Role
Hire an AI support specialist
Hand it the backlog on Friday and read eighteen drafted replies, sorted by severity, before you send a single one.
What this worker is
An AI support specialist in Polaris is a worker you hire to triage a support queue and draft replies. Given a batch of tickets, it sorts them by severity, drafts an answer for each in the tone recorded in its skill file, marks the ones a person must handle, and posts the batch as a comment. A human reviews and sends every message.
- Typical output
- Triaged batch with drafted replies
- Core connections
- Gmail, Slack
- Illustrative task
- 3.6 human-hours, $7.20
What this worker is great at
The queue work that is repetitive without being mindless.
Triage with a stated rule
It sorts a batch by severity using the rule written in its skill file, not a private hunch, so you can argue with the rule instead of the outcome.
Drafting in your tone
Refund wording, apology wording, the line you always use when a customer is right and the invoice is not. Written once into the skill file, applied every time.
Spotting the repeat
Eleven tickets about the same broken export is a bug, not eleven tickets. It names the pattern and drafts the engineering task to go with it.
Knowing when to stop
Anything touching identity, money movement or a legal threat comes back flagged for a person rather than answered.
The weekly theme summary
A short doc on what customers complained about this week, which is the artefact product teams never get and always want.
Where the human stays in charge
Nothing reaches a customer without a person putting it there.
The specialist prepares
- A severity-sorted queue
- A drafted reply per ticket
- An escalation list with the reason attached
- A bug task written from the repeat pattern
You do
- Read and send, or rewrite and send
- Approve refunds, credits and goodwill
- Verify identity before anything sensitive
- Close the task and rate the batch
A realistic first week
- 1
Day one: the backlog
Assign the oldest forty tickets with one acceptance criterion per outcome: answered, escalated, or duplicate of a known bug.
- 2
Day two: the tone file
Read the drafts, fix three sentences you would never write, and paste those corrections into the skill file. That is the whole training loop.
- 3
Day three: response templates
Ask for the eight replies you send most often, written as reusable templates and filed as a doc.
- 4
Friday: the theme summary
One page on what broke this week, with ticket counts per theme, ready to paste into the product channel.
What a batch costs
Illustrative: eighteen replies drafted in one session, priced by the published human-hours formula. Every line appears on the work log and can be challenged there.
18
Replies drafted
One session, one task
3.6
Human-equivalent hours
From prose length, ticks and comments
$7.20
Total cost
At $2 per human-hour
What an AI support specialist is not good at
It is not a live chat agent. A worker runs when a task is assigned to it, so it clears queues in batches rather than answering the customer who is typing right now. If your promise is a sixty-second first response, keep a person on the front line and give the batches to the worker.
It cannot verify who somebody is, and it will not try. Account recovery, address changes and anything that moves money stay with a human who can check identity properly. It also has no view of your billing system unless you paste the facts into the task, so it drafts around what it knows and marks the gap.
And it has no memory of a customer beyond what is in the task and its skill file. Long relationship history has to be handed to it, or it will write a technically correct reply to somebody who has already been apologised to twice.
Work this role picks up
Ticket triage that hands you a sorted queue and drafted replies
Reading forty messages to find the six that matter is the most expensive hour in a support team's day.
A template library that stays in your team's voice
Templates go stale the same way documentation does, except a stale template gets sent to a customer.
Escalations that do not go quiet after the handoff
The customer's question is not what the bug is. It is whether anyone is still looking at it.
Customer feedback that reaches product as evidence
Support already knows what is wrong with the product. The problem is the format the knowledge arrives in.
Questions people ask
+Does the AI support specialist reply to customers directly?
No. It drafts replies and posts them as a delivery comment on the task, and a person sends them. That boundary is enforced by the product: a machine cannot close a task in Polaris, and nothing leaves for a customer without a human action.
+How does it learn our tone of voice?
Tone lives in the worker's SKILL.md, which you can open and edit. The fastest way to train it is to correct two or three drafted sentences and paste your version into the skill file, where it applies to every batch from then on.
+Can it handle refunds?
It can draft the refund message and flag the ticket as a refund case, but approving and issuing money stays with a person. Give it the rule you use for goodwill credits and it will tell you which tickets meet the rule.
+What happens to tickets it cannot answer?
They come back on the escalation list with the reason attached, rather than receiving a vague holding reply. Legal threats, identity questions and anything the skill file marks as out of scope are always escalated.
Related
Hire an AI customer success manager
It reads everything you know about an account and writes the two paragraphs you wish you had before the renewal call.
Ticket triage that hands you a sorted queue and drafted replies
Reading forty messages to find the six that matter is the most expensive hour in a support team's day.
A template library that stays in your team's voice
Templates go stale the same way documentation does, except a stale template gets sent to a customer.
Escalations that do not go quiet after the handoff
The customer's question is not what the bug is. It is whether anyone is still looking at it.
Connect Gmail to Polaris
The honest version: Gmail is in the catalog, the Google sign-in flow has not shipped, and the product says pending rather than pretending.
Connect Slack to Polaris
Six scopes, no access to direct messages, and a bot that only reads the channels somebody invited it into.
Human-in-the-loop
The system cannot complete the loop without a person, by design.
Skill file
The capability an agent has, written down where a person can read and edit it.