The Ad Manager Job Is Going Away. I Know Because It Was Mine.

Quick Take

Two years ago my job was logging into ad accounts and doing things by hand. Pull the search terms, add the negatives, move the bids, write the report. That job is disappearing and I would rather say it out loud than pretend it isn’t happening. What replaced it, for me, is a system that carries every client I have ever worked on and gets better each time I use it. Clients don’t pay me for hours anymore. They pay for access to that. Below is what it is, what it has produced, where it breaks, and the part I have not built yet.

The Call That Made the Point

Last week a furniture maker in Illinois booked a call with me. He builds custom tables, five to fifteen thousand dollars each. He had just finished ten months with an agency: twelve leads, one sale.

I asked how he found me.

He said ChatGPT. He had typed in a description of his business and asked for someone who would fit a company his size instead of an agency. It gave him my name.

No ad. No referral. No Google search in the way that phrase meant something three years ago.

I have spent about a year building content specifically so machines would read it and recommend me. Forty-five answer pages sitting at /wasted-ad-spend/, written to be quoted by an AI rather than skimmed by a person. That call was the first time a stranger arrived through it and told me so on the record.

That is the whole thesis in one meeting. The front door moved. The people who noticed early are getting the traffic.

What I Used to Sell

I was an ads manager. Good one. The work was real: search term reports, negative keyword sweeps, bid adjustments, monthly reporting.

Every one of those tasks is now something I can hand to a machine and check.

I am not being dramatic about this. On a call this month I told a peer that ads managers are being replaced and that I had pivoted my whole business because of it. I meant it. If your entire value is executing a checklist inside an ad platform, the platform is coming for that job, and so is every operator who wired up a system to do it.

The mistake is thinking the answer is to do the same job faster. It isn’t. The job changes shape.

What I Run

I want to be specific here, because “I use AI” has become a meaningless sentence.

A memory bank. An Obsidian vault, 232 notes as of this morning. Every client, every project, every landing page, every mistake I have made and the rule I wrote so I would not make it twice. Plain markdown files. Nothing proprietary. I wrote about why plain files beat buying a memory product in Your AI Agent Doesn’t Need a Brain.

The notes link to each other. A landing page note for one client connects to the skill that built it, which connects to three other clients where the same approach worked. When I start work on an account, the agent reads that web first. I am not briefing it from scratch every session. It knows how that business talks, what the owner has already rejected, and what went wrong in March.

112 skills. Repeatable procedures written down: how to build a PMax campaign, how to audit a Shopify storefront, how to verify a Google Ads change before it goes live, how to write in a specific client’s voice. Each one is a file the agent loads when the task comes up.

A headless Mac Mini. Sits under my monitor with no screen attached. It runs the recurring jobs on a timer: call scoring for a law firm every weekday evening, a lead engine Monday morning, blog drafts for two stores midweek, negative keyword sweeps, monthly client reports on the first of the month, an operations brief to my inbox at 8:30 every morning.

That is the honest inventory. Files, procedures, and a cheap computer running scheduled scripts.

The Part Most People Get Wrong

Here is where I want to correct something I have said too loosely on sales calls.

When I describe this, it is easy to make it sound like an autonomous agent waking up each morning and thinking about your account. It isn’t that.

Most of what runs on a schedule is deterministic. Python scripts, cron jobs, a fixed pipeline. The language model gets used at specific points where judgment or writing is needed, and the output goes through a verification gate before anyone sees it. My monthly report system, for example, pulls the numbers deterministically, makes exactly one model call per client to write the narrative, and then runs a three-layer check: numbers against the cards, banned phrases, and anomaly smell tests. Anything that fails gets a REVIEW_ prefix and never reaches a client as finished work.

The reason it works is not that the AI is smart. It is that the system is boring and the context is good. A model with no memory of your business writes generic garbage. The same model with two years of notes about your business writes something worth reading.

Same idea I used when an agent built my tracking stack. The leverage is in the context, not the model.

What It Has Produced

Real things, with the caveats attached.

I closed an engagement this month with a manufacturer who had interviewed seven agencies before picking me. Their last vendor ran ads through the vendor’s own ad accounts, showed them charts of clicks, and produced one accidental purchase against a fifteen thousand dollar invoice. What closed it was not a clever deck. On the call I shared my screen and walked them through the vault: here is every note about your business, here is how it connects to other work, here is the rule I wrote after I got something wrong. Their technical guy’s verdict was that it looked like I was using AI properly. That was the deal.

