AI jargon explained for Customer Success Managers

LLM, agent, MCP, skill... AI is turning up in every Customer Success tool, from the CRM and the helpdesk to Customer Success Platforms. To make the right decisions, weigh up your options and talk to your product and tech teams as equals, you need to master the vocabulary. Here's the CSM's survival guide to the AI era!

1. LLM (Large Language Model)

An LLM is an AI model trained on billions of texts to understand and generate natural language. It's the engine that powers tools such as Claude, ChatGPT and Gemini. It doesn't "think" in the human sense: it statistically predicts the most relevant words in response to an input.

CS example: a Customer Success Platform such as Skalin includes an LLM that summarises a customer's situation from a range of signals and can produce an action plan or the outline of your next Business Review. What used to take 30 minutes of preparation before every call can now be read in 30 seconds.

2. Token

A token is the basic unit an LLM uses to read and produce text. It's neither a character nor a word: in English, a token is roughly ¾ of a word. "Hello" is 1 token, while a longer word like "re-engagement" can take 3. AI API usage is billed in tokens, and the size of the context window is expressed in tokens.

CS example: you use an LLM API to automatically generate a structured summary after every QBR. One hour of transcribed conversation is about 8,000 tokens. Multiply that by 50 strategic accounts per quarter and you can estimate the cost of this automation precisely before rolling it out.

💡 In practice: 1,000 tokens ≈ 750 words ≈ about 1.5 Word pages. Keep this benchmark in mind when sizing your prompts and budgets.

3. Prompt

A prompt is the instruction you give an LLM to get a response. The quality of the prompt directly determines the quality of the response. The art of writing effective prompts is called prompt engineering: a genuine professional skill, and not one reserved for developers.

CS example: rather than writing "analyse this account", a well-trained CS Manager writes: "You are a B2B SaaS Customer Success expert. Using this usage data and relationship history, identify the 3 main churn risks and suggest one concrete action for each. Be factual and concise." The result can be used straight away to prepare for a call.

💡 A good prompt includes a role ("you are..."), context, a specific task and an expected output format. The more structured it is, the more reliable the output.

4. Temperature

Temperature is a setting that controls how creative (or random) an LLM's responses are. It usually ranges from 0 to 1. At 0, the model is deterministic and factual. At 1, it's more inventive and varied, and potentially less reliable.

CS example: to generate customer health reports or QBR summaries, you set the temperature to 0: you want accuracy, not flights of fancy. To write personalised re-engagement messages or onboarding action plans tailored to each account's profile, a higher temperature produces more varied, less mechanical wording.

5. Context window

The context window is the maximum amount of text an LLM can "see" and process at once: your prompt, the conversation history, the documents provided and its own response. It's measured in tokens. The bigger it is, the more data the AI can analyse in a single pass.

CS example: with a large context window (200,000 tokens ≈ 150,000 words), you can analyse 12 months of conversations with a customer and ask: "What were the 3 warning signs we should have spotted before this churn?"

6. Embeddings

An embedding is a mathematical representation of a piece of text in the form of a numerical vector (a list of numbers). This vector captures the meaning of the text, not just its words. Two semantically similar sentences will have similar vectors, even if they don't share a single word.

CS example: your system turns all your CSMs' account notes, call summaries and success plans into embeddings. When a new CSM takes over a portfolio, they can ask a question in plain language (e.g. "what requests are still pending for this customer?") and instantly get the most relevant passages, even if they don't use exactly the same words. This is the mechanism at the heart of RAG.

💡 Embeddings are what make semantic search possible: your tool finds the relevant information, not just the content containing the same keywords.

7. Vector database

A vector database is a database designed to store and search embeddings. Unlike a traditional SQL database (which looks for exact matches), it finds the entries that are semantically closest to a query. Common options include Pinecone, Weaviate and pgvector.

CS example: you index all your Salesforce account notes, call transcripts and customer emails in a vector database. When a CSM is preparing for a call, they ask a question in plain language, such as "what integration problems has this account had?", and the system instantly finds the relevant passages across hundreds of documents.

8. RAG (Retrieval-Augmented Generation)

RAG is a technique that combines a vector database with an LLM: you first retrieve the relevant information from your internal data (using embeddings), then feed it into the LLM's context so it generates a response grounded in your up-to-date data, rather than in its training knowledge alone.

