AI ad ops, demystified

E

Every paid agency has an AI slide now. Some have a whole deck. Bid recommendations from the platform. A creative tool that drafts ad copy. An anomaly widget that emails the account manager when CPA drifts. It looks like AI. It works like 2018.

That gap, between what the deck says and what the workflow looks like, is the entire story of paid media right now. Most agencies sold their clients an AI cover band. The auction still gets run on Monday morning, by a human, in a spreadsheet.

This post is what we mean when we say ad ops should run while you sleep. Where the platform handles itself. Where the strategist shows up to make calls, not push buttons. And where humans still earn the seat, because some decisions do not survive contact with automation.

The split that matters in paid right now is whether the AI is making decisions or filling in a template. Bolt-on AI suggests bids. A human types them in on Monday. Operator AI runs the auction in real time and tells the strategist what changed at the end of the day.

Bolt-on AI is the touchscreen in a 2019 Honda Civic. It exists. It works. The car is still the same car it was before the screen got installed. Operator AI is closer to what Tesla did. The software is not bolted onto the vehicle. It is the vehicle.

For paid specifically, that distinction shows up in a few measurable places. Time from creative brief to first impression. Hours per week the team spends inside Performance Max and Advantage+ versus building strategy. Whether your QBR slide on AI initiatives is a list of tools or a list of decisions that no longer require a meeting.

We treat ad ops as three buckets. Each gets automated to a different depth, and each has a different definition of done.

• 01Search term and keyword ops. Negative list hygiene, match type expansion, brand term protection, and the daily is-the-auction-working check. Claude scans the search terms report against a banned terms list, proposes negatives, and the strategist approves them in one review pass each morning.
• 02Asset rotation and creative testing. Variant generation, performance tagging, and the kill the loser inside 72 hours rule. We run a thin n8n flow on top of Performance Max and Advantage+, with Claude writing copy variants and a human signing off before anything ships at scale.
• 03Conversion event auditing. Most paid accounts have at least one duplicated or broken conversion event somewhere in the stack. GTM rules that fire twice. Server side mappings that drop the value. Klaviyo events that double report. Boring, mechanical work. It is typically the fastest place to find wasted spend in the first 30 days of any engagement.

The strategist is not gone. The strategist is more leveraged. Three things still belong with a senior human.

Bidding strategy. Specifically, the call to switch from Maximize Conversions to Target ROAS, or to ladder in value based bidding. Get this wrong and you lose 15 percent of next quarter's revenue. The machine will optimize anything you point it at. Choosing what to point it at is still the most leveraged decision in the room.

Brand voice exceptions. The line you do not cross even if it converts. The audience you do not target even if the LTV math works. Those calls are judgment. We have not automated taste and we are not in a hurry to.

New channel decisions. Whether to spin up ChatGPT Ads now or wait 90 days. Whether Meta CAPI is worth wiring before Black Friday. Whether to retire a long running Performance Max campaign. Those are roadmap decisions, not bid decisions. Different cadence, different brain.

“The auction can run while the strategist sleeps. The strategy cannot.”

There is nothing proprietary about the stack. Anyone can buy it. The leverage is in the playbooks, not the tools.

Claude for keyword research, copy drafts, and brief generation. ChatGPT for occasional second opinions on copy. OpenClaw for the platform connections, anomaly flags, and the orchestration layer that lets agents pass work to each other. A thin custom approval layer on top, so a strategist signs off on the day in 10 minutes instead of three hours. The full picture of how that fits together lives on the Build Your Own AI System page for operators who want to install something similar in their own team.

The same architecture extends to generative search when applied to Citation Share. Same agents. Same approval layer. Different surface. The architecture compounds across surfaces because the surfaces share an operating system.

A common tell of a bolt-on agency is the 40 slide quarterly business review. If the deck is doing the work the platform should be doing, you are paying for the deck.

We ship a one page weekly. CPA, ROAS, blended CAC, and the three changes the team made this week with the reasoning attached. That is it. The dashboard is live. The client owns the credentials. Nothing is hidden behind a screenshot.

If your current agency cannot show you a live dashboard you can pull up at 11pm on a Thursday, the answer is not a new dashboard. It is a new agency.

Most marketing teams paying 7,500 dollars a month for paid media are paying for manual ad ops with an AI slide on top. Under operator AI, that work should cost roughly 700 dollars a month in software, plus a senior strategist's time on review.

That math is uncomfortable to say out loud. It is the kind of thing that costs an agency deals. We would rather lose one on price clarity than win one by mispricing the work.

If you are trying to figure out what your current paid spend would look like under operator AI, the DIY-or-Agency Quiz is a fast way to find the gap. It will not take 20 minutes and it does not end in a sales pitch.

Operator AI is not a deck. It is a workflow change. If yours still runs the way it ran in 2018, the gap between what you are paying for and what you should be getting is bigger than it looks. Humans set the strategy. The machine executes.

Want this run for you? See how we operate Paid Media.