Outcome record
Service inbox triage, timed on a 20-ticket sample
In March 2026, twelve learners in an open Document and Inbox Automation class in Fusionopolis completed the same anonymised 20-ticket pack twice: once with their current method, once with the class prompt file and checklist. Median handling time moved from 6.8 minutes to 2.9 minutes per ticket. Mislabel rate moved from 9.0% to 3.5% against a key written before the day.
Course details All outcome records
What the sample looked like
The pack was 20 incoming messages written to resemble a small Singapore service desk: facilities, access cards, invoice copies, and a handful of complaints. Names, company titles, account numbers and site codes were replaced with tokens before class. Attachments were flattened to a one-line description so that file-type variance would not dominate the clock.
Eight labels were in the key: access, billing, facilities, complaint, identity-check, out-of-scope, duplicate, and escalate. Three tickets were written as true escalations (a legal threat, a payment dispute, a missing identity). Learners who work a live desk will recognise the mix. It remains a demo pack, and the record says so.
Protocol on the day
The morning pass used whatever method each learner already had: search, folders, memory, or an assistant without a shared prompt file. The clock started when ticket 1 was open and stopped when the learner declared all 20 labelled, with a draft reply on each in-scope item. A second person, sitting with the key, counted wrong labels and missed escalations. Setup, Wi-Fi, and the instructor’s briefing were off the clock.
After the workshop rebuild, learners received ten held-out tickets of the same class (not the morning twenty) plus a second timing on a fresh shuffle of twenty for the published handling-time figure. The mislabel rate in the table is on the twenty-ticket after-pass, scored on the original key. We did not edit the key after seeing the afternoon’s mistakes.
The numbers
| Measure | Baseline | After |
|---|---|---|
| Median handling time per ticket | 6.8 min | 2.9 min (−57%) |
| Mislabel rate (wrong label or missed escalate) | 9.0% | 3.5% |
| Learners kept | 12 | |
| Date | March 2026, Fusionopolis open class | |
The −57% figure is the change between the two recorded medians. Two learners improved by less than 20% on time; one of those two also had the smallest drop in mislabels. We keep those rows in. Dropping slow improvers would make the median look tidier than the room was.
What learners built
The prompt file asked the assistant for a label, a one-sentence rationale, a draft reply, and a yes/no on the escalation rule. Blanks in the reply templates were marked so that the assistant was not allowed to invent a policy, a refund amount, or an appointment slot. The checklist forced a human read of every escalate=yes and every complaint label before send.
The measurement sheet recorded minutes and error ticks at item level, so a learner could see whether time was lost on long complaints or on easy duplicates. That sheet went home with them. Aicoursesforsg keeps only the classroom summary in this record: medians, rates, sample size, date.
Limits that belong next to the −57%
Twenty tickets is a short queue. A live desk with attachments, internal CCs, and a second system for asset numbers will run slower. The room was quiet. The key was known to the second marker and unknown to the learners, which is fair for a test and unlike a Monday when the “right” label is a negotiation with another team.
Novelty is unmeasured. We do not have a 90-day re-time on this cohort. We also did not measure customer satisfaction, only handling time and mislabels on a key. If your organisation cares about tone more than speed, this record is the wrong exhibit; look at the reporting case, or book a private sample with a tone rubric agreed in advance.
Measure this on your own tickets
A private cohort can use a stripped extract of your shared inbox if twenty comparable tickets can be produced. If they cannot, this March 2026 record is still the honest public number: classroom, demo data, twelve people. The related programme is Document and Inbox Automation.