CS example: a CSM is preparing for a renewal call. They query the system in plain language: "what commitments were made at the last QBR, and have they been met?" RAG retrieves the relevant passages from meeting notes, emails and the Success Plan, then the LLM summarises a clear answer. The CSM goes into the call fully briefed in 2 minutes rather than after 30 minutes of manual digging.

9. AI agent

An AI agent is an LLM that can act, not just respond. It can plan a sequence of actions, use tools (web search, writing files, API calls) and iterate until it reaches a goal, without human intervention at every step.

CS example: a CS agent monitors accounts at risk of churn: it analyses usage data in your CRM, detects a drop in activation, drafts a personalised re-engagement email and creates a follow-up task in your project management tool, all without manual intervention.

10. Tool

A tool is a specific function an AI agent can call to interact with the outside world: running a web search, sending an email, reading a file, querying a database. Tools turn a passive LLM into an actor that can trigger real actions.

CS example: your QBR preparation agent has 3 tools: get\_account\_data() (Salesforce), get\_usage\_metrics() (your product) and create\_slide() (Google Slides). With a single command, it generates a pre-filled business review deck for every strategic account.

11. Skill

In an AI context, a skill is a predefined module that gives an agent the ability to carry out a specific task: writing an email, analysing a PDF, calling an API, creating an Excel file... It's a reusable functional building block that you "plug into" an agent.

How do you create a skill? A skill takes the form of a text file (often a `SKILL.md`) containing two elements: a description (which tells the agent when to use the skill) and instructions (which tell it how to carry it out). You can attach additional resources: scripts to run, reference files to consult, templates to use. You write a first version of the skill, test it on real cases, assess the results and then iterate until its behaviour is reliable enough to share with the team.

CS example: your CS agent has an "NPS analysis" skill: it reads the raw responses to your latest survey, identifies recurring themes among detractors and produces a structured report with recommendations, without an analyst spending 4 hours in Excel.

12. MCP (Model Context Protocol)

MCP is an open standard (created by Anthropic) that defines how an LLM can connect to external tools and data sources securely and in a standardised way. Think of it as the USB of AI: a universal interface for plugging services into a model, without bespoke development for each integration.

CS example: an MCP-compatible Customer Success Platform lets Claude connect directly to your account data: checking the health score in real time, reading the success plan, updating goals after a call and triggering an alert to the right CSM. MCP saves you from rebuilding a specific integration for every tool in your stack.

MCP or CLI: which should you choose?

MCP is designed for interactive, continuous workflows: a user or an agent interacts with several tools in real time, over the course of a conversation. It's smooth, but every round trip with a tool consumes tokens, which can become expensive at scale.

💡 Choose MCP when a CSM interacts with their data in plain language every day, when the workflow is conversational, or when several tools need to be orchestrated on the fly.

13. CLI (Command Line Interface)

A CLI is a text-based interface where you type commands to interact with software, without a graphical interface. Tools such as Claude Code are used via a CLI. For a Customer Success Manager, understanding it means grasping why some powerful AI workflows (especially the most repetitive, high-volume ones) run in a terminal rather than through a point-and-click interface.

CS example: a CS Ops specialist uses Claude Code in the CLI to automatically generate the company's 80 monthly customer health reports in its standard format, from data exported from the Customer Success Platform. The command runs overnight, the reports are ready in the morning, and there's no interactive interface consuming tokens, so the token cost is very low.

CLI or MCP: which should you choose?

The CLI is designed for batch, repetitive, non-interactive processing: you launch a script, it runs, it produces a result. No conversational back-and-forth, no tokens wasted on conversation context. It's cheaper at scale and easier to schedule (cron jobs, automated pipelines).

💡 Choose the CLI when you're processing high volumes of data regularly and predictably (e.g. monthly report generation, batch scoring of 500 accounts, data extraction and transformation), when cost per token matters, or when you want to build AI into an existing automation chain with no human interface.

In short: MCP for real time and interaction, CLI for volume and automation.

14. Orchestration

Orchestration is the ability to coordinate several specialised AI agents to achieve a complex goal. A "conductor" agent breaks down a task, delegates to sub-agents and consolidates the results. It's the architecture behind today's most powerful AI systems.

