30 AI Terms every executive should know

An AI system that perceives its environment, reasons and executes actions autonomously to achieve a goal.

Unlike a chatbot that only replies, an AI agent can chain steps: query a database, draft an email, update a CRM and send a notification, all without human intervention. AizuaLabs deploys agents specialised by business vertical.

An instruction or message sent to an AI model to obtain a response or action.

The prompt is the 'command' you give the model. In enterprise production, the prompt is usually a long, structured text (system prompt) that defines the agent's role, tone, constraints and context. Prompt quality largely determines the quality of the result.

The smallest unit of text that an LLM processes. It equals roughly 0.75 words in English or 0.6 in Spanish.

LLMs charge per processed token. Knowing your agent's token usage helps predict costs. A typical email is ~300 tokens; a customer-support conversation, ~2,000 tokens. AizuaLabs optimises prompts to cut usage without losing quality.

The use of artificial-intelligence systems to carry out repetitive or complex tasks without constant human intervention.

It goes beyond RPA (robotic process automation) because the system can understand context, make decisions and adapt to variations. Examples: classifying and replying to emails, processing invoices, qualifying leads, generating monthly reports.

An automated conversation system. In its traditional form it follows decision trees; in its AI form it uses LLMs to reply in natural language.

Classic chatbots are rigid: if the question does not fit their tree, they fail. LLM-based chatbots understand intent and can answer unforeseen questions. An AI agent is a chatbot that can also execute actions (look up orders, create tickets, send emails).

A type of artificial intelligence capable of creating new content: text, images, code, audio or video.

LLMs are the best-known example. Unlike discriminative AI (which classifies or predicts), generative AI produces original outputs. In a business context, it is used to draft emails, generate reports, create marketing content or automate customer responses.

Large Language Model. A large-scale language model trained to understand and generate natural-language text.

GPT-4, Claude or Gemini are LLMs. They are trained on trillions of text tokens and learn patterns of human language. Companies use them as the 'brain' of their AI agents, connecting them to their own data and systems.

The process of further training a pre-trained model with data specific to a domain or company.

It lets the model learn your company's vocabulary, processes and style. It is more expensive than RAG but produces highly specialised models. For most SMEs, RAG offers a better cost-benefit ratio than fine-tuning.

The maximum amount of text (in tokens) a model can process in a single call.

Claude 3.5 has a 200,000-token context window; GPT-4o has 128,000. A large context lets the agent recall long conversations, analyse entire documents or work with multiple sources at once. It is a key factor when choosing a model.

A parameter that controls the randomness of an LLM's responses. Low values = more deterministic; high values = more creative.

Temperature 0 always picks the most likely answer (ideal for support or data agents). Temperature 0.7-1.0 produces more varied responses (useful for content generation). Business agents are usually configured with a low temperature to be predictable.

An LLM able to process several data types: text, images, audio, video or code.

GPT-4o, Claude 3.5 Sonnet and Gemini 1.5 are multimodal. They let you analyse scanned invoices, describe product images or transcribe audio calls. For business agents, vision capability is key in document-handling processes.

Retrieval-Augmented Generation. A technique that combines an LLM with search over your own documents to answer with up-to-date, verifiable information.

Instead of relying solely on the model's knowledge (which has a cut-off date), RAG first searches your knowledge base (manuals, contracts, FAQs) and then the LLM formulates the answer. It reduces hallucinations and gives the agent a 'memory' of your company.

A system instruction that defines an AI agent's behaviour, role and constraints before any conversation begins.

It is the agent's 'constitution'. It covers who it is (name, company, sector), how it should speak (tone, language), what it can and cannot do, and how to escalate complex situations. AizuaLabs writes and optimises the system prompt for each client.

A numerical (vector) representation of the meaning of a text, used to compare semantics across documents.

When an agent searches your knowledge base, it converts each text fragment into a numerical vector. When it receives a question, it also turns it into a vector and finds the most semantically similar fragments. It is the search engine behind RAG.

An LLM's ability to invoke external functions (APIs, databases) as part of its reasoning.

The model decides when and how to call a tool. Example: the agent receives 'how long will my order take?' → calls the tracking API with the order number → interprets the response → replies to the user. Without tool calling, the agent can only generate text.

