Search versus AI: how AI finds patterns, and how to feed it

What it is

There are three different ways to look for help with a piece of work, and they behave very differently.

• Boolean search matches exact words. invoice AND late NOT paid finds documents that contain both words and not the third. It is precise, but it misses anything phrased differently, such as "overdue bill".
• Web search ranks pages that already exist. Modern search engines understand some meaning, but they give you a list. You still read, compare and combine.
• An AI assistant does not hand you a list. It writes a new answer from patterns it learned in training, shaped by what you give it. That is why it can feel like a black box: you see what goes in and what comes out, but not the steps in between.

So the skill changes. With search, the skill is choosing keywords. With AI, the skill is gathering the right material and angles before you ask.

How the black box works (simplified)

This picture is a simplification, but it is a useful way to think.

An AI model turns words and ideas into vectors: long lists of numbers that place each idea as a point in a "meaning space". Ideas with similar meanings sit close together and form clusters: one area for finance terms, another for polite but firm customer tone, another for late-payment rules.

When you write a prompt, your words point the model towards certain areas. A thin prompt ("write a reminder email") points at one area, so you get a generic answer. A rich prompt that names the domain, the audience, the tone, an example and the rules points at several relevant areas at once. The model can then combine patterns from all of them, which gives a more specific and more useful answer.

To be precise: when the AI answers from its own training, it is not searching a database of clusters. It is generating text from what it learned. The vectors are real, but the tidy "clusters" you can point at are a simplification. The picture explains why richer, well-chosen context helps; it is not a literal map of the machine.

When AI searches your documents: often a real vector search

When an assistant searches a large set of your documents (a project, a knowledge base or a connector), it often uses vector search, and there vectors are used literally. Your documents are cut into short passages, each passage is turned into a vector and stored, and your question is turned into a vector too. The system then pulls in the passages nearest in meaning to your question, and the AI writes its answer from them.

This has a practical consequence. If your question uses one narrow set of words, the search finds one corner of your documents. If you describe what you need from several angles, it can pull from more of them.

Why it matters

• Search rewards the right keywords. AI rewards the right context.
• A richer, well-chosen prompt draws on more relevant patterns, so the first answer is closer to what you need.
• More is not always better. Off-topic material pulls the answer in the wrong direction. Aim for several relevant angles, not everything you have.
• AI can still be confidently wrong. Richer input raises the chance of a useful answer; it does not guarantee a correct one. You still check.

How to do it: gather before you prompt

Before you write a prompt, or brief an agent, write the task in one sentence, then collect material from eight angles.

1. State the goal: what outcome do you want, and by when? (If you cannot write the task in one sentence, the AI cannot either.)
2. Describe the audience: who will read or use it, and what do they already know?
3. Add domain knowledge: the terms, rules of the trade and context an outsider would miss.
4. Show examples: one or two pieces of good past work to copy in style or structure.
5. Give sources and data: the documents, numbers and facts it must use. Approved data only.
6. Set constraints: length, tone, format, what to leave out.
7. Name the risks: what must not go wrong.
8. Define what good looks like: how you will judge the result.

You will not need all eight every time. For a quick email, three or four are enough. For a report or an agent, use them all.

Example

Thin prompt:

Write a reminder email about a late invoice.

Result: a polite, generic email that could be from any company.

Rich prompt (same task, several angles):

You are our accounts receivable lead. Write a reminder to [customer],
a key account for five years, whose invoice [number] for [amount]
is 30 days overdue.
Tone: warm but firm, like the example below. Mention our payment
terms of 30 days. Do not threaten fees.
Offer a call with [name] if there is a problem we can solve.
Under 150 words.
Example of our tone: [paste a past email you liked]

Result: an email that sounds like your company, respects the relationship and still asks clearly for payment.

The same idea for agents

When you build several agents, give each one its own angles rather than giving every agent everything. Each agent then works from a focused, relevant set, and a coordinator (and then a person) combines their work.

Common mistakes

• Treating AI like a search box. Typing three keywords. Fix: write the task, the audience and the goal in full sentences.
• Pasting everything. Ten documents "just in case". Fix: choose the few that matter for this task.
• One angle only. Asking only "what", never "for whom" or "what good looks like". Fix: run through the eight angles.
• Trusting a rich answer more than a thin one. A better-written answer can still be wrong. Fix: check facts and numbers against the source.

Try it this week

1. Pick a task you would normally search the web for help with.
2. Write the task in one sentence.
3. Spend five minutes noting the eight angles on paper.
4. Ask the AI twice: once with a thin prompt, once with your angles.
5. Compare the two answers, and keep the better prompt as a template.

Going further

• Ask the AI for the angles: "Before you answer, list what else you need to know to do this well." Then supply what it asks for.
• For connected documents: ask the same question in two or three phrasings, or ask it to "find passages about [topic A], [topic B] and [topic C]", so the search covers more of your material.
• For agent environments: write each agent's angles into its instructions, and pass only the relevant material in each hand-off.

Key words

• Boolean search: search with exact words joined by AND, OR and NOT.
• Embedding: another name for the vector a model makes from a word or passage.
• Vector: a list of numbers that places a word, passage or idea in a "meaning space".
• Cluster: a group of vectors close together because their meanings are similar.
• Vector search: finding the stored passages whose meaning is nearest to your question.
• Black box: a system where you see the input and output but not the steps in between.

Examples and figures in this guide are illustrative. AI tools change often: check what your organisation's approved tools offer today. Use approved tools only, and a person checks and decides.

Prefer to print or share it? Download this guide as a PDF (399 KB), or the complete AI Field Guide (1.7 MB).