CS example: to prepare your Business Review, an orchestrator simultaneously launches a Data agent (analyses the KPIs), a Risks agent (detects churn signals) and an Opportunities agent (identifies upsell potential). It compiles the results into a coherent report in under 2 minutes.

15. Level of autonomy (level of automation)

The level of autonomy describes how far an AI system can act without human approval. There are generally three tiers: suggest (the human decides), execute with confirmation (the human approves before the action) and execute with full autonomy (the agent acts alone and the human only sees the result). The higher the level, the greater the benefits (and the risks).

CS example: Level 1: the AI suggests a renewal action plan, and the CSM adapts and sends it. Level 2: the AI automatically sends onboarding sequences to new SMB accounts, and the CSM only steps in on accounts that don't activate. Level 3: the AI fully manages follow-up for low-touch accounts below a certain ARR, with no human intervention except for explicit escalation. Each level requires a different degree of operational maturity and trust in the model.

💡 The right question isn't "how far can we automate?" but "what does a mistake cost at this level?"

16. Human-in-the-loop

Human-in-the-loop (HITL) is a design principle that keeps a human in an AI system's decision loop at strategic points in the workflow. The AI processes, proposes or acts, but some decisions remain subject to human approval before they're carried out.

CS example: your churn detection agent flags an at-risk account and drafts a re-engagement email. Before it's sent, the CSM responsible receives a notification and can approve, edit or block it. The AI gains speed, and the human keeps control of decisions with a big impact on the relationship.

17. Guardrails

Guardrails are rules or filters applied around an LLM to frame its responses: banning certain topics, enforcing an output format, preventing the disclosure of sensitive information or blocking out-of-scope answers. They can be added at prompt level, system level or in a separate application layer.

CS example: your renewal preparation AI agent is configured with guardrails that stop it from making contractual commitments, mentioning discounts not approved by management or commenting on product weaknesses in an uncontrolled way. If the model tries to generate an unapproved sales argument, the guardrail blocks the response and redirects to a neutral message. The result: AI remains a lever for efficiency, not a legal or commercial risk.

18. Hallucination

A hallucination is when an LLM generates false information and presents it confidently, as if it were true. It isn't a "bug" in the traditional sense; it's a structural limitation: the model optimises for linguistic plausibility, not factual truth.

CS example: an AI agent without RAG generates an account summary with invented usage data that doesn't match reality, and the CSM uses it as is in a QBR. Their credibility takes a hit. That's why any AI agent handling customer data must be grounded in a source of truth (RAG) and include a human review step before any external communication.

19. Observability

Observability means being able to monitor in real time what an AI system is doing, why it's doing it and where it's failing. It covers conversation logs, agent execution traces, performance metrics and alerts on abnormal behaviour. Without observability, an agent in production is a black box.

CS example: your observability dashboard reveals that your churn detection agent generates 15% false positives on accounts in an expansion phase: it reads a temporary dip in usage as a risk signal, when it's actually a documented seasonal lull. You adjust the model accordingly. Without this monitoring, your CSMs would have over-contacted healthy accounts, damaging the relationship for no reason.

💡 Observability is a prerequisite for continuously improving any AI system in production. Deploying without monitoring means giving up on improving the model.

20. Evaluation (eval)

An eval is a testing protocol that measures the quality of an LLM's responses against a set of reference cases. You define typical questions, expected answers and scoring criteria. Evals let you compare models, detect regressions after an update and validate a deployment before it reaches your customers.

CS example: before deploying your customer health scoring agent, you build a set of 150 real accounts whose outcome you know (renewal, churn, expansion), along with the signals that should have predicted it. You run the model on these cases and measure its accuracy. If it agrees with your best CSMs' judgement less than 80% of the time, it doesn't go into production.

21. Fine-tuning

Fine-tuning means retraining a general-purpose LLM on a dataset specific to your field, so it adopts your vocabulary, tone and processes. It's an expensive exercise, justified when prompt engineering alone is no longer enough.

CS example: a B2B SaaS company fine-tunes its customer health scoring model on three years of history annotated by its CSMs: usage data, relationship signals, contract context and final outcome (churn or renewal). The model learns to weight the signals specific to their vertical market, a level of precision that's impossible to achieve with a general-purpose model and a prompt, however well written.

💡 Before talking about fine-tuning, try RAG and prompt engineering: they cover 80% of CS needs at a fraction of the cost and complexity.