Firing the Ad Agency

The owner of a small marketing agency told me, in our first meeting, that she had just fired her ad-operations vendor. For years she had paid an outside firm a meaningful monthly fee per client to run search keyword ads — the sponsored links that appear when someone searches for her clients' services. A few months earlier she'd started downloading the monthly performance exports herself and feeding them to Claude for analysis and recommendations. Two months of that outperformed the vendor. So she cut it.

Which is the moment most versions of this story end, and the moment the real problem starts. Because from the following month, the work the vendor used to do was hers — and her calendar was already full of sales meetings. "If it isn't this automated," she told me, "I won't do it. My schedule will eat it." She wasn't celebrating; she was cornered.

The job is four steps, and she'd automated one

The first useful thing we did was decompose the job, because "running the ads" hides four distinct tasks:

1. Collect — log into the ad platform, account by account, and download the raw performance data.
2. Report — analyze it, write up what worked, propose changes.
3. Execute — log back in and actually make the changes: add keywords, delete keywords, adjust bids.
4. Monitor — watch rankings and competitor bids daily, because the neighbor raising their bid quietly demotes you.

Laid out like that, her situation was obvious. She had automated step two — the analysis — and it was genuinely good. But steps one, three, and four were still hands on a keyboard. She had fired the vendor on the strength of a better report, while keeping the vendor's actual labor for herself. That's not replacing outsourcing; that's inheriting it.

The temptation was browser automation: teach the AI to log in and click what a person clicks. We tested it, honestly, and the honest verdict was: fragile. Login challenges, layout changes — anything built on clicking breaks on someone else's schedule. The better path was hiding in plain sight: the ad platform ships an official API. Performance queries cover collection. Bulk keyword operations — add, remove, pause, re-bid — cover execution. Scheduled queries with alerts cover monitoring. All four steps close, without a robot pretending to be a person at a login screen.

So that's the shape now. Data pulls itself on schedule. The report is a template — drop in the raw data, get the same structure every month, with proposed changes at the end. The proposals become API calls. Ranks and bids are collected daily, and she hears about them only when something actually moves.

The part that wasn't about ads at all

Halfway through the engagement, it became clear the ad pipeline was the smaller half of what she actually wanted. The bigger half came out in one sentence: her head planner is exceptional, and she wanted the effect of ten of her.

Her agency runs on a two-person shorthand built over years. The planner drafts, anticipates the owner's objections, and revises before showing her — so what reaches the owner is nearly final. That shorthand is the company's real asset, and it existed nowhere. Not in a wiki, not in templates. In two heads.

So we started the least glamorous project I ever recommend, and the one I believe in most: recording. Every meeting, every feedback session, every "no, not like that — like this" now gets captured and turned into text. The explanations the planner would give a new employee — why this layout for a conservative client, why that message order for a brand launch — get spoken out loud precisely because the AI, like a new employee, doesn't know the shorthand. From that corpus, two things are being distilled: written planning principles the AI drafts against, and a reviewer persona trained on the owner's own criteria that gives every draft its first pass, in her voice, before she ever sees it. The contrast is plainest at day's end. Before: drafts queued behind her sales meetings, waiting for the only pair of eyes that could judge them — reviewed at night, or not at all. After: a draft comes back in minutes carrying the objections she would have raised, and what reaches her is a second draft. She keeps the final look. She stops being the first filter for everything.

I gave her the same warning I give everyone building a persona from their own judgment: it only works if the judgment is consistent. If you call something red on Monday and blue on Thursday, the machine learns your noise, not your taste. Being forced to articulate stable criteria turned out to be valuable independent of the AI — it was the first time some of those rules had ever been stated.

What was actually replaced

I'd summarize the engagement this way: firing the vendor was an event, but it wasn't the achievement. Plenty of people cut a vendor, drown in the inherited work, and crawl back. What made it stick was moving each of the four steps into a system she owns — and starting the slower project of moving the judgment, hers and her planner's, into documents and a persona that don't leave when a person is busy, sick, or gone.

None of it works because the AI reviews better than she does. It works because first-pass review is the kind of job a person doesn't fail at so much as quietly abandon — the tenth draft at night gets a skim, a week of back-to-back meetings gets nothing. Written criteria get the same attention on the tenth draft as on the first. What was replicated wasn't her taste at its best; it was her taste showing up every time.

The caveats are real. The API doesn't cover every task on the platform; a few things remain manual, and we said so plainly. The persona is only as good as the recorded judgment behind it, and that corpus is months old, not years. But the direction is the point, and so is the real prize: the saving was never the vendor fee — it's that her own review instinct got copied into a system that works while she sells. She didn't stop paying for ad operations. She stopped renting it.