Retiring an automation
Decide when an automation should stop, against the task's own monthly sum, and switch it off without losing a single item in flight.
On this page
An automation should be retired on its sum, not on a feeling. Three things end one: the task changes under it, a person becomes cheaper again, or its monthly cost crosses the line that justified it, the task's own Sum 1. Each arrives as a run of small, reasonable decisions, which is why the retirement date exists to force the question, and why switching off needs an order.
Nine months of reasonable changes
The lead handler launched at $305 a month: the $262 priced in how to use this course plus the owner's 20-minute review. Follow it through a year in which the agency changes how it sells, and by month 9 it costs this:
| Line | Launch | Month 9 |
|---|---|---|
| Model calls and platform | $22 | $22 |
| Leads handed to a person | 40 × 6 min, $120 | 100 × 8 min, $400 |
| Reading replies | 1 in 5, 80 × 3 min, $120 | every draft, 300 × 3 min, $450 |
| Rewriting drafts | none | 100 × 10 min, $500 |
| Owner's review | 20 min a week, $43 | 40 min a week, $87 |
| Founder's prompt fixes | none | 6 h × $100, $600 |
| Total | $305 | $2,059 |
Every change was sensible when it was made. In month 3 a wrong reply cost a deal, and the owner moved from sampling to reading every draft before it sends, as sampling review and confidence thresholds advises when a wrong reply costs $375 on average. In month 6 the agency moved to monthly retainers, and leads began asking for a monthly price, which the handler may never quote: a quarter went to a person instead of a tenth, at 8 minutes each because the account manager read the failed draft first. The reviewer rewrote one draft in three, the founder spent six hours a month rewriting the prompt, and the review stretched to 40 minutes.
Each review looked at that week's change; none added up the column. The budget line covered money, which never moved; the people lines, which carried the total to $2,059, had no alarm.
The line is the task's own sum
The line is Sum 1, computed today: 400 leads × 10 minutes = 66.7 hours × $30 = $2,000 a month, or $5 a lead. It moves with the task: at 300 leads it is $1,500.
The automation's side is everything it costs this month: model calls, platform, every person-minute it causes, the owner's review and any fixes. The $1,200 build is not on it: it had to pay back inside the year at launch, and running longer does not recover it. A fix the automation now needs is on it, as a new build that must pay back before the retirement date or justify a new one.
Put an alarm on the people line well below the line. At half of it, $1,000, the saving still funds a fix that pays back within a month or two. The lead handler passed $1,000 around month 6, three months before anyone asked.
Three ways an automation stops paying
The task changes. The handler was built for a reply within 5 minutes that never quotes a price; retainer leads need a scoping call and a price range. The signature is hand-offs and rewrites climbing together while the eval set, built from launch traffic, still passes, because it measures the old task. No prompt edit fixes a changed procedure: six founder hours a month bought no fall in the rewrite rate.
A person is cheaper again. Sum 1 falls when volume or minutes fall, but the automation has a floor that does not: the $20 platform and the $43 review, $63 a month. At launch settings a lead costs about $0.61 (tokens $0.0054 plus 1.2 person-minutes: a tenth of leads at 6 minutes, a fifth sampled at 3) against $5 by hand, so the handler pays down to $63 ÷ ($5 − $0.61) ≈ 14 leads a month. Volume alone rarely retires it; volume plus a needed fix does. At 60 leads it saves 60 × $4.39 − $63 ≈ $200 a month, so a $1,000 connector rebuild takes five months to pay back, past a date three months away.
The cost crosses the line. Usually through people rather than tokens, and usually as a symptom of one of the other two. When the cause is one part of the traffic, retire that part: one source or category handed back to people, the rest kept, both halves re-priced. When the cause is the task, retire all of it.
The decision in month 9
The cause is the task, so there is no cheap fix to price: a fix means a new procedure and a new build. By hand the leads cost $2,000, $59 a month less than the handler. The account managers' share rises from the 48 hours a month they spend on hand-offs, reading, rewriting and review today to 66.7, about 19 hours more, while the founder gets six back.
Retire it now, three months before its date. A month of handling retainer leads by hand is also how the new procedure gets written, and any new automation is a new decision with its own two sums.
Switching it off without losing the queue
The tempting switch-off is deleting the workflow at 17:00 on a Friday. At that moment the handler holds nine leads: six drafts awaiting review and three hand-offs open. Deleting it loses their state and the record of which were answered, and the form provider's retries hit an endpoint that no longer exists. If an unanswered lead is lost one time in four, the odds module 4 used for a wrong reply, nine leads at $1,500 put about $3,400 at risk.
The order instead:
- Name who takes the work, with the hours. About 19 more a month for the account managers, said before the date.
- Stop intake at a set hour. Point the three sources at the account managers' queue. Change the trigger; delete nothing.
- Drain what is in flight. The six drafts finish through the reviewer; the three hand-offs stay in their queue, clocks running.
- Hold through the retry window. Webhooks arrive at least once, so duplicates and late retries keep coming, and only an idempotency key stops an answered lead being answered twice. Keep the endpoint and its keys for 72 hours, forwarding anything new to people.
- Archive, then revoke. The procedure, prompt versions, eval set and review log go in one folder, where any rebuild starts. Revoke the credentials; cancel the plan at the end of its billing period.
- Log it. The catalogue row gets the date, the reason and the sums at retirement.
In practice
Tell the team the decision with its sum: "The lead handler costs $2,059 a month to save a $2,000 task. It retires on Monday at 09:00; from then the three sources land in your queue, about 19 hours a month more between you, and drafts already in review finish as usual." Tell any vendor running it to keep the endpoint up 72 hours after intake stops.
The mistake that costs money is keeping an automation because its model bill is small while its people lines have passed the task's own sum. The lead handler's $22 never moved while the people lines, $2,037 of its $2,059, went unwatched.
Check yourself
1. Paid ads stop and leads fall to 150 a month, with the handler at launch settings. Is a person cheaper again?
No. Sum 1 is 150 × $5 = $750; the handler costs about 150 × $0.61 + $63 ≈ $154, so it still saves about $596 a month. At these settings a person is cheaper only below about 14 leads a month.
2. At those 150 leads the form provider retires its webhook, and the connector needs a 10-founder-hour rebuild. The retirement date is two months away. Rebuild or retire?
Rebuild. $1,000 ÷ $596 ≈ 1.7 months, inside the two left. As a new build it comes with new sums and a moved date; over a fresh year it returns about $596 × 12 − $1,000 ≈ $6,150.
3. The owner plans to retire the handler by deleting the workflow at 17:00 on Friday, with six drafts in review and three hand-offs open. What do you change?
The order. Nine leads are in flight, about $3,400 at risk. Stop intake at a set hour, let the six drafts finish through the reviewer, leave the three hand-offs in their queue, and keep the endpoint and keys for 72 hours, forwarding late retries to people. Delete nothing until the queue is empty.