# The AI Assistant


Built into Primebrick there is an assistant you can talk to — in plain
language, like a colleague who knows the product inside out. You can ask it
how something works, ask it to find a record, or ask it to *do* something
for you — and it will tell you exactly what it is about to do before it
does it.

## Two kinds of questions, two kinds of answers

Everything you can ask falls roughly into two families:

| You ask… | The assistant… | Example |
|---|---|---|
| **About the product** | Answers from the product's own documentation — the same pages you are reading now | *"How does the audit history work?"* |
| **About your data, or to act** | Looks things up in the application itself — and, with your permission, performs actions | *"How many customers do we have?"*, *"Rename Acme Spa to Acme Ltd"* |

The difference matters: the first kind is a well-read librarian; the second
kind is a pair of hands that can actually touch your records.

## When the assistant wants to do something — it asks you first

Reading is free; **doing always goes through you**. When the assistant
decides an action is needed — renaming a customer, changing a setting,
deleting a record — it does not just do it. It tells you exactly what it is
about to change, and waits for your confirmation.

<Mermaid chart={`
flowchart TD
    Q["You ask the assistant<br/>to change something"] --> D["It explains what it<br/>is about to do"]
    D --> C{"You confirm?"}
    C -- "Yes" --> A["The action runs —<br/>audited like any other change"]
    C -- "No" --> N["Nothing happens —<br/>the conversation continues"]
`} />

Two consequences worth remembering:

- **Nothing the assistant does is invisible.** Changes made through it are
  tracked in the audit history exactly as if you had made them yourself.
- **It can only do what *you* can do.** It acts with your permissions —
  if a button is unavailable to you, it is unavailable to the assistant too.

## What it is good at (and what it is not)

**Ask it to:**

- Explain how a feature works, in plain language
- Find records: *"show me all customers in Italy"*, *"who is John Doe?"*
- Summarise a record or its history
- Take a concrete action you could take yourself — after your confirmation

**Do not expect it to:**

- Know things outside the product and its documentation
- Be right 100% of the time — it is helpful to verify important answers,
  and every answer it gives can be rated
- Replace your judgement on irreversible actions — the confirmation step
  exists precisely so the last word is yours

## Your conversations, your history

Each conversation is saved so you can pick it up later — you will find your
past chats in the conversations list, you can rename them, and delete them
when they are no longer useful. Conversations are yours: other users cannot
see them.

## Help it help you — the thumbs

Under each answer you will find a 👍 / 👎 rating. Use it: those votes are
read by the team maintaining the assistant and are the fastest way to flag
an answer that was wrong, vague, or unhelpful — or to confirm it nailed it.

## A worked example

> *"Jane Doe needs to update ACME Company's billing address. Instead of
> opening the customer record, finding the field, and typing it herself,
> she asks: 'change ACME Spa's address to Via Roma 12, Milano'. The
> assistant finds the customer, shows her the exact change it is about to
> make — old address, new address — and waits. She confirms; the record is
> updated and the change appears in the audit history under her name, as
> if she had done it manually."*