The operating model and the catalogue · Worked automations
One catalogue entry with a constraint changed
Redesign the recruiting screen for ten times the applicants while keeping a person on every rejection, and cost the new routing and review.
designhardExercise · 4 min · 75 of 75
On this page
A worked automation is a set of decisions that held for one set of numbers. Change one number by ten and some decisions break while others must not move; the design is knowing which.
The agency's recruiting screen handles about 150 applicants a role. Of those, 30 fail a hard rule; 24 score 18 or more out of 25 and are invited straight to a 20-minute first call with the founder; 36 score 12 to 17 and go to the founder at about 3 minutes each; 60 score under 12. Every proposed decline is decided by the operations manager: 1 minute for a rule failure, 1.5 minutes for a low score. Assume staff time at $30 an hour and the founder's at $100. Scoring costs $0.0094 an applicant. By hand, the founder reads a CV in 8 minutes.
One constraint changes. A job-board listing for a junior designer draws 1,500 applicants, about 500 a day over three days, in the same shares. Every applicant is still promised an answer within three days of applying, and the founder can hold at most 40 first calls for the role.
Redesign the screen for 1,500 applicants: decide what changes in routing, gates and staffing, what stays fixed, and what the role now costs against doing it by hand.
Scale the shares: 300 rule failures, 240 at 18 or more, 360 at 12 to 17, 600 under 12. Two things break, and neither is tokens. The founder's band is 360 × 3 minutes, 18 hours in three days. And 240 automatic invites exceed 40 calls, so 200 invited applicants would be turned away later, a worse rejection than a decline.
Routing moves from thresholds to capacity. Rank the 1,200 who pass the rules by total; the top 60 go to the founder, 3 hours or $300, and she chooses up to 40 calls. Nobody is invited without her now, because calls are the scarce resource.
A person on every decline stays. Proposed declines are the other 1,140 plus 300 rule failures: 1,140 × 1.5 + 300 × 1 = 2,010 minutes, 33.5 hours, $1,005, about 11 hours a day. Two trained reviewers, the operations manager and an account manager, each take about 5.6 hours a day for three days, with authority to move anyone into the founder's queue. Declines ranked above the cut say they lost on numbers, not fit.
Redaction, the quote check, seeded cards in each reviewer's daily queue, deletion at six months and the counterfactual pairs do not change; none depends on volume. At 1,500 applicants the selection-rate ratio finally means something for one role.
The role costs 1,500 × $0.0094 = $14 of tokens, $1,005 of review and $300 of founder ranking, plus the lesson's per-role lines scaled tenfold: 30 review requests, $500, and 50 unreadable CVs, $125. About $1,940. By hand it is 200 founder hours and 50 of the operations manager's, $21,500, and impossible in three days. If the founder never picks below rank 40, cut her queue to 50.
"Decline everyone under 12 automatically and save 15 hours." It saves 600 × 1.5 minutes, $450, and brings back the decline nobody decided. At the lesson's 10% wrong and $300 a wrongful decline, that is $18,000 of expected loss to save $450, and it is the decision Article 22 is written about.
"Keep the thresholds and give the founder more time." Eighteen hours of reading is not the bottleneck; the 40 calls are. With 240 automatic invites, 200 people are invited and then refused, which the reviewer reads as a design that never checked capacity.
"Raise the invite threshold to 22 so that about 40 are invited." A threshold tuned to one pool fails on the next, and every applicant between 18 and 21 lands in the founder's band, which grows instead of shrinking. A ranking capped at the call count says what is meant.
