What got automated and what did not

Generation made most marketing tasks cheaper and almost none of them free. Sort a team's recurring work by what each task still needs from a person (a reader, an owner, or a decision) and the hours saved come to about half of what a count of tasks suggests. The same sort says which jobs the change created, and so where the next hire should come from.

Three columns, and what puts a task in each

Every recurring task lands in one of three columns, decided by what it still needs from a person.

• Prompt. One request, one output, and a person reads every output before it is used. The task's cost falls to the cost of reading. Ad variants, a call summary, a first draft.
• Workflow. A trigger, fixed steps and a check, with no person on each item. Someone owns it: builds it, handles the exceptions, fixes it when an input changes. The weekly report, campaign QA, routing a lead.
• Judgment. The output is a decision, the information it needs is in no system, or being wrong costs more than the hours saved. Choosing the segment, reading a holdout, approving a claim.

The test between the first two is whether a person still looks at every item. The finer line between a script, a workflow and an agent is drawn in scripts, workflows and agents, and which task to automate first is the scoring frame. A marketing team needs only the three columns and the hours in each.

The ledger for a $3M ARR company

The running company for this module is the invoicing app from the funnel lesson, two years on: 5,000 paying customers at $50 a month, which is $3M of annual recurring revenue (ARR), and a marketing team of three. Three people at the course's 176 working hours a month have 528 hours. The 30 recurring tasks below take 330 of them; meetings, planning, holidays and leave take the rest. Every figure is an assumption, priced at the course's $100 loaded marketing hour.

The blog row is the course constants at work: a written post is four hours, a generated one is a 30-minute review, so 8 posts go from 32 hours to 4. Every prompt row follows that pattern. The drafting goes; the reading stays.

The mistake: counting tasks instead of hours

Count rows and the case for a smaller team writes itself. 21 of the 30 tasks, 70%, are now a prompt or a workflow, and those 21 took 252 of the 330 hours. A founder reading that sees three marketers doing one marketer's work.

Count hours after the change instead, and add the two costs the rows leave out:

prompt rows        150 h  ->   34 h    the reading stays
workflow rows      102 h  ->   17 h    exceptions only
workflow upkeep           +    40 h    ten workflows, each breaking about once a month
review standard           +    30 h    sampling and drift checks, next lesson
judgment rows       78 h  ->   78 h    unchanged
total              330 h  ->  199 h    131 hours freed, about $13,100 a month

131 hours is less than one person's 176. The automatable tasks fell by 80%, from 252 hours to 51, but the team's hours fell by 40%, because judgment did not shrink and the workflows and the review standard brought work of their own.

Upkeep is the figure most often left at zero. Ten workflows that each break about once a month, at four hours to find and fix the cause, is 40 hours, and a broken workflow fails quietly: the wrong email goes to the wrong list until someone notices, the subject of drift and silent failure.

Nor do freed hours have to leave as headcount. Customer interviews sit at 20 hours a month, about ten conversations. The judgment rows are the work a team never had time for, and the place freed hours compound.

The roles the ledger creates

Each column needs an owner, and two of the three owners are jobs a team of this size rarely had before.

Expand

The task ledger for a marketing team of three: 30 recurring tasks and 330 hours a month split into prompt tasks (150 hours down to 34), workflow tasks (102 hours down to 17, plus 40 hours of upkeep) and judgment tasks (78 hours, unchanged), each column owned by a role100%−+FitSVG (opens in a new tab)CloseEsc

Pinch to zoom · drag to panScroll to zoom · drag or arrow keys to pan · 0 fits · Esc closes

Marketing ops that builds automations. Owns the workflow column: builds the ten workflows, carries the 40 hours of upkeep, and answers for each one's exception rate and rollback. It is the campaign-QA job from agents in marketing operations run across every workflow: a permission boundary, a check, a way back. Measured on hours returned to the team and on sends that went wrong.

Brand steward for generated output. Owns what the prompt column ships: writes the review standard, reviews every paid ad, samples the rest, and checks for drift. In a team of three it is about 30 hours of one person's month, and it has to be named, because otherwise everyone who generates also approves their own work and nobody measures the total. Brand stewardship of generated output sets the standard.

Growth engineer. Writes code in the product and the site: onboarding experiments, the product events that trigger lifecycle sends, the plumbing for holdouts. Marketing ops builds on top of tools; a growth engineer changes the product the funnel runs through. Measured on a funnel rate: activation from 30% to 33% is a 10% lift, about ten paying customers a month.

The judgment column stays with the marketing lead, and grows: more output means more claims to approve and more results to read.

A task's column is not permanent

A task moves from judgment towards a workflow when its decision becomes a rule with a number in it. Once "pause any ad whose cost per sign-up passes $100 for three days" is written down, budget pacing is a workflow, and the judgment left is setting the $100.

A task moves the other way when a wrong output starts to cost more. A social post is a prompt until it carries a customer's words; then approving it is judgment, because a testimonial carries legal weight, set out in ethics and disclosure.

Re-run the ledger each quarter. A prompt task whose reader approves 98 outputs in 100 unchanged is a candidate for a workflow with a sampled check; a workflow whose exceptions take more hours than the task used to is a candidate to move back.

In practice

To a founder or a board, present the ledger in hours, with upkeep and review counted: "Generation took the automatable tasks from 252 hours to 51, and the team's total from 330 to 199. The 131 freed hours go to customer interviews and the lifecycle programs we stopped running, and the next hire owns the workflows." To the team: every row has one owner, and every workflow has a named person who hears first when it breaks. To a vendor selling "an AI marketer": which rows does it move, and who maintains it when an input changes?

The mistake that costs money is cutting headcount on the task count. "70% of tasks automated" reads as two of three marketers redundant. On hours the team freed 131 of 330, less than one person, and cutting two leaves ten workflows with no owner and the judgment rows with no time.

Check yourself

1. A vendor offers to move the 20 hours of customer interviews into the prompt column by interviewing synthetic personas. Do you move the row?

No. The interviews are judgment because what they produce, a customer's reason in their own words, exists in no system yet. A persona can draft questions and hypotheses, the line synthetic research draws. Use the tool for the hour of question drafting and keep the 20 hours of talking.

2. The founder will fund a hire that returns at least $150,000 a year. Does the ledger justify a marketing-ops builder on hours alone?

No. With a builder carrying the upkeep, the workflow column returns 102 − 17 = 85 hours a month to the team: $8,500 a month, about $102,000 a year. The rest of the case has to come from what the builder makes possible, such as the lifecycle sends nobody has time to run, and that is a number to estimate, not assume; hire a marketing-ops builder or a growth engineer works it.

3. Social posts are read by whoever generated them, and 97 in 100 are approved unchanged. Do you move the row to a workflow?

Partly. Posts with no customer, number or claim in them move to a workflow with a random sample the brand steward reads, taking the 5 hours of reading to about one. Posts carrying a quote, a price or a result keep a reader on each, because one wrong claim there costs more than the 4 hours saved.