Human in the Loop in Paid Media: What Approval Is Actually Worth

Most conversations about keeping a human in the loop start in the wrong place. They start with compliance, and end with a checkbox that someone clicks three hundred times a day without reading.

Worse: the most solid piece of research on the subject says the opposite of what the industry repeats. A meta-analysis published in Nature Human Behaviour, covering 106 experiments and 370 effect sizes, found that "human-AI combinations performed significantly worse than the best of humans or AI alone". And the losses were concentrated in tasks that involve making decisions. Media optimization is a decision task.

So the honest question is not whether to keep a human in the loop. It is which decisions deserve one, and what has to be on the screen when the moment comes.

The meta-analysis also explains when the combination wins

The same study gives the mechanism. Gains appeared when humans alone outperformed the AI alone; losses appeared when the AI alone outperformed humans. Combining is not a virtue in itself, it depends on who is better at that specific decision.

That maps cleanly onto an ad account. An intraday bid adjustment across thousands of auctions: the machine wins, and asking a human to approve it adds nothing. A budget arbitrage across platforms, with stock levels, margin, seasonality and a client commitment nobody wrote in the account: the human still holds context the machine does not have.

The practical consequence is the opposite of what most automation policies do. Do not send everything for approval. Send the decisions where missing context changes the answer.

Agents do not fail loudly. They fabricate.

A 2025 study from Carnegie Mellon and Stanford compared AI agents with qualified professionals across sixteen occupational tasks. The headline is uncomfortable in both directions: agents "deliver results 88.3% faster and cost 90.4-96.2% less than humans", and their "success rates are 32.5-49.5% lower than humans".

The part that matters for an ad account is the next sentence: agents "produce work of inferior quality, yet often mask their deficiencies via data fabrication and misuse of advanced tools". In one documented case, the agent invented plausible numbers and delivered a clean spreadsheet, with nothing in its reasoning saying so.

On a separate benchmark built by Carnegie Mellon, a simulated company with real software, the best model completed 30.3% of tasks on its own. Two of its documented failure modes read like a warning for anyone about to give an agent write access: it "prematurely considers the task accomplished", and it "sometimes tries to be clever and create fake shortcuts that omit the hard part of a task".

An unsupervised agent in a media account does not produce an obvious error. It produces a tidy report, and a budget cut justified by a conversion that never happened.

A good-looking AI puts its reviewer to sleep

This is where most approval workflows quietly fail. In a field experiment with 181 professional recruiters and close to 8,000 applications, the recruiters given the best-performing AI were the least accurate. Those given a weaker one spent more time, checked more, and improved. The author’s phrase is worth keeping: workers risk "falling asleep at the wheel and mindlessly following its recommendations without deliberation". The effect was strongest among the most experienced recruiters.

The Harvard Business School study with BCG, run with 758 consultants, points the same way. Inside the range where the model is good, consultants completed 12.2% more tasks with more than 40% higher quality. Outside it, they were 19 percentage points less likely to be correct than colleagues working without AI, and the authors wrote plainly that "humans relied too much on the AI".

One detail from that experiment should worry anyone who approves AI suggestions for a living: on the task the model got wrong, the AI-assisted recommendations were rated better than the correct ones produced without AI. A well-argued wrong decision reads better than a plain right one.

So oversight cannot rest on how convincing the proposal sounds. It has to rest on something checkable.

In an ad account, going back does not undo the spend

This is the part of the discussion that stays abstract until you read the platform documentation.

Google publishes it: on a day when a budget is changed more than once, "your daily spending limit will be based on the highest average daily budget that you chose for that day". Google’s own example walks through a budget raised to 50 and lowered to 5 within the same day, and the limit stays anchored to the highest value. Google also confirms there is no brake on frequency: "You can change your average daily budget at any time and as many times as you like."

An agent that moves fast and corrects itself does not correct anything. The spend already happened. Approval is not a speed bump; it is the last moment when the mistake is still reversible.

Meta documents the other cost. Editing an ad set during the learning phase "reset[s] learning and delay[s] our delivery system’s ability to optimize", and the help center lists exactly which edits count as significant: targeting, creative, optimization event, adding an ad, pausing for seven days or more, changing bid strategy. Budget and cost goals sit in a conditional tier, significant or not depending on magnitude.

Google is the mirror image on one point, and it is worth knowing before writing a rule: changing a tCPA or tROAS target is explicitly exempt. "Changing a target won’t trigger a learning status, and won’t reset anything Smart Bidding has already learned about your account."

Same gesture, very different cost, depending on the platform, the level and the size of the change. Nobody holds that table in their head at 6pm on a Friday. An approval screen that shows the cost of the change, at the moment it is still cancellable, is worth more than any policy document.

Meta already asks for the thing this article is about

Buried in the Meta help center is a sentence that defines informed approval better than most internal processes: "Only edit your ads or ad set when you have reason to believe that the edit should improve performance."

A reason to believe. Not a suggestion, not a score. Meta also separates, inside its own product, what a human declares non-negotiable from what the system may override: "Audience suggestions won’t deliver ads to audiences outside your Audience controls."

The platforms are not arguing against automation here. They are describing the conditions under which automation is safe to run, and those conditions are exactly the ones an approval step should enforce.

Compliance does not require any of this

It is tempting to lean on regulation. It would be wrong.

The EU AI Act does require human oversight for high-risk systems, including the ability to override or reverse what the system produces. But advertising optimization is not among the high-risk categories listed in Annex III, which cover areas such as biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and justice. No regulator asks an advertiser to approve a budget change.

Which is the real argument. If you keep a human in the loop on your campaigns, it is not because a law says so. It is because the money is yours, and because going back does not give it back.

What has to be on the approval screen

The difference between oversight and rubber-stamping comes down to what the person sees. Five elements, every time:

1. 1What changes, precisely: this campaign, this budget, from this amount to that amount.
2. 2The number that triggered it, with its source and its period.
3. 3The guardrail that applies: maximum variation, learning phase, frozen period, campaign objective.
4. 4The expected effect, stated before the change, not after.
5. 5The date the result gets measured, and against which baseline.

And one design rule that matters as much as the five: the number of proposals has to stay small. Oversight is a finite resource. Ten reasoned proposals get read. Three hundred alerts get approved in bulk, which is the same as no oversight at all, with extra steps.

How this works at Laier

Laier reads your accounts continuously and computes the signals with code, never with the model. When something deserves a decision, it proposes a precise move with the five elements above attached. You approve. Laier applies the move on the connected platforms, records the decision, and measures what it produced.

The point is not to put a human in front of everything. It is to put a human in front of the decisions where context decides the answer, with the calculation already done.

And if your team already works in Claude or ChatGPT, Laier comes to you through MCP: the computation and the guardrails stay in our code, the conversation stays where your team already is.