This AI glossary for companies exists to answer just one question: what each AI and automation term means for those running a business — not for those writing papers. AI agent, LLM, RAG, NLP, process automation: each concept explained in one direct sentence, with an example of how it appears in practice. In the end, every term points to the same thing: an Digital Employee from XMACNA that EXECUTES the process — understands, decides, acts, and records on the Intelligent Dashboard. You don’t need to become an AI expert. You need to know what to ask for.
The term is just a name. What matters is what it EXECUTES.
Client contacts via WhatsApp, voice, or active campaign.
The Digital Employee reads history, memory, and data from the Intelligent Dashboard.
It chooses the next action: reply, qualify, schedule, collect, or escalate.
The conversation turns into structured data, private notes, and updated opportunity.
Your team takes over when a human is needed, already with full context.
Almost every AI conversation for business revolves around three words. Understand these and the rest of the vocabulary will fall into place — because they explain, in order, what AI thinks, how it responds, and how it acts. Each becomes, in practice, a part of the Digital Employee.
An LLM is a model trained on large volumes of text to handle language: interpret what the customer wrote, even with typos or slang, and draft a natural reply. It can misread an ambiguous sentence, so what it does in a company depends on the instructions, the data it receives and the rules around it. XMACNA chooses the model for each task based on criteria for the case, such as test quality, cost and response time.
01 · The engineAn AI agent is software that uses a language model and tools to move a task forward: understand the request, check a system, log, book. It acts within the permissions and rules the company sets, and hands to a person whatever falls outside them. At XMACNA, the agent becomes a Digital Employee, hired for a role and connected to the channels and systems integrated during implementation.
02 · What it acts onRAG is the technique that retrieves passages from the company's documents (catalog, price list, policies) and hands them to the model before it answers. This helps ground what the agent says in what the company wrote, but it does not eliminate errors: the result depends on an up-to-date base, on retrieval bringing the relevant passage and on critical information being checked. Updating the agent's knowledge becomes, to a large extent, keeping those documents current.
03 · The knowledgeDictionary definitions don’t close sales. What changes results is what the concept EXECUTES when it becomes a Digital Employee. See the term on one side and what it truly delivers on the other.
NLP (natural language processing): AI understands text written any way.
Reads the customer’s message with slang, audio, and broken sentences — and responds in the brand’s voice.
Process automation: repetitive tasks run on their own, with rules and integrations.
Serves, qualifies, schedules, updates CRM, and follows up — no typing needed.
Agentic AI: AI not only chats but plans steps and uses tools until completion.
Checks the schedule, books the time, confirms, and sends reminders — closes the cycle.
CRM with AI: the relationship system fills and organizes itself.
Each conversation becomes a contact, stage, and note in the Intelligent Dashboard — pipeline always updated.
A token is the smallest unit of text a language model processes: a word, part of a word or a character. How many characters fit in a token depends on the model and the language.
In practice (illustrative): a long customer conversation, plus the agent's instructions and catalog excerpts, can reach thousands of tokens.
Watch out: many model providers charge for input and output tokens. Instructions and documents sent along with the messages also count, so the cost of an agent depends on how it is designed, not only on the number of conversations. See how much AI costs a company.
The amount of text the model can consider at once when generating a reply: its working memory at that moment. It is different from what the model learned in training.
In practice (illustrative): to answer an exchange request, the window needs the relevant parts: the exchange rule, what the customer has already said about the order and the order details.
Watch out: what is not in the window, the model does not see. In long conversations, older parts may drop out. Remembering a customer from one week to the next requires storing information outside the model, with clear rules on what to keep. See the risks of AI memory.
A prompt is the text sent to the model: the question, the context and the instructions. In an agent, the instructions on role, tone of voice, rules and limits (the so-called system prompt) are sent to the model on each run, according to how the agent is designed.
In practice (illustrative): "You serve store X. Only share delivery dates that are in the system. Discount requests above 10% go to the manager."
Watch out: written instructions guide the model but do not guarantee it always complies. That is why important rules also live in permissions and systems, and the prompt is tuned against a success criterion and tests with real conversations. See system prompts for AI agents.
An embedding represents a piece of text as a list of numbers (a vector). The distance between two vectors suggests how related the texts are: texts with similar meaning tend to sit close together.
In practice (illustrative): "I want to return the product" and "how do I exchange the product" tend to sit close, because they are about the same subject, although the intent is not identical: a return means money back, an exchange means another item.
Watch out: closeness suggests a relationship; it does not guarantee that the text found answers the question. See AI knowledge management.
Search that uses embeddings to bring back passages potentially relevant to the question, even without the exact word. It is one of the ways RAG, described above, chooses what to hand to the model.
In practice (illustrative): a customer asks "can I return it after using it?" and the search may bring back the policy paragraph about used products, even though it never says "return".
Watch out: the search may bring back a passage that is only similar, not the right one. An outdated document can contaminate the answer, and duplicate or contradictory bases raise the risk of error. Maintaining the base and checking what was used is part of the job.
Also called function calling. The model decides, based on the request and a tool's description, that it needs an external action and returns a structured request. The software responsible for the tool executes it, which may be the company's application or a service run by the model provider, under the configured permissions; the result goes back to the model.
In practice (illustrative): a customer asks where their order is; the agent calls the "look up order" tool with the order number, the system returns the status and the agent writes the reply.
Watch out: if a required piece of data is missing, the model may try to infer a value. Actions that change something (cancel, charge, update a record) need confirmed data and well-defined permissions. See system integration.
When the model confidently states something wrong or made up. NIST calls it confabulation.
In practice (illustrative): without access to the delivery date, the model answers "it arrives in three days" because it sounds plausible.
Watch out: letting the agent say "I don't know", answering from company documents (RAG, above) and checking information at the source reduce the problem but do not eliminate it. Critical information (price, deadline, contract) must come from a system or document, and quality should be measured by verified resolution. See AI customer service metrics.
A design in which a person reviews, approves or takes over defined steps of an AI agent's work, with clear roles and responsibilities.
In practice (illustrative): a discount above the limit goes to the manager for approval; a complaint goes to a person with the full history.
Watch out: review that turns into a rubber stamp does not protect anyone. NIST points to risks such as automation bias and over-reliance on the system; whoever approves needs the time, criteria and authority to say no. See human intervention in AI customer service.
Not sure where to start? The free assessment has seven questions and shows a transparent estimate of the time your team could regain. If you want to go further afterwards, Hermes continues the conversation on WhatsApp.
It’s a list of artificial intelligence and automation terms — AI agent, LLM, RAG, NLP, process automation — explained simply, by what each concept means in practice for your company.
LLM is the engine that understands and generates language; the AI agent uses this engine to decide and execute the task. XMACNA’s Digital Employee is the agent — the LLM is just one piece of it.
RAG makes the agent retrieve passages from your own documents before answering. This helps ground answers about products and prices in what the company wrote, but it does not eliminate errors: the base must be up to date and critical information must be checked.
No. XMACNA delivers the Digital Employee installed, trained, and operating. You only need to know what your business bottleneck is — the technical translation is our problem.
to deliver extraordinary results in your Digital Employee.
Combining the best of each model for superior results.
Free assessment. In 30 days it is answering customers.
Still unsure where to start? Take the AI Assessment — 7 questions, 90 seconds, free.