Role
Hire an AI research analyst
Assign the question on Tuesday afternoon and read the brief before your Wednesday call, with every claim carrying the URL it came from.
What this worker is
An AI research analyst in Polaris is a worker you hire in chat to answer open questions with sourced written briefs. Given a question and acceptance criteria, a cloud machine runs live web searches, drafts the brief in a versioned doc, attaches a typeset PDF when the work is board-facing, lists its sources as plain URLs, and posts the result as a comment for a person to close and rate.
- Typical output
- Sourced brief, doc plus PDF
- Core connection
- Web search
- Illustrative task
- 3.4 human-hours, $6.80
What this worker is great at
Turning a vague question into a written answer somebody can act on. You assign “what are the three cheapest ways to run scheduled jobs on Cloudflare, and what breaks at each price point” and you get back a brief that names the options, quotes the prices it found, and links each number to the page it came from.
The brief that comes back has a shape, because the shape lives in the skill file. Claims carry sources. Conflicting sources are flagged rather than averaged into mush. There is a short section on what the analyst could not verify, which is the part most human briefs quietly omit and the part that stops you presenting a guess as a finding.
Written work is drafted in Docs by default, so it lands as a versioned page with comments rather than a wall of chat text. When the brief is going in front of a board or a client, it also arrives as a typeset PDF.
What to connect it to
Connections come from the fixed catalog, are authorised once for the whole org, and are stored server-side.
Web search
The one this role cannot work without. It is what turns an assertion into a sourced claim, and it is the only connection in the catalog that needs no credential.
Google Drive
Give it Drive when the research has to sit alongside existing material your team already keeps there.
Notion
Useful when the prior art lives in a Notion wiki and you want the analyst working from the same page as everyone else.
Slack
Worth adding if the questions arrive as messages: the Inbox turns those signals into prefilled task suggestions you approve with one click.
A realistic first week
- 1
Monday: a competitor pricing brief
Four named competitors, current published prices, what each price includes, and the URL for every figure. Acceptance criteria: one row per competitor, no unsourced number.
- 2
Tuesday: the question your board asked
Whatever came up in the last meeting that nobody had time to chase. Ask for a two-page answer with a stated confidence level per section.
- 3
Thursday: a market primer for a new hire
The regulations, acronyms and players in a market your team is entering, written for somebody who starts on Monday.
- 4
Friday: a follow-up on Monday's brief
Comment on the task with your three sharpest objections. The latest human comment becomes the brief for the next run, so the second pass answers you rather than restating the first.
Where the human stays in charge
The analyst prepares
- The question decomposed into answerable parts
- Findings with the URL behind each one
- A note where two sources disagree
- An explicit list of what it could not verify
- The brief as a doc, and a PDF when it needs one
You decide
- Whether the finding is good enough to act on
- What goes in front of a client or a board
- Which recommendation the company actually takes
- Closing the task and rating the delivery
What a brief costs
Hours are human-equivalent, estimated from observable effort by the published formula and logged job by job. Illustrative example, computed from that formula rather than measured on a customer.
| Observed effort | Human-equivalent minutes |
|---|---|
| Picking up the task | 15 |
| 5 live web searches at 12 min each | 60 |
| 6,300 characters of finished prose at 90 chars/min | 70 |
| 4 acceptance criteria ticked at 8 min each | 32 |
| 1 progress comment, 1 doc, 1 PDF | 25 |
| Total: 3.4 human-hours at $2 | $6.80 |
Work this role picks up
Research memos waiting for you in the morning
The questions worth researching are the ones you never have a free afternoon for.
Competitor monitoring that actually happens every week
Everyone agrees competitor tracking matters and nobody has done it since the last time a deal was lost over it.
Twelve interviews recorded, two of them read
Themes with participant counts and verbatim quotes, plus a flag wherever the evidence is thin.
Account briefs waiting for you before the first call
Ten minutes of research changes a first call. Nobody has ten minutes before a first call.
Questions people ask
+How does an AI research analyst cite its sources?
Every factual claim carries the URL it came from, listed as plain links in the delivery comment and inside the doc. The runtime records each web search in the work log, so the trail of what was searched and what was read sits next to the finished brief.
+Can it research private or internal material?
It works from the open web plus whatever you connect it to and whatever you put in the task. It cannot reach paywalled databases, gated analyst reports or anything requiring a personal login that is not part of the connection catalog.
+How long does a research brief take to come back?
A machine session runs in minutes rather than the hours the same brief takes a person, and progress is posted on the task while it works. Long questions are better split across several tasks than pushed into one session.
+What stops the analyst from inventing a statistic?
The skill file requires a source for every claim and an explicit note where verification failed, and the work log shows how many live searches were actually run. A brief with no searches logged and confident numbers in it is a brief worth challenging.
+Can I edit how the analyst writes briefs?
Yes. The role's capabilities live in a SKILL.md file you can open and edit, so a house rule such as always including a confidence level or never using a source older than eighteen months becomes permanent after you write it once.
Related
Hire an AI competitive analyst
Assign it every Monday and you get a written record of what your market did last week, with a URL behind every line of it.
Hire an AI data analyst
It settles what active user means, writes the query that matches the definition, and turns the numbers you hand it into a paragraph an executive can read.
Research memos waiting for you in the morning
The questions worth researching are the ones you never have a free afternoon for.
Competitor monitoring that actually happens every week
Everyone agrees competitor tracking matters and nobody has done it since the last time a deal was lost over it.
Twelve interviews recorded, two of them read
Themes with participant counts and verbatim quotes, plus a flag wherever the evidence is thin.
Web search in Polaris
Nothing to authorize, nothing to store, and every query a worker runs appears in the activity feed while it works.
Work log
The record that makes an unwatched run reviewable afterwards.
How long an agent session can run, and what happens when it ends
Every agent runtime has bounds. The useful thing a vendor can do is tell you what they are.