The scoring frame
Score candidate tasks 1 to 5 on frequency, cost, tolerance for error and input structure, multiply, and defend the order against an hours-first list.
On this page
The two sums say whether one task pays for its automation, not which of twelve candidates to build first. Ranking by hours saved is the usual answer and the wrong one, because it tends to put the most expensive mistakes at the top. Four factors, each scored 1 to 5 and multiplied, rank a list in ten minutes and survive the argument afterwards.
Four factors and their bands
Frequency is how often the task runs. Cost is the person-hours it takes each week, the first of the two sums before it becomes dollars. Tolerance for error is what one mistake costs once it has reached a client or the books. Input structure is how much of what the task reads arrives as fields rather than prose. Write the bands down before anyone scores, so two people scoring one task get one number. A band includes its lower bound.
| Score | Frequency, runs a week | Cost, hours a week | One error costs | Input |
|---|---|---|---|---|
| 5 | 100 or more | 8 or more | under $25 | fields from a system: a feed, an export |
| 4 | 30 to 99 | 4 to 8 | $25 to $249 | a form with fixed fields |
| 3 | 10 to 29 | 2 to 4 | $250 to $999 | documents with a usual layout, or a form plus email |
| 2 | 2 to 9 | 1 to 2 | $1,000 to $4,999 | free text with a predictable shape |
| 1 | under 2 | under 1 | $5,000 or more | free text with no shape |
| 0 | not permitted unattended | not written down anywhere |
Frequency and cost look like one number counted twice, since cost is frequency × minutes. They are not. Cost is what the automation can save; frequency is how often it is tested, and how many runs share the build. A task run 35 times a week shows a 2% error rate within a month; a month-end job shows it every four years or so.
Tolerance is scored from the cost of one error, not its likelihood, which belongs to a build that does not exist yet. Input is scored from what arrives today: a free-text email scores 2 at most even if a form could replace it, and the form is a separate, cheaper project.
Why the four are multiplied
Summing feels fairer and is the mistake. A task scored 5, 5, 1, 4 sums to 15 of a possible 20, 75%, and looks strong. Its 1 is an error that costs $5,000 or more. Multiplied it is 100 of a possible 625, 16%: the 1 has capped the task at a fifth of what its other three factors would allow, which is what a $5,000 error should do.
The 0 is the case summing cannot express. Some tasks may not run unattended at all: a contract forbids it, a regulator does, or the bank will not let money leave without a named approver. Summed, a 0 costs five points and three good factors outvote it. Multiplied, one zero kills the candidate however good the rest looks. RICE multiplies reach, impact and confidence for the same reason: a weak factor should drag the whole score, not a quarter of it.
Six tasks side by side
The agency from the running example keeps a list of twelve tasks people have asked to hand to software. Six are scored here; the other six, the invoice run among them, are the module's first exercise.
| Task | /week | min | h/week | One error costs | Input |
|---|---|---|---|---|---|
| Chase overdue invoices | 35 | 4 | 2.3 | $100, a reminder to a client who has paid | receivables list from the accounting system |
| Meeting notes and actions | 30 | 15 | 7.5 | $150, a missed action | the call transcript |
| Apply ad budget changes | 120 | 5 | 10 | $10,000, a misplaced zero | a sheet with fixed columns |
| Status replies to clients | 40 | 8 | 5.3 | $600, a wrong date promised | a client's free-text email; the answer is in the project tool |
| Month-end close pack | 0.23 | 360 | 1.4 | $500, a second pass with the accountant | accounting exports |
| Pay supplier bills | 40 | 5 | 3.3 | not permitted: the bank wants a named approver | bills with parsed fields |
Three rankings that disagree
By hours, the list starts with ad budget changes: 10 hours a week, and the task where one error spends $10,000 of a client's money overnight.
Summed, the chase leads on 16 with ad budgets a point behind. Supplier bills, which may not run unattended at all, sit on 11, level with the month-end pack and still on the list.
Multiplied, the chase leads on 240, nearly double meeting notes on 128. Ad budgets fall to third: the 1 on tolerance cost them the place their hours bought. Supplier bills score 0 and leave the list. The month-end pack scores 30 despite perfect input, because 12 runs a year are too few to test on or to spread a build over.
A 0 says re-scope, not abandon. "Prepare the bill run for you to approve in the bank" is a different task with a different error cost, and gets its own row next quarter.
The second pass: build cost
The frame ranks tasks, not builds. Two tasks with one score can differ fivefold in build cost, and that belongs in a second pass with one rule: the build estimate may re-order the top three, and never promotes anything from the bottom three. A cheap build does not rescue a low-ranked task; it delivers a bad automation sooner.
Estimate the top three in founder hours. The chase needs a connection to read the accounting system and one to send mail: about 10 hours, $1,000. Meeting notes need the transcript in and the notes out to the project tool: about 16 hours, $1,600. Ad budgets need four ad platforms, each with its own API and sign-in, plus a hold on large changes until a person approves: about 40 hours, $4,000. The pass confirms the order, and a tolerance of 1 says ad budgets need an approval gate before a build; gates are module 4.
Then the two sums for the winner. By hand: 2.3 hours a week × 4.33 × $30 ≈ $303 a month. Automated: $1,000 to build; $20 for the platform and about $1 of model calls, the model reading only clients' replies; and about 35 minutes a week on the exceptions those replies raise, 2.5 hours a month, $76. That is $97 a month, a saving of $206, payback in $1,000 ÷ $206 ≈ 4.9 months. The next two lessons turn the chase into an SOP and a definition of done.
In practice
Show the team the bands before the scores, and score in the room. The argument that follows is useful when it is about a column, not the total. When an account manager says ad budgets should go first because they take the most time, answer with the column: "Its tolerance is 1, because one misplaced zero costs a client $10,000. If every change over 20% waits for a person, the error that gets through is small, tolerance goes to 3, and it scores 300 and goes first." The frame says what would change the order, which is worth more than the order.
The mistake that costs money: ranking by hours saved. The longest task on this list is the one where a single error costs $10,000.
Check yourself
1. The account managers add a task: first drafts of client blog posts, 8 a week at 2 hours each. A factual error published under a client's name costs about $2,000, and the input is a one-paragraph brief. Where does it rank?
Frequency 2 (8 a week), cost 5 (16 hours a week, the most on the list), tolerance 2, input 1: 2 × 5 × 2 × 1 = 20, between the month-end pack and supplier bills. Not this quarter. A brief form with fixed fields would lift input to 4 and the score to 80, still below status replies on 96, so build the form first and score again.
2. The bank agrees that supplier payments under $500 can go out on a standing approval. Thirty of the 40 weekly bills qualify. Does the task move?
Bills under $500 become a task that may run unattended, with a recoverable error under $500: tolerance 3. Frequency 4, cost 3 (30 × 5 min = 2.5 hours), input 4: 4 × 3 × 3 × 4 = 144, second, ahead of meeting notes. The build pass keeps it there, since it reuses the chase's accounting connection. Bills over $500 stay at 0.
3. Your operations manager wants to drop frequency because cost already contains it. Estimate how many errors the chase and the month-end pack would each show in a year at a 2% error rate, and decide.
The chase runs 35 × 52 = 1,820 times a year: about 36 errors, three a month, a pattern within weeks. The pack runs 12 times: 0.24 errors a year, one every four years, found when it matters. Keep the column: frequency is how quickly an automation tells you it is broken.