Assistants in the browser

In Ajutant for Chrome, any assistant can read the page the person has open when their message is about it, such as “summarise this”. To let an assistant read other pages, move between or act on pages itself, give it a scope and switch on what it may do. A team admin then approves that for their team.

Its sites

Open the assistant, then Tools → Browser. Add the sites it works on. Each entry is a host, and optionally a path it must start with:

• linkedin.com with path /in/ covers profile pages such as www.linkedin.com/in/ada, and not the feed.
• docs.example.com with no path covers the whole site and its subdomains.

Keep the scope as narrow as the job. Sites your organisation has blocked can never be used, whatever you add here.

What it may do

Three switches, each limited to its sites:

• Read pages: read the text and the controls of the page open in the tab.
• Move between pages: open other pages of its sites in the same tab.
• Act on pages: click, type, choose options and scroll.

When it needs a page outside its sites, it asks the person, who can allow that page or that site for the session. That covers reading and opening only: acting outside its sites is never possible.

Note

Whatever you switch on, anything that sends, posts or submits stops for the person to review. They see what will be sent and where, can change the text on the page, and confirm or cancel. A button Ajutant cannot identify waits too.

Test it, then ask for approval

Your changes are a draft until you publish. People who can test the assistant can choose Run the draft in the panel to try it on real pages; those test runs are recorded.

When you publish a change that lets the assistant do more, each team that has it gets a request to approve, with your test runs attached as evidence. Until a team approves, its members get what was approved before. Publishing a change that does less applies at once, with no approval.

Tasks as buttons

A task can be a button in the panel. In the task’s Trigger step, switch on Show as a button in Ajutant for Chrome and give it a short label, such as Draft intro. Pressing it runs the task with the page attached; write {{page.title}} or {{page.url}} in the instructions to name the page.

The button works on the assistant’s sites. Elsewhere it is disabled with a note of where it works (Shown, disabled), or Hidden, as you choose under Tools → Browser.

For a team that has not approved acting on pages, the task reads and drafts: the assistant writes the message or answers in its reply, and the person pastes and sends them.

An example: a connection note on LinkedIn

An assistant with your product and tone of voice in its knowledge, scoped to linkedin.com /in/, with all three switches on and approved for the sales team, and a Draft intro task:

1. On a profile, the person presses Draft intro.
2. The assistant reads the profile and the pages it links to.
3. It clicks Connect. This is an outward action, so it waits for the person to confirm.
4. In LinkedIn’s dialog it writes a personal note, then stops at Send for the person to review, edit if they like, and confirm.

So the task holds twice: once at Connect, once at Send. Nothing is sent without the person’s go-ahead.