Anthropic's Marketing Team Cut a Two-Day Report to Two Hours. The Job That Survived Is Checking the Numbers.

Most companies describe their AI wins in percentages and leave the plumbing out. On July 8, Anthropic did the opposite and published how its own marketing operations team rebuilt its two most manual workflows, naming the systems, the failure modes, and the person whose week changed.

Ian Chan, on the marketing operations team, owns the weekly metrics report. Before: “Ian used to spend a day to two days every week tracking down data and validating it.” After: “the entire process, which used to take up to two days of work, takes up to two hours.”

That’s a real number, and it’s the least interesting thing in the post.

The interesting part is which two hours survived. Not the data pulls, which now run against Salesforce, HubSpot, Swoogo, Asana, and Gmail on their own. Not the assembly. What’s left is verification, and Anthropic says so plainly: “Human validation has become an integral part of both workstreams.”

Automation didn’t remove the work. It removed the half a person could do half-asleep.

The mechanism is worth understanding, because it’s the part that transfers. The team works through agents, software that can take actions across connected tools rather than just answer questions in a chat window. Each recurring job is packaged as a “skill,” a saved set of instructions the agent reuses so it runs the task the same way every time. Ten of them exist across the two workstreams: prep, proofreading, action items, dispatcher, event build, audit, webinar landing page, apply to attend, approval support, data import.

Two design choices in that list are the whole story. A proofreading skill exists as a separate step, because the system that assembles the numbers isn’t trusted to also confirm them. And when sources disagree, “When the numbers don’t line up, Claude flags the mismatch instead of guessing.” On the campaign side, Annabel Custer, who focuses on campaign operations, keeps the last word: “Annabel reviews each result before it ships.”

So the architecture encodes an admission. The output is good enough to build on and not good enough to send.

PPC Land’s account of the same post lands on the same conclusion from the outside, describing marketing operations moving from execution toward governance and enablement, with the remaining human work concentrated on defining data, validating outputs, building processes, and explaining why two systems disagree. It also names the tension the post is really about: “marketing operations teams own automation as their mandate, yet a large share of the day-to-day remains stubbornly hands-on.”

Robert Gillespie, who writes the Marketing with AI newsletter, put the industry version of this in a July 12 roundup:

“There’s a single thread running through all three: the work is shifting from driving AI to supervising AI that acts on your behalf. The agent now has write access to your inbox, your ad account and your files—which makes the review step, not the drafting step, the job.”

He’s describing a permission change, not a capability one. And it isn’t hypothetical for anyone running paid media: Google’s July terms now let its systems generate targets, ad copy, and destination URLs on the advertiser’s behalf while the advertiser keeps the liability. Agencies running research on synthetic audiences draw the same line at the same place, with humans vetting the last 20%. Three unrelated corners of the industry, one shape: the machine produces, the person certifies.

Here’s the part that should worry a founder-marketer more than the headcount math. Checking is a senior skill built from junior repetitions. You learn that a number is wrong by having produced a few thousand of them, and the production is what just got automated. We tracked the front end of that in the vanishing entry-level marketing role. A team can buy the two-hours-instead-of-two-days outcome immediately. It can’t buy the judgment that makes those two hours worth anything, and nobody has proposed where the next Ian Chan learns it.

There’s also an honest limit on how far this evidence carries. This is a lab publishing about its own product, written by people with unusually deep access to it. The post is a demonstration as much as a report. No independent time study confirms the numbers. What makes it useful anyway is the specificity: named systems, a named proofreading step, a named human sign-off, and a stated failure mode. That’s more than most vendor case studies risk putting in writing, and every claim is checkable against the workflows it describes.

Gillespie’s line about what expertise now means is the one to keep:

“skill meant producing. Now that agents produce and act, skill means supervising—briefing well, setting guardrails, and checking the work before it ships.”

The move that follows is small and specific. Before automating a recurring marketing task, write down who checks the output and against what source, and treat that as part of the build rather than a step you add after something goes wrong. Anthropic’s own advice for getting there is the cheapest process change on offer: “When you find yourself correcting Claude on the same thing more than once, that feedback belongs in a skill.” A team that does this ends up with two hours of real work. A team that skips it ends up with a report nobody has read, arriving faster than before.