Executive · Research
Research memos waiting for you in the morning
The questions worth researching are the ones you never have a free afternoon for.
What the worker does
Strategic research in Polaris means assigning a question to an AI worker with the web search connection. A cloud machine wakes for the task, runs a live tool loop across the open web, ticks the acceptance criteria you set, and delivers a memo as a comment with the file attached and every claim linked. The work continues after you close your laptop, and a person decides what to do with the answer.
- Connection
- web search
- Runs
- On a cloud machine, after you log off
- Delivered as
- A sourced memo file on the task
The questions that never get answered
Should we enter this adjacent market. What are the three companies doing that we keep hearing about in deals. What does regulation in that country actually require. How do other companies in this category price.
Each of these is answerable with several hours of careful public reading. Each of them loses every day to something urgent, which is why founders make these decisions on impressions rather than on reading.
This is the strongest case for a cloud agent rather than one on your laptop: the work happens during the hours you were never going to spend on it anyway.
A recorded research delivery
The competitor-pricing memo in the Polaris delivery demo. A real machine session with live web research, sped up rather than staged.
~7 min
Machine time, end to end
$2.80
Billed for the delivery
~1h24m
The human clock it replaced
About $70 at $50 an hour
How to write a research task that comes back useful
Ask a decidable question
How three named competitors price their team tier produces evidence. What is the future of our market produces an essay.
Name the sources you trust and distrust
In the skill file, once. Whether vendor blogs count, whether you want primary regulatory text, what to do with a number that appears only in a press release.
Set acceptance criteria
Every claim linked, contradictions between sources reported rather than resolved silently, an explicit list of what could not be found.
Ask for the argument against
A memo that only supports the direction you already lean is worse than no memo.
Questions people ask
+How do I know the memo is not invented?
Every claim carries a link, and the acceptance criteria include an explicit list of what the worker could not verify. Reading the sources for the two claims that matter most is a five-minute check, and it is the right check to keep doing.
+Can it read documents we already have?
Yes, if the material is in Docs in Polaris or in a connected Google Drive. A memo that combines public research with your own numbers is usually the more useful one.
+How long can a single research task run?
Sessions are clamped between five minutes and eight hours, so a very large question is best split into several tasks with narrower scopes. That also produces better memos.
+What does a research memo cost?
About $2 per human-equivalent hour, computed from observable effort including how many searches ran and how much finished prose was produced. The recorded demo memo was billed at $2.80, and every job is itemised on the work log.
Related
Running the company with an AI teammate on the reporting
The information you need to run the company exists. Assembling it every week is what nobody has time 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.
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.
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.
Web search in Polaris
Nothing to authorize, nothing to store, and every query a worker runs appears in the activity feed while it works.
Queueing work at night and reading it in the morning
The realistic version of overnight work, including the part where there is no scheduler.
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.
Human-equivalent hours
A billing unit denominated in the work replaced, not the compute consumed.