Session 2 · Core concepts
Core concepts for using an AI agent on QuickSight
AI concepts, Claude, and the Factor BI skill
Before you start asking business questions, it pays to be clear on five concepts: which model to choose, what a skill is, why the agent needs tools and permissions, how projects work, and why everything valuable gets documented.
Project scope
The goal is to connect Claude to the QuickSight dashboards your company already uses, so you can ask them business questions in plain language.
- It starts with a single dashboard. Usually Sales: it covers sales, orders, delivery notes and quotes, and compares against any prior period you pick, not just last year.
- Any other management dashboard (purchasing, items, stock, accounts receivable) can be connected in later phases. Each one needs its data mapping done first.
- Each user has their own QuickSight access, and that access is what connects to the dashboard. The agent sees exactly what that user is allowed to see.
- One scope at a time. Any other analytics projects you have running stay separate from this one; they do not get mixed into the same conversation.
What you can ask it
The agent does not just hand back a table: it interprets the question, builds the query, and analyzes the result.
- Business metrics in your own vocabulary: sales, orders, delivery notes, quotes, margins, profit.
- Acronyms and shortcuts:
YTD(year to date) or "give me the 80/20": the agent understands them without being told. - Cuts and filters: by product family, top 10 customers, top 10 sales reps, portfolio concentration.
- Comparisons: against whatever period you name, with changes in percent and in percentage points.
- Business insights: when you ask for them explicitly, it flags which movements matter, connects findings to each other (for example, one sales rep's drop against one customer's drop) and proposes hypotheses you can confirm or rule out.
- Follow-up questions: on the same result, without restating the whole context.
Use the word dashboard in your question, so the agent knows to go to QuickSight. And ask for the query plan before it runs: that way you check it understood the question before it pulls any data.
> From the Sales dashboard, give me YTD sales for the "ABC" line, with top 10 customers and top 10 sales reps against last year. Show me the query plan first. At the end, give me business insights.
Insight the agent returns: "This is quality growth: profit is rising faster than sales, so it is not just volume. Two things to look into: one customer went from $2.1 M to $4.7 M (+124 %), an unusual jump worth explaining; and one sales rep's decline matches one customer's decline almost one for one, which points to a reassigned account rather than a real loss."
The figures above are made up to show the answer format. In your implementation, the data comes straight from your dashboard.
Models: choosing who you talk to
Claude is the application. The model is the brain that processes the question, and it does not live on your computer: it lives on Anthropic's servers. You pick which one to use when you start each conversation.
Haiku is like an intern three months into the job: quick to answer, but the reasoning is not deep. Opus is the specialist you consult on a complex matter. It makes no sense to ask the CEO to answer a simple question, or to ask an intern for a portfolio analysis.
What a skill is
It is the answer to the obvious question: if I type "give me sales", how does the agent know it has to go to my company's dashboard? Because it has prior training called a SKILL.
- A skill is a package of instructions sent to the model along with your question, without you seeing or writing it.
- Ours is called Factor BI QuickSight Advisor. It holds the agent's identity, its communication rules, and a numbered sequence of steps it must run to answer.
- It is installed once from Claude's settings; from then on it works in every conversation.
- There are skills for many things besides dashboards: it is a general way to teach the agent a specific task.
Tools and permissions
The same model that talks with you can also work with tools: use resources on your computer and connect to outside services. That is what lets it query the dashboard.
Conversation only
It talks, reasons and answers with what it knows.
- Talks and reasons
- Can search the web
- Creates files, but does not change any existing folder or file on your computer
With tools
The same model, with controlled access to resources on your computer and to outside services.
- Reads and creates files: Word, Excel, PowerPoint, text
- Runs code in a controlled way
- Connects to Google Drive, Gmail, Calendar, meeting-notes apps and more
- It is what you need to query the dashboard
- Permissions are explicit. The agent does not walk into any folder it likes: you authorize each one. And each connector is authorized once; after that it does not ask again.
- Example from the session: it was asked to connect to Google Drive, find every file related to a client (Comercializadora Bajío, in the example) and build an index of them in Word, grouped by category: manuals, data dictionaries, videos, technical documents. That index did not exist; it was built in the conversation.
Projects
A project groups several conversations on the same topic, with fixed instructions and its own folder of files.
- Project instructions: a text the agent reads at the start of every conversation. That is where the goals, the decisions already made, the current state of the topic and the operating rules go.
- File folder: the agent can consult any of those files whenever it needs to. It does not read them all every time.
- They can be shared: a project can be shared with other users in your organization, and they can open new conversations from what is already documented.
- One per topic. Keep sales, accounts receivable and commissions in separate projects rather than mixing them.
Inside a project you open several conversations, just as you open several email threads with a client: one for billing, one for the contract, one for the project definition. Nobody would think of mixing all three in a single thread.
Document everything important
Conversation 2 cannot read conversation 1. And the agent cannot search through its past conversations.
That is how the agent works today in a local Cowork project, which is how the QuickSight connector runs: even inside the same project, conversations do not share history. If something valuable came up in one conversation and stays there, for practical purposes it is gone.
- The fix: as soon as a useful finding comes up, ask the agent to document it in a file inside the project folder. Literally: "document this topic precisely and save it to a file".
- Format matters. Use
.mdfiles: plain text, very easy for the agent to read and edit, and where it keeps adding information. Word and Excel are harder work for it. .mdfiles look like code, but there are viewers that display them nicely, and you can edit them by hand whenever you want.- A simple rule: any valuable information produced in a chat has to land somewhere: an
.md, a spreadsheet, a document. Never only in the chat.
The value is not in any single answer, but in the knowledge that builds up in the project folder: definitions, analysis criteria, confirmed findings. That is what makes the next question come out better than the last one.
And once the analysis is right, the agent can deliver it in whatever format you need:
Topics for later sessions
With the concepts above you can start asking. These topics are covered in the following sessions, as your use calls for them.
