AI automation terms, in plain English.
The words you will meet on this site and in AI proposals, explained without jargon.
- AI system
- Software that uses AI models to complete a piece of work end to end, such as answering an enquiry or processing a document, together with the integrations, checks, and monitoring around it. We use “AI system” rather than “AI agent” because buyers want the outcome, not one component.
- AI agent
- A component that can work out the steps of a task and take actions with tools you allow, such as looking up a record, booking a slot, or sending a message. Good agents work within limits, record what they did, and hand over to a person when unsure.
- Multi-agent system
- Several specialised AI agents, each responsible for one stage of a process, coordinated by an orchestrator. Useful when a process has distinct stages that need different skills; unnecessary when one well-built system will do. When multi-agent makes sense.
- Orchestration
- The layer that passes work between agents or steps, keeps them in order, retries failures, and pauses at checkpoints for a person to approve.
- Human in the loop (checkpoint)
- A point in an automated process where a person reviews, approves, or takes over before the work moves on. Every system we build has them wherever judgement or approval matters.
- RAG (retrieval-augmented generation)
- Assistants that answer from your own documents, with sources. The system first finds the relevant passages in your material, then writes an answer based only on them and shows where each part came from. Knowledge assistants.
- Knowledge graph
- A structured map of the people, topics, documents, and decisions in your records and how they relate, so related information can be found and explored together.
- LLM (large language model)
- The kind of AI model that reads and writes text. Different LLMs vary widely in quality, speed, and cost, which is why choosing the right model for each task matters.
- Evaluation
- Testing an AI system’s output against known good answers or checks, before launch and continuously afterwards, so you can see where it is right, where it is unsure, and when quality drifts.
- Model routing
- Sending each request to the most suitable model or path, so simple requests use cheaper, faster options and only harder ones use larger models. AI cost optimization.
- Caching
- Storing answers to repeated requests so the system does not pay to compute the same result again.
- Cost per process
- What it costs to run one instance of an automated process, such as one enquiry answered or one onboarding completed, including model, telephony, messaging, and hosting costs.
- BYOK (bring your own key)
- Using your own account and API key with an AI model provider, so usage is billed to you directly and stays under your control. Harnad Connect’s AI agents work this way.
- WABA (WhatsApp Business Account)
- The account Meta uses to give a business access to the official WhatsApp Business Platform, its phone numbers, and message templates.
- Embedded Signup
- Meta’s official flow for connecting a WhatsApp Business number to a third-party platform in a few steps, without manual setup.
- PWA (progressive web app)
- A web app that installs on a phone like a native app and works on any device, which suits floor and field teams.