One store I have run since last summer went from roughly $20,000 a month in orders to $49,000 within two months of the ads going live. This January through March, with Google and Meta running together on about $4,000 a month of combined ad spend, it did $87,131, then $92,365, then $116,129. I had been quoting those three months from memory on sales calls, so before writing them here I re-pulled them from the raw order records. The recall held to the dollar. And to keep this honest: the account in the next paragraph, the one whose Performance Max fell off a cliff, is the same store. The curve is real. So is the dip that followed it. Both live in the same order records.

The system also catches things I would have missed. On one account this month it surfaced that Performance Max efficiency had fallen off a cliff, ROAS from roughly 14.9x down to 3.5x and cost per acquisition from about $100 to $408, while click volume went up. Traffic quality collapse, not a demand problem.

Then it found the worse thing. That account’s historical numbers had been inflated for months, because the Shopping campaigns were counting add-to-cart and begin-checkout events as conversions. One month showed $146,618 in “conversions” with zero purchase value behind it. Real purchases over five months were about ten.

I had to go tell the client that the good numbers were not real. That is the least fun conversation in this business and it is exactly the conversation the system exists to force.

Where It Breaks

If I only told you the wins this would be another AI hype post, and I have written about spotting that.

The agent on that Mac Mini crash-looped 262 times in a single day because one invalid line got written into its config. I found out because it went quiet.

Before that it ran the API bill to about $56 a day by dragging a 207,000 token conversation into every single call. I fixed it by resetting the session nightly, not by anything clever.

It has run out of API credits mid-week and silently stopped.

I now have a watchdog that checks its health every five minutes, restores the last known good config, restarts it, and emails me what it did. I built that because the thing failed, not because I planned well.

The other honest limit: the AI is worse than me at knowing what matters. It will happily produce a competent report about the wrong thing. Every client-facing output still goes through me. When I tell a client they are paying for access to me, the access is the part that keeps the machine pointed at the right problem.

Where I Think This Goes

My read, stated plainly so you can hold me to it.

The single-channel specialist role goes first. Nobody is going to pay a person a monthly fee to log into Google Ads and add negative keywords, because that is now a scheduled job.

Most mid-market agencies follow. Not the top tier working on enormous accounts, but the ones charging a startup two grand a month to have three junior people each touch one platform. That model is a stack of individual jobs, and the jobs are what is being automated. I think five years is the outside number.

That is a claim about a pricing model, not about the people inside it. The good shops already know. Some of the sharpest operators I know work at agencies and are rebuilding the same way I did, and the ones who move will come out of this fine. The ones billing the same retainer for work that now runs on a timer are the ones with a problem.

What replaces it is a subscription to a system that has context. Not a person selling hours, and not software you have to operate yourself. Someone who has built the accumulated record of what worked across many businesses, plugged you into it, and stays personally responsible for the outcome. The value compounds because every client makes the system better for the next one.

Here is the part I have not built and will not claim I have: landing pages that update themselves against live campaign performance, season, and trend, with no human in the loop. Right now I can build a page fast, informed by data about which pages converted before, and change it quickly when the numbers move. That is assisted, not autonomous. The autonomous version is where I am pointing everything, and I will write about it when it is real rather than when it is a slide.

The reason I am doubling down anyway is that the market has not caught up. Almost nobody selling marketing services right now has this built. That gap closes, and the people who started early will have years of context that the late arrivals cannot buy.

What This Means If You’re Buying

You do not need to care about any of my tooling. Most of my clients don’t, and I have stopped showing them the vault unless they ask.

What you should care about is the shape of what you’re buying.

If you are paying an agency and your point of contact is a junior account manager reading you a dashboard, you are paying for a job that is being automated, at the price of a person doing it by hand.

If you are being sold “AI-powered” anything with no explanation of where the context comes from, it is a model with no memory of your business writing generic output. You can tell because it reads generic.

The question worth asking a vendor is simple. What do you know about my business that you would still know six months from now if the person on this call left? If the answer is a folder of PDFs and one account manager’s memory, that is the risk.

I am not selling you a quick fix, and I will never promise you a number. You are paying for the work and for the system the work runs through. That is the whole offer.

If you are still working out which shape you are buying, the comparison is at freelancer, agency, or fractional marketer, and the number at what a fractional marketer costs.

If you want the diagnosis before the engagement, that is what the paid audit is for. It is written-first, no call required, and if I think the fix is upstream of anything I would run, I will tell you that instead of selling you a retainer.