AI now makes most of the in-account decisions media buyers used to make by hand, and it makes them faster and better. What it cannot do is tell you the objective is wrong. It optimizes toward the goal and the data you give it, so the humans who set those two things now decide whether the spend works.
That is the argument, and it is not an anti-AI argument. Automated bidding and AI campaign types have absorbed a large share of the manual labor in paid media, and most accounts are better for it. The interesting question was never whether the machine is good at its job. It is. The interesting question is who draws the boundary the machine optimizes inside, and what happens when that boundary is drawn badly.
What the automation actually absorbed
Meta’s Advantage+ and Google’s Performance Max work on broadly the same publicly documented principle. You supply an objective, a budget, a set of creative assets and, for catalog businesses, a product feed. The system then handles audience selection, placement, bidding and asset combination inside the boundary you set. Automated creative tools generate and assemble variants. Agentic analytics assistants sit on top of the measurement stack and turn a plain language question into a query against your data.
Taken together, that is most of the daily work of a media buyer circa 2019. Pulling the audience, splitting the ad sets, watching the delivery, nudging the bids, cutting the loser, building the report. A competent buyer did all of that on a Tuesday. A machine now does the same work continuously, across more combinations than a person can hold in their head, and does not get tired at four in the afternoon.
This is a real gain and it deserves to be treated as one. The mistake is reading it as the job disappearing. What actually happened is that the job moved. Every one of those automated decisions is downstream of a decision a human still makes, and those upstream decisions got more consequential, not less, because the machine now executes against them at full speed.
What AI does well, and what still needs a person
It helps to be specific rather than philosophical about this. Here is the split as it shows up in a real ecommerce ad account.
| The decision | What AI does with it | What still needs a human | Why |
|---|---|---|---|
| Bid and budget allocation | Adjusts continuously across placements, audiences and times of day at a speed no team can match | Setting the objective and the efficiency boundary it optimizes inside | The system optimizes toward the target it is given. It has no way to ask whether that target is the right one. |
| Pattern matching across signals | Finds combinations of audience, placement, creative and timing that convert, at a scale no analyst reaches | Deciding which signals are trustworthy enough to learn from | A model cannot distinguish a real purchase event from a double counted one. It treats both as truth. |
| Creative variant generation | Produces and tests permutations of copy, crops, formats and asset assemblies at volume | Choosing the concepts and the variables worth testing in the first place | Variants explore a space efficiently. They do not choose which space is worth exploring. |
| Anomaly detection | Flags a spend spike, a CPA drift or a conversion drop within hours instead of at month end | Deciding what the anomaly means and what to do about it | An alert is a question. The answer is usually somewhere outside the ad account. |
| Turning a question into a query | Returns an answer in seconds where the same question used to cost an analyst an afternoon | Knowing which question to ask, and whether the answer is plausible | A fast wrong answer is more dangerous than a slow one, because it gets acted on before anyone checks it. |
| Catalog and feed delivery | Serves whatever the feed says is available to whoever looks most likely to buy it | Feed integrity, margin flags, stock accuracy | The system will sell an unprofitable or unshippable item exactly as efficiently as a good one. |
The failure mode is not bad optimization. It is excellent optimization toward the wrong thing.
This is the spine of the argument, and it is worth stating carefully because it is easy to caricature.
It cannot tell you the objective is wrong. Point a campaign at revenue when the business needs contribution margin, and it will buy you unprofitable growth with real skill. Every decision it makes will be defensible against the goal it was given. The reporting will look healthy. The bank account will disagree, and the disagreement will take a quarter to become obvious, by which point the machine has been compounding the error the entire time.
It cannot audit the signal it learns from. Feed it a purchase event that fires twice, or a conversion action that also counts a subscription renewal, or a pixel that has been double firing on one template since a theme update, and it will optimize brilliantly toward a fiction. This is the least glamorous failure and by far the most common one. Nothing in the system is designed to notice that its own training signal is wrong, because from the inside a corrupted signal and a clean one look identical. This is why tracking and attribution work has become more important as automation has improved, not less. The model is only as good as the events you hand it.
It will sell what the feed says you have. A product feed full of items you cannot ship, cannot fulfil profitably, or have already discontinued is not a data problem to an automated campaign. It is a shopping list. The system will find the buyers, and you will find out later.
The pattern in all three is the same. The garbage-in problem does not shrink as the models improve. It gets faster and more expensive, because a better model executes a bad instruction more completely. A mediocre system pointed at a broken objective wastes some budget. An excellent one pointed at a broken objective wastes all of it, on schedule, with a clean dashboard.
So the scarce skill moved up
If the machine handles execution inside the boundary, then value concentrates in drawing the boundary. In practice that is five things:
- Defining the objective. Not “more revenue” but the specific number the business actually runs on, usually contribution margin or a blended efficiency target that survives contact with the P&L.
- Guaranteeing the integrity of the signal. Deduplicated events, one agreed source of truth, a reconciliation between what the platforms claim and what the business banked.
- Setting the margin floor. Knowing which products and which customer cohorts can be bought profitably, and making sure the system is not allowed to scale the ones that cannot.
- Deciding what to test. Automation explores variations. Humans decide which hypotheses deserve budget, which is a judgment about the customer and the offer rather than about the account.
