Tasks
Web search and research
AI staff can search the public web, read a page, and run multi-round research that comes back as a report with numbered sources.
Three tools cover the open web, from a single lookup to a full research pass. They belong to AI staff: give a staff member the web tools and it can find current information on its own, cite where it came from, and hand you a report.
Where it runs: in AI staff runs, for staff given the web tools.
What a staff member can do
| It can | What that looks like |
|---|---|
| Search the web | Ranked results — title, URL, snippet — to find out what exists |
| Read a page | One URL fetched and reduced to clean, readable text |
| Research a question | Several rounds of searching and reading, synthesised into a short report citing its sources as [1], [2] |
Search finds URLs; reading turns one into text; research is the loop of both, run for you. For an open question — “what are the current rules on X”, “who else is doing Y” — research is the tool to ask for, because the output is a sourced answer rather than a list of links.
Reading the report
Deep research returns a concise report with numbered citations. Treat those numbers as the point of the exercise: they are what separates a researched answer from a confident one. Open the sources before acting on anything that matters.
Rounds are bounded — this is a few passes over a handful of sources, not an exhaustive survey. For a question that deserves twenty sources, ask for the research in parts.
What it will not fetch
Only public http(s) pages. Requests to private or internal network addresses
are refused outright, so a staff member cannot be talked into reaching into your
network by being handed a clever URL. Pages that need a login or that only render
under JavaScript need the sandbox browser instead — see
Data, code and calculations.
Availability
Web search depends on a search provider being configured for your deployment. Where it is not, staff can still read a URL you give them, but cannot discover one.
The assistant in a normal task does not browse the web. If you want a question answered from current public sources, run it through an AI staff member with web tools.