How to use this course
Price any task with the two sums every page repeats, what it costs a person and what its automation costs to build and run, then decide from them.
On this page
Every page in this course ends with you choosing: automate this task or leave it, on which rung, with a person where, at what budget. Each choice rests on the same two sums: what the task costs a person now, and what the automation costs to build and then to run. A proposal that cannot show both has not earned its decision.
The two sums
Sum 1 is what the task costs a person, per month.
person cost per month = tasks per month × minutes per task ÷ 60 × loaded hourly cost
Sum 2 is what the automation costs, in three lines: one paid once, two paid every month.
build, once = build hours × loaded hourly cost of whoever builds it
run, per month = model tokens + platform fees
people, per month = person minutes it still needs ÷ 60 × loaded hourly cost
The people line is the one that gets left out. Someone still reads a sample of what the automation produced, and someone handles the items it hands back, at the same rate as the work being replaced.
The decision comes from setting the two side by side:
monthly saving = Sum 1 − run − people
payback months = build ÷ monthly saving (never, if the saving is zero or less)
An automation is worth building when it pays back well inside its own life. Module 9 gives every automation a retirement date 12 months out on the day it ships, so a payback longer than 12 months is a no.
The constants, as assumptions
Every sum uses the same named assumptions, each a placeholder for your own figure, so no two pages disagree.
| Name | Assumption |
|---|---|
| loaded hourly cost | $30 for staff, $100 for the founder |
| large model | $2 per million input tokens, $8 per million output |
| small model | one-fifth of the large model's prices |
| cached prefix | one-tenth of the input price |
| tokens per word | 1.3 |
"Loaded" means everything an hour carries: salary, taxes, benefits, the desk. The founder's $100 is what that hour is worth elsewhere in the business, so a founder building on a Saturday is not free. To convert from US dollars, replace the money figures; token counts and minutes are the same everywhere.
Running example
Most pages use one company: a 14-person marketing agency, with you, the founder, at $100 an hour, and an operations manager who keeps the books, two account managers and ten people doing client work across 40 active projects, all at $30. It gets about 400 inbound leads a month from three sources (the website form, the general inbox and a partner-referral form), and an account manager spends about 10 minutes on each: read the message, find the company in the CRM, draft a reply. A page that uses a different company says so.
Both sums for one task
The lead reply, built as the one-call workflow from scripts, workflows and agents, with the founder building it.
Sum 1: 400 leads × 10 minutes = 4,000 minutes, 66.7 hours, × $30 = $2,000 a month.
Sum 2, once: 12 founder hours × $100 = $1,200. Sum 2, monthly:
- run: 1,700 tokens in and 250 out per lead, so 1,700 × $2 per million + 250 × $8 per million = $0.0054 a lead, $2.16 for 400, call it $2; plus an assumed $20 a month for the hosted workflow tool. $22.
- people: one lead in ten matches no customer and goes to a person, 40 × 6 minutes = 4 hours; at launch an account manager reads one sent reply in five, 80 × 3 minutes = 4 hours. 8 hours × $30 = $240.
Run and people together are $22 + $240 = $262 a month. The saving is $2,000 − $262 = $1,738 a month, and the payback is $1,200 ÷ $1,738 = 0.7 months, about three weeks.
Tokens are under 1% of the monthly cost; person time is 92%. The small model would save $1.73 a month. Reading one reply in twenty instead of one in five saves 3 hours, $90, about 50 times as much. The arithmetic of a task finds the same effect on the invoice run, where the review minute decides the payback.
Whether a smaller sample is safe depends on what a wrong reply costs, which is module 4's question, and for the lead reply the answer runs the other way. Sampling review and confidence thresholds prices a wrong reply at about $375, so from month 3 every reply is read before it sends: $600 a month in place of the $120 sample. With module 9's 20-minute weekly review, $43, the handler costs $305 and saves $1,695 a month at launch, about $785 and $1,215 from month 3, and still pays back in about a month.
The sum usually done instead
The sum done in most meetings is tokens against salary: "$2 a month of model calls replaces $2,000 of someone's time". It omits the build and the people line, which is 120 times the tokens. On the lead reply the corrected sum still says yes; on a smaller task it says no.
The operations manager spends 3 hours a month assembling the board report: $90. Automating it takes 10 founder hours, $1,000, on the platform already paid for, and someone still checks the result for an hour, $30. The saving is $60 a month and the payback is $1,000 ÷ $60 = 17 months, past the retirement date. It stays manual.
The opposite mistake is comparing against a whole salary. The lead reply is 66.7 hours a month, about 0.4 of a full-time job; Sum 1 prices those hours.
Lessons decide; exercises make you decide
Each lesson teaches one decision and its number, and ends with ## In practice, what to tell the vendor or the team and the mistake that costs money, and ## Check yourself, questions that force a decision or an estimate. Answer before reading the blockquotes.
Each module then has exercises, in five kinds:
| kind | What you produce |
|---|---|
case | a decision for a situation, and the evidence that would reverse it |
estimate | a number, with the arithmetic shown from stated assumptions |
design | a mechanism, such as a role, a graph or a gate, and its defence |
critique | what is wrong with a given brief or result, and what to do instead |
code | one solve() against a stated contract, proved by a test runner |
Read the situation and the question, then stop and write your answer: the decision, the sum behind it, and what would change your mind. Only then open the strong answer and mark yourself against the rubric, the things a reviewer scores. The weak answers listed last often include yours. The five code exercises are in module 6.
The route through the course
Module 1 picks the task, module 2 the rung, modules 3 to 7 shape and connect the build, and modules 8 and 9 run it: evals, cost, security, an owner and a retirement date.
A parallel ladder, ordered by how much of the loop lives in code, is five eras, one ladder in the developer's course, which teaches the mechanisms this one applies: the tool-call loop and its stop conditions, evals, the MCP protocol, harness design. Queueing arithmetic comes from Little's law. If you already run automations, test out of modules 1 and 2 by their exercises, but do not skip 4 and 8, where automations that worked in a demo lose money.
In practice
When a vendor, a contractor or your own team proposes an automation, ask for both sums on one page before discussing tools: the task's monthly person-hours, then the build hours, the monthly tokens and fees, and the person minutes it will still need. A proposal with no people line has priced less than a tenth of the monthly cost above.
The mistake that costs money is pricing the tokens and not the people. Review and exception minutes are most of what an automation costs each month, and they grow when it gets things wrong.
Check yourself
1. A contractor quotes 30 hours at $80 to automate overdue-invoice reminders: 150 a month, 4 minutes each by the operations manager. Tokens are about $1 a month on the platform you already pay for, and the operations manager would check one reminder in five at 2 minutes. Build it?
Sum 1: 150 × 4 = 600 minutes, 10 hours × $30 = $300 a month. Build: 30 × $80 = $2,400. Run $1; people 30 checks × 2 minutes = 1 hour, $30. Saving $269; payback $2,400 ÷ $269 = 8.9 months, inside the 12-month life with little room: a 40% overrun makes it 12.5. Build it only at a fixed price.
2. At launch the lead workflow costs $262 a month to run and staff. Which one change cuts it most: the small model for the draft, caching the 1,200-token prefix, or reading one reply in twenty instead of one in five?
The sample rate. The small model saves $2.16 − $0.43 = $1.73. Caching saves 480,000 tokens a month × ($2 − $0.20) per million = $0.86. Reading 20 replies instead of 80 at 3 minutes saves 3 hours, $90. Whether that is safe turns on what a wrong reply costs; at module 4's $375 it is not, and from month 3 every reply is read before it sends.