Glossary

The words that come up in every automation project

Written for the person on the buying side of the call. Each entry says what the term means, and where it matters, in the fewest words that stay accurate.

The basics

Automation: A system that carries out a business process without somebody driving it step by step. The test is whether it runs correctly at 2am with nobody watching, which is what separates it from a shortcut or a macro.
Business process: A repeatable sequence of steps with an owner, an input and an output. Most automation projects fail at this definition rather than at the code, because the process turns out to have four undocumented variants.
Trigger: The event that starts an automated run: a form submitted, an email received, a row added, a clock reaching a time. Everything downstream is a consequence of choosing the right one.
Webhook: A URL one system calls the moment something happens, so the other system reacts immediately instead of polling for changes. The usual way a form, a CRM or a payment provider tells your automation that something is worth doing.
Integration: A connection that lets two systems read and write each other's data. Integration count drives project cost more reliably than anything else, because each one carries its own authentication, rate limits and edge cases.
API: The documented way to ask a system for data or tell it to do something, without a human clicking through its screens. When a vendor has no API, automating around them means screen-scraping or a file drop, and both are fragile.
No-code / Low-code: Tools that build working software through configuration rather than typing most of the code. They cover the common 80 percent quickly, and the remaining 20 percent is usually where the business logic that matters lives.

AI and agents

LLM: The model underneath tools like Claude and GPT. It predicts text, which is why it is excellent at reading, classifying, drafting and extracting, and why it needs checking on arithmetic and on facts it was never given.
AI agent: An LLM given tools, permissions and a goal, so it decides which steps to take rather than following a fixed path. The difference from a chatbot is that an agent changes something in a real system.
Prompt: The instructions a model is given for a task, including its role, its constraints and the format of the answer. In production it is a versioned asset, closer to code than to a message.
Context window: How much text a model can hold in mind at once, prompt and answer together. Long documents get chunked or summarized because of it, and quality falls off before the limit rather than at it.
RAG: Fetching the relevant passages from your own documents and handing them to the model with the question. It is how a system answers from your contracts and procedures instead of from what the model absorbed during training.
Hallucination: A confident answer that is wrong, usually because the model was asked something its input did not cover. The engineering answer is to give it the source and require a citation, rather than to ask it to try harder.
Fine-tuning: Further training of a model on your own examples so it adopts a format or a style. Rarely the right first move: prompting and RAG solve most of what people reach for it to fix, at a fraction of the cost.
MCP: An open standard for connecting a model to tools and data sources, so one integration works across different AI clients. It is the reason an agent can read your calendar and your database without a bespoke bridge for each.
OCR: Turning an image of a document into text a system can process. Hebrew, handwriting and scanned tables are the hard cases, and they are exactly the ones Israeli document projects run into.
Structured output: Forcing a model to answer in a fixed shape, usually JSON against a schema, so the next step can rely on the fields existing. Without it, one unusually chatty answer breaks the workflow downstream.

Tools and platforms

n8n: An open-source automation platform where a process is built as connected nodes, with real code available wherever the visual blocks run out. It can run on your own server, which is why regulated clients tend to land on it.
Workflow: One automated process from trigger to result, as built inside a platform. Keeping each one small and single-purpose is what makes a failure debuggable a year later.
Node: A single step in a workflow: read a row, call an API, ask a model, branch on a condition. Node count is the rough unit of complexity when scoping a build.
Self-hosted: Running the software on infrastructure you control rather than the vendor's cloud. It costs setup and maintenance, and it is often what makes a project possible at all under a security review.
Zapier / Make: Hosted automation platforms priced per task or per operation. Fast for simple connections between two apps, and the per-run pricing is what pushes high-volume processes elsewhere.
Vector database: Storage that finds text by meaning rather than by exact words, which is what makes RAG work over documents phrased differently from the question. Usually a component inside a system rather than a product a business buys directly.

The commercial side

ROI: What the project returns against what it cost, usually expressed as the month it stops being an expense. Hours recovered and money recovered are different claims, and only the second one reaches the accounts.
IP: Who owns the code and the system after it is built. Worth settling in writing before work starts, because the difference between owning a system and renting access to one shows up the day you want to change vendor.
SOW: The document that says what will be built, what counts as finished, and what it costs. The acceptance criteria in it are the part that decides arguments later, so they are worth reading twice.
POC: A small build that answers one question about whether the full thing will work. A POC with no stated question is just a small project, and it tends to become the system by accident.
Retainer: A standing monthly amount of a supplier's time rather than a fixed scope. It suits ongoing improvement and a queue of small processes, where writing a scope document per change costs more than the change.
SLA: The committed response and repair times when something breaks. For an automation running unattended, the monitoring that detects a failure matters as much as the hours promised for fixing it.

Questions

Which of these terms actually matter in a first call?

Process, trigger, integration and IP. The first three decide what a project costs, because scope is really a count of processes and the integrations they touch, and the fourth decides what you are left holding at the end. The AI vocabulary matters later, during the build.

What is the difference between a chatbot and an AI agent?

A chatbot answers. An agent has tools and permissions, so it changes something: books the meeting, updates the record, sends the document. The moment a system can write to another system, the questions that matter become permissions and auditing rather than answer quality.

Do we need RAG or fine-tuning to use AI on our own documents?

Almost always RAG. It puts the relevant passage in front of the model at the moment of the question, so the answer follows the current version of the document. Fine-tuning teaches format and style, and cannot keep up with content that changes.

Is self-hosting worth the trouble?

It depends on who is reviewing you. If personal data, health data or client contracts pass through the system, self-hosting is usually what makes the security review passable at all. For a marketing workflow moving public information, it is overhead with no return.

How do we tell a real automation from a shortcut?

Ask what happens when it fails at 2am. A real one has a trigger, error handling, a retry policy and somebody who gets told. A shortcut runs when a person remembers to run it, which means it stops working the week they go on holiday.

Something here you want applied to your own process?

A mapping call turns the vocabulary into a ranked list of your processes, with hours and owners attached.

Tell us what's slowing your business down.

30 minutes. No pitch, no deck. Just listening to your needs and seeing how we can help.

Schedule a discovery call

orMessage us on WhatsApp