The discipline of designing and optimising instructions to get the best results from an AI model.

It includes techniques such as few-shot prompting (giving examples), chain-of-thought (asking for step-by-step reasoning), role prompting (assigning the model a role) and retrieval augmentation. It is a critical skill for building reliable production agents.

A technique of including input-output examples in the prompt to guide the model toward the desired format or style.

Instead of explaining what to do, you show the model how to do it with 2-5 examples. It is especially useful for tasks with a specific format (classifying emails by type, extracting data from invoices, generating responses with a fixed structure).

A technique that asks the model to 'think step by step' before giving an answer, improving accuracy on complex tasks.

It is triggered by phrases such as 'reason step by step' or 'explain your process before answering'. It significantly improves performance on logic, mathematics or multi-step analysis problems. Some models do it internally (reasoning models such as o1 or Claude 3.5).

A database specialised in storing and searching embeddings efficiently.

Pinecone, Weaviate and pgvector (PostgreSQL) are examples. They let you find 'the most semantically similar' among millions of fragments in milliseconds. It is the infrastructure piece that makes RAG possible at scale.

Application Programming Interface. An interface that lets two software systems communicate with each other.

AI agents access data and services through APIs: your CRM's API to read customers, your ERP's API to check stock, Stripe's API to verify payments. Without API integration, the agent can only generate text, not take action.

A mechanism by which one system sends data to another in real time when an event occurs.

When a customer makes a purchase in your store, Stripe sends a webhook to your AI agent with the order data. The agent can then send a welcome email, create a task in the CRM and notify logistics. It is the foundation of real-time automation flows.

The phenomenon whereby an LLM generates false or made-up information with apparent confidence.

It is the main risk of AI systems in production. It is mitigated with RAG (the model answers based on real sources), response validation and human oversight. AizuaLabs designs agents with guardrails that reduce the likelihood of hallucination.

Constraints and filters that limit an AI agent's behaviour to prevent harmful or out-of-scope responses.

A customer-service agent with guardrails will not answer about politics or religion, will not make up prices and will escalate to a human when it detects dissatisfaction. They are the difference between a chatbot that 'goes rogue' and one that acts like an employee of your company.

Meta's official API for integrating WhatsApp into business systems and automating conversations at scale.

It lets you send and receive WhatsApp messages from external platforms. It is the foundation of the AizuaLabs WhatsApp Sales Bot. It requires Meta approval and a dedicated phone number. The cost includes a conversation fee (~€0.04-0.13 per conversation depending on type).

A component that coordinates multiple AI agents or tools to complete complex tasks.

In a multi-agent system, the orchestrator receives the task, decides which agent or tool to use at each step and combines the results. For example: a sales agent orchestrates a CRM search, a product-catalogue query and the drafting of a personalised quote.

An AI agent specialised in a specific sector or business function, with knowledge and workflows tailored to that domain.

A real-estate vertical agent understands the sector's vocabulary, knows how to qualify based on whether the customer is buying or renting, and is familiar with the steps of the property sales process. It is more effective than a generic agent because its responses and actions are tuned to the reality of the sector.

Extracting text from images or PDFs using vision models, with semantic understanding of the content.

Modern vision models not only read characters (classic OCR) but understand the context: they know whether a number is a price, a date or a product code. AizuaLabs uses this in its Accounting Agent to process scanned invoices automatically.

The time an AI agent takes to process an input and return a response.

In enterprise chat applications, latencies of 1-3 seconds are acceptable. For background automation flows (processing invoices, generating overnight reports), latency is not critical. The choice of model and the length of the prompt directly impact latency.

Software as a Service based on artificial intelligence. An AI service accessible under a monthly subscription, with no infrastructure to manage.

The AizuaLabs model is SaaS: the client pays a monthly fee and receives the agent configured, operational and maintained. They need no servers of their own, nor a technical team. The provider (AizuaLabs) manages the LLMs, the integrations and the updates.

A Spanish government grant programme to digitalise SMEs and freelancers, with vouchers of up to €29,000.

Within Kit Digital there are categories compatible with AI solutions: 'Artificial Intelligence and Analytics' (up to €6,000), 'Secure Communications' and 'Customer Management'. AizuaLabs supports companies in applying for these grants as part of its consulting service.