- Knowing when the machine is confidently wrong. The hardest one. Automated systems fail with the same tone of voice they succeed with, and the tell is usually a mismatch between the dashboard and something you know about the business.
None of these live inside the ad platform. All of them determine whether the ad platform produces anything worth having. As Jason Lu puts it: “It’s not the strategies within the ad accounts that drive the business, it’s the strategic layer above the ad account. Make sure the growth infrastructure is all in place so that ads can succeed.”
That thesis predates the current wave of agentic tooling, and the tooling has made it more true rather than less. When the in-account work was manual, a strong buyer could partially compensate for a weak strategic layer by grinding. That escape hatch is closing. When execution is automated and fast, the quality of the instruction is close to the whole game, which is why this has become the core of what an executive marketing partner is actually for.
What good supervision produces
Three examples from brands we work with, described by category rather than by name, all traced to Triple Whale exports held on file. In each case the in-account work leaned heavily on automated campaign types. The difference was upstream.
- An athletic apparel brand grew sales 35.7% to $9.27M between January and June 2026 while paid media spend rose 42%, with blended ROAS holding at 3.36x across the period. Holding blended efficiency flat through a 42% spend increase is the part that matters. The default for a scaling account is decay, because the incremental dollar reaches a colder audience than the last one did. No bidding algorithm produces that outcome on its own, because it is a portfolio decision made above the account.
- A fashion apparel and accessories brand grew sales 249% year over year with net profit up 205%, landing at a 41% net margin on a 29% MER. Growth and margin moving together is the signature of an objective that was defined in margin terms from the start.
- A womens fashion brand grew sales 99% with new customer CPA down 21% and new customer ROAS up 58%, while net profit rose 136% on a net margin of around 3%. At a 3% net margin, an order that looks perfectly healthy on platform ROAS can be a loss on contribution. An automated campaign optimizing to purchases cannot see that distinction. A human setting the margin floor can.
Will AI replace media buyers?
It has already replaced a good deal of what media buying used to mean, and it will keep going. Manual bid management, audience construction, placement testing and routine reporting are largely gone as billable human activities, and defending them is not a strategy.
What has not been replaced is the person who decides what the account is for. Buyers whose value was speed and thoroughness inside the platform are genuinely exposed. Buyers whose value is judgment about objectives, measurement, margin and testing are worth more than they were, because the machine now amplifies whatever instruction they give it. The same shift is visible in creative. Generating variants is close to free, so the scarce input is the concept and the testing structure around it, which is why performance creative increasingly looks like a research function rather than a production one.
The honest version of the answer is that AI replaced a set of tasks, raised the value of a different set, and moved the boundary between them upward. That is what most automation does. It is only unsettling if your role was defined entirely by the tasks on the lower side of the line.
Where to start
If you are running Advantage+, Performance Max or any agentic layer on top of your data, the useful audit is not of the campaigns. It is of the instructions. Can you state, in one sentence, the number those systems are optimizing toward, and is it the number the business actually runs on? Does your purchase event reconcile to banked revenue? Does the feed exclude items you cannot ship or cannot sell profitably? Do you know which of your products can be bought at a profit and which cannot? If any of those answers is uncertain, that uncertainty is being executed at machine speed right now.
Getting those answers straight is unglamorous work and it is where the return is. Automation has made the ceiling higher and the floor lower at the same time, and which one you land on is decided above the ad account.
If you are spending $50,000 or more a month on paid media and want an outside read on what your automated campaigns are actually being told to do, book a growth audit. You get the diagnosis and the efficiency picture either way, and it is yours to act on with us or without us.
Jason Lu is the founder of Plaid Testing and a Meta Business Partner. He presented “Meet Moby 2” on Triple Whale’s Customer Education Series, a session on agentic AI inside the ecommerce attribution stack, and has spoken on panel at The Whalies. To review your own setup, start with a growth audit.
Should I still use Advantage+ and Performance Max if AI can optimize toward the wrong goal?
Yes. These campaign types generally outperform manual equivalents at the job they do, which is finding buyers inside the boundary you set. The risk is not the automation, it is an unexamined objective feeding it. Fix the conversion event, the margin floor and the product feed first, then let the system run. Well supervised automation beats manual buying in most ecommerce accounts.
How do I know if my conversion signal is corrupted?
Reconcile platform reported revenue against what the business actually banked for the same period. Persistent gaps in one direction usually mean double counted events, a conversion action capturing more than purchases, or overlapping attribution windows across channels. Also check whether a single theme or template change coincides with the divergence. Most signal corruption traces to a deployment nobody connected to the reporting.
What does agentic AI change about ecommerce measurement specifically?
It collapses the time between a question and an answer, which is a real gain and a real risk. Analysis that took an analyst an afternoon now takes seconds, so more questions get asked and more decisions get made on the results. That only helps if the underlying data is trustworthy. Speed applied to a broken measurement stack produces confident wrong answers faster than anyone can check them.
Does this mean I need fewer people on paid media?
Usually you need a different mix rather than fewer heads. Hours spent on manual bid and audience management fall sharply. Hours spent on measurement integrity, margin analysis, creative strategy and testing design go up. Brands that cut headcount without reallocating that time tend to end up with efficient automation pointed at objectives nobody has revisited in three quarters.