Agents in marketing operations
Put each action of a marketing-operations agent on a permission rung by what a wrong one costs, and cost a campaign-QA agent in human hours a week.
On this page
Four marketing-operations jobs now run with an agent in the loop: the weekly report, campaign QA, lead enrichment and outbound email. They differ less in how hard the work is than in what a wrong action costs and whether it can be undone, which decides how much the agent may do without a person. The worked case: a campaign-QA agent checking 200 ads a week for two hours of human review.
Four jobs, four costs of being wrong
| Job | What the agent does | A wrong action | Can it be undone? |
|---|---|---|---|
| Reporting | pulls spend and conversions, drafts the weekly summary | a wrong number in the board deck | yes, before it is sent |
| Campaign QA | checks each new or edited ad against a checklist | a broken ad passed, or a good ad flagged | a flag costs a reviewer a minute |
| Lead enrichment | fills company size and industry on CRM records | a lead routed to the wrong rep | yes, from the field's history |
| Outbound | writes and sends first-touch emails | a false claim in 2,000 inboxes | no |
The order is the cost of the worst plausible mistake, not the difficulty. Outbound is the easiest to build and the hardest to permit: a sent email stays sent. A wrong enrichment write is one restore away; what costs money is the routing rule that fires on the new value within minutes.
The permission ladder
An agent's permissions sit on four rungs:
- Read. It sees ads, pages, spend and records, and changes nothing.
- Draft. It produces something a person sends, applies or discards.
- Act with approval. It makes a change after a named person approves that change.
- Act. It makes the change and tells someone afterwards.
The rung belongs to an action, not to the agent. The campaign-QA agent reads every ad, drafts a flag for every failed check, and for exactly one action, pausing an ad whose landing page is down, acts on its own. One rung for the whole agent forces a choice between an agent too timid to stop a dead ad and one trusted to rewrite approved copy. Where a person signs is the subject of approval gates; the role itself, written as six fields (inputs, outputs, definition of done, manager, escalation path, budget), is the virtual employee frame.
An action moves up a rung on evidence, weeks of clean audits, not a good demo, and down at the first wrong one.
Worked: the campaign-QA agent
The team launches or edits 200 ads a week across two ad platforms, and each ad is checked against 12 rules. Seven need no model: the landing page loads, the tracking parameters follow the convention, the headline fits the character limit, the offer has not expired, the ad's price matches the page's, and the countries and daily budget match the brief. Five need judgement, which is where a language model earns its cost: no claim the product cannot prove, tone against the brand's reference set, image and copy describing the same offer, a disclosure line on any endorsement, and a competitor named only in wording legal has cleared.
The seven run as code, which gives the same answer every time. Asking a model whether a URL resolves invites a confident "looks fine"; workflow scripts and deterministic harnesses makes the case for determinism in the harness.
The arithmetic, at the course's $100 marketing hour:
checks a week 200 ads x 12 checks = 2,400
flags 3% of checks fail = 72
review 72 flags x 1.5 minutes = 108 minutes, about 2 hours, $200
by hand today 2,400 checks x 15 seconds = 600 minutes, 10 hours, $1,000
model cost assume under $0.05 an ad = under $10 a week
About eight hours, $800, come back each week for under $10 of model calls, provided the queue stays small.
The mistake: an agent that fixes what it flags
The tempting design gives the agent write access so it repairs what it finds: pauses the failing ad, shortens the headline, corrects the budget. It breaks on the flags that are wrong.
Suppose one flag in four is false: 18 wrong changes a week to live ads. A wrong pause on an ad spending $300 a day, made on a Friday evening and noticed on Monday, withholds $900 of delivery, close to three customers at the course's $333 each. A wrong headline rewrite is worse in kind: it puts unapproved copy live, and the check most likely to misfire, the unprovable claim, is the one whose wrong fix creates the problem it exists to catch. The coordinator now audits edits instead of flags.
The fix is a rung per action. Every failed check is a draft: the rule, the evidence and a proposed fix, in the coordinator's queue. Copy, budget and targeting change only with approval. One action acts alone: pausing an ad whose landing page fails twice, 10 minutes apart. A dead page converts nobody, so a wrong pause costs hours of delivery, while a late one costs every hour the spend runs. It is capped at five a day; a sixth means the site is down, and the agent stops and pages the coordinator.
Knowing what it misses, and taking it back
The queue shows the 3% of checks that failed. The agent's misses are in the 97% that passed, which nobody looks at unprompted. Each week the coordinator audits 10 passed ads at random, 5% of the week's 200, on all 12 rules: 120 checks at 15 seconds, 30 minutes. If 40 audited ads over four weeks show no miss, the miss rate is still plausibly up to 3 ÷ 40 = 7.5% (the rule of three: zero failures in n bounds the rate near 3/n); "under 1.5%" takes 200 clean ads, 20 weeks. Each miss found becomes a test case, which is how an evaluation set grows in production; evals as the unit test of LLM work builds one deliberately.
Rollback needs two things written before launch. Each applied action logs the prior state and the rule that fired, so one command restores it, and running the restore twice changes nothing more, the property idempotency names. And the write credential is separate from the read one, so revoking it drops every action to draft without stopping the checks.
In practice
Present an agent to the team as a list of actions with their rungs: "it reads everything, flags into the queue, and pauses a dead-page ad on its own, five a day at most", with the human time beside it, about 2 hours of queue and 30 minutes of audit. Ask a vendor which actions the agent takes without a person, how each is undone, and where the log is. "It figures out the right action" is the top rung for every action at once.
The mistake that costs money is granting write access because the flags looked right in the demo. At one false flag in four, write access turns 72 flags into 18 wrong changes to live ads every week.
Check yourself
1. The enrichment agent writes company size into 1,500 CRM records a month, and routing sends companies over 200 employees to the enterprise rep. An audit of 50 records finds 4 sizes wrong. Which rung does the write sit on?
Split it. 4 in 50 is 8%, about 120 wrong records a month. A wrong size that stays on one side of 200 changes nothing downstream and is restorable, so those writes act. A write that crosses 200 misroutes a lead within minutes, so only those, a small share of the 1,500, go to approval until an audit of crossing records comes back clean.
2. Ad volume doubles to 400 a week and legal adds three checks. At the same 3% failure rate and the same review and audit times, how many hours a week does supervision take?
400 × 15 = 6,000 checks; 3% is 180 flags × 1.5 minutes = 4.5 hours. The 5% audit is 20 ads × 15 checks × 15 seconds = 75 minutes. About 5.75 hours a week, under a quarter of the 25 hours the larger checklist would take by hand. Find the check producing most false flags and fix it before adding a reviewer.
3. The vendor offers to let the agent rewrite over-length headlines automatically. Do you allow it?
No. The platform truncates a long headline, so waiting costs little, while an automatic rewrite puts unapproved copy live and can introduce the very unprovable claim a check exists to catch. Keep it at draft: the flag carries the shortened headline, accepted in the same 1.5 minutes.