly proposal: the focus for the week, the missions worth running with a reason for each, and what to skip. It proposes and never acts; approval happens in the Dock.; Autopilot
Brand kit: The visual identity captured from a company's live site: ranked colour palette with roles, logo in light and dark, typefaces and social profiles. Used so generated documents look like they came from the company rather than from a template.
Buying committee: Everyone whose agreement a B2B purchase needs. Typically an economic buyer, a champion, one or more technical or functional evaluators, and at least one person whose only role is to be able to say no.; ABM campaign guide
Buying signal: An observable event that changes whether now is the right moment to reach out: a funding round, a relevant hire, a competitor's price change, a leadership move. The test of a good signal is that it gives you something specific to say.
Champion: The person inside the account who wants the purchase to happen and will argue for it when you are not in the room. A champion has something to gain personally from it succeeding; a supporter merely does not object.
Context window: The total amount of text a model can consider at once, counted in tokens. Everything the agent knows for a given request has to fit inside it: the instructions, the shared business context, the documents, and the conversation so far.; A bigger window is not the same as better recall. Material in the middle of a very long input is measurably less influential than material at either end.
Credits: The unit of metered work. One credit is anchored to roughly a penny of real provider cost, so a plan's allowance is a cost ceiling rather than a made-up currency. Reviewing and reading never consume any.; Pricing
Dark funnel: The parts of a buying process you cannot see: private communities, group chats, podcasts, peer conversations. Most of a modern B2B decision happens there, which is why last-touch attribution reliably credits the wrong thing.
Deliverable: A finished document an agent produced, saved as a versioned artefact rather than left in a chat thread. Deliverables can be reviewed, revised section by section, exported, and filed once approved.
Enrichment: Filling gaps in a record from an external source: adding a job title, a company size, a verified email. Enrichment decays, so a record enriched a year ago should be treated as a hypothesis rather than a fact.
Firmographics: The company-level attributes used to describe a market: industry, headcount, revenue, geography, structure. The company equivalent of demographics, and on its own a weak predictor without situational context.
Grounding: Tying what an agent says to source material it was actually given, and requiring it to say when it has none. The practical form is a rule in the instructions: name the document behind a material claim, and never attribute to something you were not shown.
Guardrail: A constraint that holds regardless of what the model decides: a claim it may not make, a system it may not write to, a spend it cannot exceed. Guardrails are mechanical. An instruction politely asking a model not to do something is not one.
Hallucination: A model asserting something specific and false with the same confidence it asserts true things. Most commonly a citation, a statistic or a quotation that does not exist.; Usually a briefing failure rather than a model failure. Instructing an agent to be authoritative while giving it nothing to be authoritative from is a direct instruction to invent.; The Briefing Playbook
Human in the loop: A design where a person approves an action before it takes effect, rather than reviewing it afterwards. The distinction matters: reviewing after the fact is auditing, not approving, and it does not prevent anything.; The Dock
Ideal customer profile (ICP): A description of the accounts worth pursuing, specific enough to exclude accounts you would otherwise chase. Firmographics plus the situational conditions that make now the right time.; An ICP that excludes nobody is a market description, not an ICP. The exclusions are the useful half.; Build an ICP with AI
Intent data: Evidence that an account is researching a problem you solve: content consumption, review-site activity, job postings, or visits to your own site. First-party intent is what happens on your properties; third-party is bought.
Lead scoring: Ranking contacts or accounts by fit and by engagement, usually as two separate scores. Combining them into one number is the most common way scoring stops being actionable, because a great-fit account doing nothing needs a different response to a poor-fit account doing everything.; Lead scoring
Marketing qualified lead (MQL): A contact whose behaviour crossed a threshold marketing set. Its usefulness depends entirely on whether the threshold predicts anything, which is worth testing rather than assuming.; Lead scoring
Messaging: The words that carry a position to a specific audience in a specific place. Positioning is decided once and rarely; messaging is written many times and varies by role, channel and stage.
Mission: A multi-step piece of work run across several agents, planned before it starts and producing one or more documents at the end. The alternative is a single agent answering a single ask, which is what most requests actually need.
Model Context Protocol (MCP): An open standard for exposing tools and data to AI assistants, so a capability written once can be called from Claude, Cursor, ChatGPT and others. To a user it looks like new tools appearing inside the assistant they already use.; MCP setup
Multi-touch attribution: Dividing credit for a deal across the touchpoints that preceded it. Structurally limited: it can only credit what it can observe, and the most influential parts of a B2B decision are usually unobservable.
Net revenue retention (NRR): Revenue from existing customers this period against the same cohort a year ago, including expansion and after churn. Above 100 percent means the customer base grows without any new logos.
Orchestration: Deciding which agents run, in what order, and what each one receives from the ones before it. In practice it is a dependency graph: some steps can run in parallel, some must wait, and a final step usually needs to see everything.
Playbook: A reusable mission template: the phases, which agent runs each one, and what each phase receives. A playbook is to a mission what a recipe is to a meal.
Positioning: The decision about what you are an alternative to, and for whom. It is a choice with a cost, not a description. If your positioning statement contains no trade-off, no positioning decision has been made.
Product-led growth (PLG): A motion where the product itself does the acquiring, converting and expanding, usually through a free tier or trial. It works when someone can get real value alone, in one sitting, without a call.
Prompt caching: Storing the unchanging front portion of a prompt, such as a system prompt and a knowledge base, so repeat requests are cheaper and faster. It is what makes injecting a large shared context on every call affordable.
Retrieval-augmented generation (RAG): Searching a document store for the passages most relevant to a request, and putting only those into the model's context. It exists to solve a size problem: too much material to include wholesale.; Below roughly 200,000 tokens of source material, injecting everything with prompt caching usually beats retrieval, and skips a whole class of retrieval-quality bugs. RAG is a scale answer, not a quality one.
Reverse trial: Giving new users the paid experience for a fixed period, then dropping them to a limited free tier rather than cutting them off. It shows the ceiling first, which makes the eventual upgrade a decision about keeping something rather than buying something.; Pricing
Sales-led growth: A motion where a person qualifies, demonstrates and closes. The right choice when the purchase needs several people to agree, the configuration is genuinely complex, or the price makes a call cheaper than a trial.
The Dock: The review lane where finished work waits for a decision. Notes are pinned to the artefact, unresolved notes carry into the next round, and every approval is recorded against a specific version. Reviewing never costs credits.; The Dock
Token: The unit models read and bill in, roughly three quarters of an English word. Useful mainly as the currency of cost and limits: a 120KB knowledge base is about 30,000 tokens.
Total addressable market (TAM): The revenue available if every company that could buy did. Useful for judging whether a market is worth entering, and close to useless for planning a quarter, where the serviceable and obtainable subsets matter far more.
Value proposition: What a buyer stops doing, stops paying for, or stops worrying about once they buy. Stated from their side of the transaction, not yours.
World: The shared profile of one business that every agent works from: what it sells, who buys, its competitors, positioning, brand kit and uploaded knowledge. Built from the company's own site and research, and correctable by hand.; One world per brand. Agencies run a separate world per client so context, credentials and documents never mix.; For agencies

Seeing the words is not the same as seeing the thing

The interactive demo runs the whole loop on an example business: build a world, brief a specialist, review its work, approve a week. No signup.

Open the demo Read the guides