Rollout, handover and writing · Rollout

Set the gates

Size a decline-letter crawl from reviewer time, gate each reason on its own sample, set the walk trigger, and decide whether any reason earns run.

estimatemediumExercise · 3 min · 54 of 64

On this page

Phase 2 drafts the letters nobody can take back. The gates have to be set before the first draft, and the VP has already named the date he wants them gone.

The situation

Phase 2 of the claims-triage pilot, from week 9: drafting the letter sent when a claim is declined, most of them after adjuster review, about 15% of 1,200 claims a day, so 180 letters. A handler writes each from one of nine templates in about six minutes; checking and signing a correct draft takes about one and a half. A wrong letter, one citing the wrong decline reason or using a phrase the legal team prohibits, is a complaint to the ombudsman, and a sent letter cannot be recalled. Decline reasons split three ways: policy exclusion 60% of letters, late notification 25%, non-disclosure 15%. A reviewer grades a draft against the legal team's rubric in about four minutes, and the two senior handlers can each give two and a half hours a day. The VP wants letters "going out without a handler" by week 14. Assume a reading near p on n graded letters wobbles by about 2 × √(p(1 − p) ÷ n), roughly ±3 points at 95% on 200, and 250 working days a year.

The question

Set the crawl, letters graded per reason and the number of days, the walk gate for each decline reason and the walk's rollback trigger, and decide whether any reason earns a run phase by week 14.

A strong answer · Rubric · Where answers go wrong

Write your answer first, then open

A strong answer

Reviewer time sets the crawl, not letter volume. Two reviewers at 2.5 hours is 300 minutes a day, 75 letters at four minutes. A gate read on 200 letters per reason carries about ±3 points, so three reasons need 600 graded letters: 8 working days. Non-disclosure, at 27 a day, reaches 200 in 7.4 days only if every letter is graded; the other two are sampled to 200. Crawl runs 8 working days.

The walk gate is per reason, as in crawl, walk, run: at least 97% of the 200 pass the rubric, so the bottom of the band is still above 94%, and none cites the wrong reason or a prohibited phrase. One such letter holds that reason in crawl whatever its percentage.

In walk the handler edits and signs every letter. At 97%, about 5 of 180 drafts a day need an edit, which is walk working. A miss is a signed letter later found wrong in a weekly regrade of 50 sent letters, and the trigger is one: that reason goes back to crawl the same day, because each miss is already a complaint.

No reason earns run, by week 14 or later. Unsigned at 97%, about 5 flawed letters a day reach claimants, around 1,350 a year; even at 99.5%, about 225. Week 14 can give the VP the saving instead: signing at 1.5 minutes instead of writing at 6 frees 180 × 4.5 = 810 minutes, 13.5 handler-hours a day.

What would reverse it: a reason whose letter is fixed text with two fields from the claim record is a template merge, not a model, and needs no gate.

Rubric

derives the crawl's length from reviewer-minutes a day, not from letter volume
sets a sample per reason from the stated band and checks that the smallest reason can supply it
gates each decline reason separately, with a zero-tolerance condition on wrong reasons and prohibited phrases
defines the walk miss as a signed wrong letter rather than an edit, and sets the trigger from what one costs
declines a run phase with the yearly count of flawed letters it would send
offers the VP the handler-hours that walk saves, with the arithmetic

Where answers go wrong

"Crawl two weeks at 500 a day, as the triage pilot did." Loses because there are 180 letters a day and the binding constraint is 300 reviewer-minutes; the numbers were copied from a deployment whose constraint was a different person's time.

"Gate at 95% overall, then autonomous at 99%." Loses twice: an overall gate lets non-disclosure, 15% of letters, pass on exclusion's numbers, and 99% unsigned is still about 450 flawed letters a year that no handler reads before they go.

"Roll back after three bad letters in a day." Loses because the triage trigger was sized to an error a handler catches the same day; a signed letter has already gone, so the trigger for an irreversible error is one.