Skip to content
Stage 01—02 · the assay

8 things people actually ask for

Nobody wrote this list. The backlog was grouped by what each ticket says, then every group was ranked by the human time it costs — how often it comes in, multiplied by how long it takes to close. Together they account for about 250 hours of handling across 218 tickets.

250h
recoverable handling time

Positions come from the ticket text alone — subject, description and the agent's closing note, run through TF-IDF and projected onto the top three principal components. The grouping was never shown to the projection, which is what makes the clumps worth anything: on this corpus a ticket's nearest neighbour in this space shares its request type 100% of the time.

What you are looking at
  1. Every dot is one support ticket. This is a whole helpdesk backlog at once — every ticket that was raised, worked and closed. Nothing has been filtered out.
  2. Nothing told it what these tickets were about. A dot's position was worked out from the words in the ticket alone — the subject line, the description, and the note the agent wrote when they closed it. No categories, no tags, no labels were used.
  3. They landed in clumps anyway. Each line joins a ticket to the ticket that is worded most like it. On this backlog, that nearest neighbour turns out to be the same kind of request every single time. The clumps are the things people actually ask for, discovered rather than declared.
  4. Bigger dots are requests that repeat. A dot grows with how crowded its neighbourhood is. Big dots sit inside a request that comes in over and over — those are the ones worth automating. Small lonely dots are one-offs, and they are left alone.
  5. Colour is the job each ticket does. When a skill gets built, tickets are used three different ways: the well-handled ones teach the steps, the ones that went wrong teach it when to stop, and a sealed set is kept back to test the result. Only the ones that taught the steps are lit right now — that is what the buttons under the panel do.
02

Ranked by what they cost

volume × median handling time

New Employee Onboarding

struck

Request to set up a new hire with required accounts, hardware, and access.

38
97h
recoverable
40 tickets146m median8 taught the steps8 taught it to stop8 sealed for testingopen the skill →

Laptop Hardware Issues

not struck

Problems with laptop hardware such as screen, keyboard, overheating, battery, or power.

61h
recoverable
38 tickets96m median8 taught the steps7 taught it to stop7 sealed for testing

VPN Connection Issues

not struck

Problems establishing or maintaining a VPN connection.

25h
recoverable
40 tickets38m median8 taught the steps8 taught it to stop8 sealed for testing

Salesforce Access Request

not struck

Requesting read‑only or other access to Salesforce.

19h
recoverable
8 tickets145m median5 taught the steps2 taught it to stop1 sealed for testing

Datadog Access Request

not struck

Requesting access to the Datadog monitoring workspace.

16h
recoverable
10 tickets98m median6 taught the steps2 taught it to stop2 sealed for testing

MFA Reset

not struck

Resetting or reconfiguring multi‑factor authentication devices or codes.

16h
recoverable
39 tickets24m median8 taught the steps8 taught it to stop7 sealed for testing

Account Unlock

not struck

Unlocking a user account that has been locked.

9h
recoverable
35 tickets16m median8 taught the steps7 taught it to stop7 sealed for testing

Miro Access Request

not struck

Requesting a license or access to Miro.

6h
recoverable
8 tickets48m median4 taught the steps3 taught it to stop1 sealed for testing

The three-way split is what stops this being a summariser. Averaging all twenty tickets in a group would produce the median procedure — which encodes the team’s bad habits as if they were policy. Instead the best-handled tickets teach the steps, the ones that went wrong teach the stop rules, and the sealed set is never seen until the trial.