People

Review cycles that finish, without a worker forming an opinion

The administration of a review cycle is enormous and the judgment inside it is entirely human. A worker takes the first part and touches none of the second.

What the worker does

An AI worker in Polaris runs the administration of a performance review cycle: creating a task per participant, booking the conversations in Google Calendar, chasing managers and reviewers who have not submitted, and assembling each person's packet from what was submitted. It writes no assessment, forms no view on anyone's performance, and reads nothing beyond the material submitted for the cycle.

Connections: Google Calendar · Gmail · Docs
Assembles: One packet per participant
Assesses: Nobody. Ever

Cycles fail on completion rates, not on the form design

Every review cycle has the same shape: a self-assessment, some peer input, a manager write-up, a conversation, an outcome. Every review cycle also has the same failure: on the deadline, sixty percent of it is in, and the people chasing the other forty percent are the same people who owe write-ups themselves.

So the cycle slips three weeks, the feedback goes stale, and the conversations happen with half the input missing.

Chasing is the whole job. A worker can do it every day, to everyone, without anyone taking it personally, and can tell you on day three that four managers have not started.

Running the cycle

1
Set up the cycle as a workstream with a task per person
One task per participant, owned by their manager, in a workstream for the cycle. Everything for that person, including the packet, ends up on their task.
2
Restrict what the worker can read, in writing
State in the SKILL.md that the worker reads only the documents submitted for this cycle. It should not be reading a person's chat history or their commits to form a picture of them, and the brief is where you make that explicit.
3
Let it chase on a schedule
Daily reminders to whoever has an outstanding item, escalating to the manager's manager on the date you set, and a running completion figure posted to the cycle task.
4
Have it assemble the packet
Self-assessment, peer input and the manager's write-up collected into one document per person, in a consistent order, with anything missing named at the top.
5
Humans hold the conversations and close the tasks
Ratings, feedback, pay and progression are decided and delivered by people. The manager closes the task when the conversation has happened.

The line, drawn hard

Performance is the clearest case in this whole cluster where an AI worker must not have an opinion.

The worker does

Create, assign and chase the tasks
Book the conversations
Report completion rates and who is behind
Assemble each packet in a consistent order
Name what is missing from a packet

The worker does not

Summarize or paraphrase someone's feedback about a colleague
Rate, rank or calibrate
Suggest an outcome
Read work systems to build evidence about a person
Send anything to the person being reviewed

What to measure about your own cycle

Completion on the deadline, not two weeks after
The number that tells you whether the cycle is working. The worker reports it daily during the window.
Days from deadline to last conversation held
This is the figure that quietly stretches every year.
Packets missing an input
A packet delivered with a gap named at the top is honest. A packet that hides the gap is not.

Questions people ask

+Can the worker draft a manager's review for them?

No. A review is a manager's own assessment of a person they are responsible for, and outsourcing the words to a machine hollows out the one conversation of the year that is supposed to be personal. The worker chases the manager. The manager writes.

+Is the review material private?

It lives on tasks in your workspace with the access controls you set, and Polaris stores workspace data in Postgres with row-level security. Treat access design as your decision: put the cycle in a workstream visible to the people who should see it, and no wider.

+What does the worker actually cost for a cycle?

Hours are computed from the observable effort in each job and billed at roughly $2 per human-equivalent hour, itemized on the work log. Chasing and assembly are small repeated jobs, and any line can be challenged from the log.

+Can it run a lightweight cycle for a ten-person company?

That is the case where it helps most, because a ten-person company has no HR function and the founder is both the chaser and one of the people being chased. A worker chasing on a schedule removes the awkwardness that stalls small-company cycles.

HR work an AI worker can prepare, and where it must stop

Scheduling, onboarding logistics, policy drafts and review-cycle admin, prepared by a worker. Every decision about a person stays with a person.

Policy documents that stay current and show their changes

A policy nobody has updated in three years is worse than no policy. A worker keeps the drafts moving and the versions visible. A qualified adviser signs them off.

OKRs that stay visible after the offsite

OKRs do not fail at the writing. They fail in week three, when nobody has looked at them since the offsite.

Dependencies as dated items, not as things people said in a meeting

Two teams agree something in a call. Neither writes it down. Three weeks later each is waiting for the other. A worker makes the promise a tracked item.

Hire an AI project coordinator

The unglamorous half of running projects, done every week without anyone having to be the person who nags.

Human-in-the-loop

The system cannot complete the loop without a person, by design.

Work log

The record that makes an unwatched run reviewable afterwards.

Connect Google Calendar to Polaris

Scheduling is the one thing an assistant is asked for first, and it is the one connection still waiting on Google sign-in.