Choosing the Right Kind of AI: Chat, Search, Notebooks and Assistants

Search for "best AI tool" and you will find an endless supply of lists, all out of date by the time you finish reading them. People who ask the question usually want one of two things: a single tool to commit to, or reassurance that they are not missing out. Neither list gives them either.

A more durable question is: what kind of help do I need for this task? The kinds of help change slowly. The brand names do not. Learn the categories and you can classify any new product in a minute.

• Six categories of AI help, and what each adds beyond a plain chatbot.
• Which category suits which kind of task.
• Which questions to ask about any tool: grounding, data handling, permissions and cost.
• A five-task test for choosing tools yourself.
• How to keep up without chasing every release.

You type, it replies. It works from what it learned in training plus whatever you put in the conversation. Many such products also offer extras like file upload, image input, saved instructions or memory, but the core is conversation with a language model.

• Good for: drafting, rewriting, explaining, brainstorming, structuring, translating, tutoring, code help.
• Weak at: current facts and exact figures (unless the product adds search), and anything about documents you have not supplied.
• Examples (as of October 2026): the main chat products from the large AI labs.

Within many assistants you can choose a model or setting that spends more effort working through a problem in steps. These usually help most on multi-step problems: maths, planning, tricky code, analysis with several constraints. They tend to be slower and may cost more, or use up usage limits faster. As noted in article 3, the makers' own prompting guidance can differ for these models. OpenAI's prompt documentation, for example, says general models do well with very precise instructions while reasoning models do better with high-level goals. A "reasoning" label is not a guarantee of correctness.

These combine a language model with a web search step. The model reads retrieved pages and writes an answer with links to the pages it used. Perplexity's developer documentation describes its approach as web-grounded answers with built-in citations. Many general assistants now offer a search option too.

• Good for: current information, quick orientation on a topic, finding sources to read.
• Weak at: quality control. The answer can only be as good as the pages retrieved, and the summary can still misstate what a page says. Links prove that a page exists, not that it supports the claim.
• Your job: open the cited pages for any claim that matters, and prefer primary sources, which Module 7 explains.

You add your own sources, such as PDFs, notes, web pages or transcripts, and the tool answers questions using only those, typically pointing back to the passages it used. NotebookLM is an example from Google. Google's Workspace announcement of January 2026 also described notebooks becoming usable as sources inside the Gemini app, which shows how boundaries between products move.

• Good for: studying from your own materials, summarising a set of documents, finding where in a source something is said.
• Weak at: anything outside your sources, and subtle misreadings. Fewer invented facts is not the same as none.
• Your job: click through to the cited passage for important claims.

These live in your email, documents, spreadsheets or meeting tools, and can use data from your account. Microsoft's documentation for its Copilot, for example, describes it as combining language models with content you have permission to access in Microsoft 365, and says prompts, responses and data accessed through Microsoft Graph are not used to train its foundation models. Other suites offer comparable features with their own terms.

• Good for: summarising your own threads, drafting from your own documents, saving context-switching.
• Weak at: the same hallucination issues as any model, plus a new risk: they can surface information you did not realise you were allowed to see, or draw from content that includes untrusted text. Permissions hygiene matters.
• Your job: use the approved version at work, understand what it can access, and check anything that leaves your hands.

Tools aimed at programming, in your editor, terminal or code-hosting platform. GitHub's documentation describes Copilot as able to suggest code as you type, answer questions about code, edit or refactor on request, review pull requests and, in agent mode, work through multiple steps toward a goal. Its documentation also says you remain responsible for reviewing and approving.

• Good for: boilerplate, explanation, tests, refactoring suggestions, navigating unfamiliar code.
• Weak at: judging your system's real constraints, security, and subtle bugs. See Module 5.

These stay useful even when the product list changes completely.

1. What does it see? Only my message? Files I attach? My whole mailbox? My code repository? More access means more power and more exposure.
2. Where do its facts come from? Training only, the web, my sources, or my account? Can I see the sources?
3. What happens to my data? Is it kept, for how long, can it be used to improve models, can I turn that off, and is there a business version with different terms?
4. What can it do on its own? Only talk, or also send, edit, delete, buy or run things?
5. How do I check its work? Citations to open? Code I can run? Nothing at all?
6. What does it cost, and what are the limits? Free tiers and limits change. Do not build a habit around a limit you have not read.
7. Can I leave? Can I export my notes, prompts and history? Dependence on one product is a risk that grows quietly.

If a vendor's documentation does not answer questions 1 to 4 clearly, treat the tool as higher risk for anything sensitive.

Rankings reflect someone else's tasks. Your tasks are the only ones that count. Here is a simple, honest way to compare two or three tools in about an hour.

1. Pick five real tasks that represent your work or study. For example: summarise an article you have read, explain a concept you know well, draft a short email, fix a small bug you already solved, answer a factual question where you know the right answer.
2. Write each request once, using the Describe skills from the next module, and paste the same text into each tool.
3. Score each answer from 1 to 5 on the dimensions that matter: accuracy (you can check, because you chose tasks you know), usefulness, clarity, and how well it followed your constraints.
4. Add two notes: did it show sources, and did anything feel unsafe or surprising?
5. Repeat in a month if you care. Results change.

A scorecard might look like this:

The numbers here are made up to show the format. Your results will differ, and a small test like this tells you about your tasks, not about the tools in general.

• Follow capabilities, not brands. When news arrives, ask which category it belongs to and which of your tasks it would change.
• Update your tools quarterly, not weekly. A short review every few months is enough for most people.
• Re-test before you re-trust. A model update can change behaviour on tasks that worked last month. Keep a few known-answer tasks as a "canary" to re-run after big changes.
• Keep your skills tool-independent. The templates in your playbook should work in more than one product.
• Do not let free-tier limits dictate your workflow. If a limit stops you mid-task, your process is too dependent on one tool.

• Using a general chatbot for current facts, when a search-grounded tool or a primary source would serve better.
• Treating citations as proof. They are a pointer, not a verdict.
• Using a consumer app with work data because it is convenient.
• Switching tools whenever a new one trends, and never getting good at any.
• Paying for a plan before you know your own use, or believing a premium plan removes the need to check.
• Judging a tool from one task, especially one you did not know the answer to.

• Next: Prompt Fundamentals: Goal, Context, Constraints and Format begins Module 2, where you learn how to ask.
• For using AI at work in particular, see Using AI Tools at Work Without Getting Into Trouble.
• For privacy questions about any tool, go to Privacy and Security: What Not to Paste, Connected Tools and Prompt Injection.