Insight series · Part 1 · 14 September 2026 · Part 2 · Part 3 · Part 4

From your question to an answer — AI jargon in plain English

The easiest way to understand the terminology is not as a dry glossary, but as what happens from the moment you type a question until an answer appears.

Think of a business AI assistant like Aadi. You type:

“What is our policy for refunding a customer who cancels after installation has been scheduled?”

Your question → Prompt → Retrieval / RAG → Context → LLM → Response → Checks / actions → You

Here is what each part means.

Follow one question all the way through

You ask Aadi:

“Can Mrs Smith cancel her solar installation and get her deposit back?”

1. Your question becomes part of the prompt

A prompt is not always just what you typed. The system may add instructions such as:

You are the customer assistant for this organisation. Use approved company information. Do not invent contractual terms. Escalate when uncertain.

So there is a user prompt (what you asked) and often a system prompt (how the assistant should behave).

2. The system works out what information it needs

The LLM knows a great deal of general knowledge from training, but it does not automatically know your latest contract terms or Mrs Smith's particular case.

That is where RAG — retrieval-augmented generation matters. The name sounds complicated; the idea is simple:

Retrieve the right information first; then give it to the AI to help generate the answer.

3. RAG searches your knowledge

Suppose your business has hundreds of documents. Instead of sending all of them to the LLM, the system searches for pieces most relevant to:

cancellation + solar installation + deposit refund

Embeddings help match meaning, not just keywords. A vector database is commonly used to store and search those representations.

The system might retrieve:

Customer terms, section 7.2 — cancellation after installation scheduling…

That material is added to the context for the LLM:

• your question
• instructions
• relevant retrieved company information

4. The LLM reasons over context and generates language

The LLM — large language model — is the language intelligence at the centre. It might conclude that Mrs Smith may be entitled to cancel, but whether the full deposit is refundable depends on costs already incurred.

Running the trained model to produce that answer is inference. That is different from training:

• Training — building or improving the underlying model from large datasets
• Inference — using the already-trained model to answer something now

When someone asks, “Is my company data being used by the AI?”, using company data as context for inference is not the same as using it to train the model. Policy, contracts, and architecture decide what actually happens — ask, do not assume.

From assistant to agent

Suppose instead you say:

“Deal with Mrs Smith's cancellation.”

Generating an answer alone is not enough. A digital worker might need to:

• identify Mrs Smith in the CRM
• check contract and installation status
• retrieve cancellation policy
• calculate any allowable deduction
• prepare confirmation
• ask a human to approve
• update systems and send confirmation

That is agentic AI — the LLM is still important, but it sits inside a larger system.

A useful split:

• LLM — intelligence
• RAG — access to relevant knowledge
• Tools — ability to interact with other systems
• Workflow — defined process
• Agent — intelligence using knowledge and tools to pursue a job
• Guardrails — boundaries on what it may do

That maps to how we talk about Aadi:

Process → job → digital worker → outcome

The LLM is not the product in that picture. It is one important component inside the worker — alongside models, knowledge, tools, workflows, permissions, and governance.

RAG vs training (often confused)

Imagine giving a new colleague your handbook. You have not retrained their brain; you gave them something to consult. That is broadly the intuition behind RAG.

Training or fine-tuning changes the model itself. RAG usually leaves the base model alone and supplies relevant information when needed.

When somebody says, “We trained our chatbot on all our company documents,” they often mean:

“We indexed our documents and use RAG to retrieve relevant information for the model.”

Those are technically different things.

One mental picture

• Prompt — what I'm asking
• Context — what the AI can see right now
• RAG — how it finds relevant knowledge
• LLM — the language / reasoning engine
• Inference — the engine running
• Tools — what it can use
• Agent — LLM + context + tools + objective
• Guardrails — what it is allowed to do
• Workflow — how the job should proceed
• Digital worker — all of that packaged around a defined business job

You are not buying access to an LLM.
You are packaging models, knowledge, tools, workflows, permissions, and governance around work that needs to get done.