HomeTasksE-commerceCustomer support replies
Automating customer support replies
Customer support replies in E-commerce. Capability-gated, 75% repeatability, Costly-tier risk. Here is what the framework says about the shape of this work, and what supervision it would need if you ran it.
What the shape of this task tells you
With the shape of this task alone, the readiness score lands between 30 and 54 out of 100. Where it falls inside that range depends entirely on your volume: the economics half of the score is worth 45 points and it is earned at the scale you actually run. Enter your numbers to fix the score.
Repeatability 75%. A substantial share of instances deviate from the standard path. That is the part that decides this verdict: the exceptions are the work, and they are what a pilot measured on routine volume will not show you.
Judgment medium, risk Costly. Together these set both the readiness penalty and the supervision this work needs. Errors cost real time or money to undo, so the supervision line belongs in the business case as a cost, not a footnote.
The failure mode to test for
A refund, discount or delivery date promised that your policy does not actually allow
This is the one to build your sample around. Errors that survive review are the ones that look correct, so a review pass that only inspects items flagged as unusual will systematically miss it.
The oversight contract
Delegation level 2 of 4 — Draft and approve. Read production, write to a draft queue a person releases.
| Period | Sample | Step down when |
|---|---|---|
| Weeks 1–2 | 54% | Error rate under 3% for the full period |
| Weeks 3–6 | 27% | Error rate under 3% for the full period |
| Week 7 onward | 11% | Steady state — do not reduce further |
Step down on measured performance, never on the calendar date. If the error rate breaches the ceiling at any point, go back up a step.
Sample adequacy. Seeing no faults in a sample of n only rules out error rates above roughly 3 ÷ n. To have evidence for a 3% ceiling you need about 100 reviewed units a month. At a 11% steady-state rate that means a monthly output of roughly 927 units before a clean month means anything. Below that, treat a clean month as an operational check rather than proof.
Capability-gated
This verdict rests on what models can currently do reliably. It is re-scored after every frontier release and it is expected to move.
Get the full verdict for your numbers
The score above is the shape of the work. Enter headcount, salary, hours and output and the audit returns the readiness score, three costed scenarios and a supervision schedule sized to your volume.
Audit this taskCommon questions
Should customer support replies be automated?
It depends on your volume as well as the shape of the work. On shape alone this task scores between 30 and 54 out of 100, with 75% repeatability, medium judgment and Costly-tier risk. The remaining 45 points of the score come from your cost per unit and whether your manual spend covers the fixed costs, so a low-volume version of this task can score well below a high-volume one.
How much output would need to be reviewed?
At steady state, about 11% of output, after a ramp starting at 54% for the first fortnight. The error ceiling is 3% — above that, automation pauses.
What is the most likely way this goes wrong?
A refund, discount or delivery date promised that your policy does not actually allow.
Will a better model change this verdict?
This verdict rests on what models can currently do reliably. It is re-scored after every frontier release and it is expected to move.