Automation did not remove judgement, it moved it

The common story about automation in UA is a subtraction story. Algorithms took over bidding and targeting, so the human decisions shrank, and the UA manager's job got smaller or vanished. The story is wrong in an instructive way. The decisions did not disappear. They moved to places where fewer people are looking, and the quality of those decisions now determines more of the outcome than bidding ever did.

That is the thesis: automation made each human judgement count for more while moving it out of the tools that used to display it. A practitioner could reasonably disagree, and the platforms' own marketing implies otherwise. But the operating evidence points one way.

What the algorithm actually took

Consider what a buyer used to decide, campaign by campaign: bid level, audience definition, placement, geography splits, budget pacing, creative rotation. Across AppLovin's Axon as much as Unity Vector or Moloco or the social platforms' automated products, most of those levers are now either gone or advisory. Unity's closure of the ironSource Ads network on 30 April, announced in late March as part of a refocus on Vector, was one more step in the same direction: fewer manual controls, more model.

What did not get taken is the set of decisions the model cannot make because they are inputs to it, not outputs of it.

• Which event the model optimises toward, and how that event is defined in the game's code.
• What the target value is, and on what horizon.
• What creative exists for the model to choose between.
• Which markets and platforms the studio is in at all.
• How success is measured independently of the platform's own reporting.

Every one of those is a judgement call. Each has a bigger effect on results than any bid a buyer used to set, because the algorithm amplifies it across every auction it enters.

The mechanism: why relocated judgement is more dangerous

A bid mistake in 2021 cost you a bad week on one campaign. You saw it in the dashboard and fixed it. The relocated decisions have three properties that make errors more expensive.

They happen less often. A studio defines its optimisation event once and reviews it yearly at best. A wrong choice persists across every campaign on every network.

Different people make them: event definitions live with analytics or engineering, target values come from finance, and creative comes from a studio team with its own incentives. The person accountable for UA results often does not own any of the decisions that now drive them.

They are hidden by the tools. Automated products report on what they did, not on whether the input they received was sensible. A campaign optimising toward a poorly chosen event will report excellent event-cost efficiency all the way to the write-down.

Creative is the clearest case. When the model chooses placements and audiences, the creative becomes the targeting. An ad that appeals to a particular kind of player is functionally an audience definition, and the model will find those players. The creative strategist is now making the decisions a media planner used to make, usually without knowing it and without the data the planner would have demanded.

The second-order effect nobody budgets for

The relocation created a skills gap in a specific shape. Studios staffed UA for the old job: buyers who could read auction dynamics and manage campaigns. Studios have cut or reassigned many of those people. Meanwhile the new decision points need people who can reason about event design, statistical validity of targets, and creative as a targeting instrument. Those are different skills, and they sit in different departments with no one coordinating them.

The result is that a lot of automated spend runs on inputs nobody has examined for years. The algorithm is very good; the question it received is stale.

There is a trade-off in fixing this. Centralising the relocated decisions under a growth function gives coherence but slows the game team and creates a bottleneck. Leaving them distributed keeps speed and loses accountability. Neither is free, and the choice depends on studio size more than on principle.

A framework for finding the decisions

A useful exercise for a UA lead is to trace each automated campaign backwards and ask who made each input decision, when they made it and what evidence they had. The structure that emerges usually has four layers.

• Objective layer: the event and target. Owned by whoever can prove the event predicts long-run value. Reviewed quarterly against fresh cohort data, not annually.
• Supply layer: which platforms, markets and formats the studio is present in. Owned by UA. This is the residue of the old media planning job and still matters.
• Creative layer: the portfolio the model selects from. Owned by creative, but briefed with the explicit understanding that each concept is an audience hypothesis.
• Verification layer: the independent measurement that checks whether the algorithm's reported efficiency is real. Owned by measurement, deliberately outside the platform's reporting.

The decision rule is that no automated campaign runs without a named owner for each layer, and no layer's decision is older than the most recent cohort data that could invalidate it.

An illustrative example

Take an illustrative mid-core studio running Vector and Axon campaigns optimised toward a "level 10 reached" event chosen two years ago because it correlated with day-30 payers at the time. Since then the game added a new early-game economy and level 10 now arrives in a third of the sessions it used to. The correlation has broken. Both networks report strong cost-per-event efficiency, and ROAS has drifted down for two quarters without a clear campaign-level cause.

Nobody bid wrongly. The relocated decision, the event definition, simply expired. Under the framework, the objective layer owner would have caught the break at the quarterly review and moved the event to a purchase-adjacent signal, and the algorithms would have redirected within days.

The lesson for structuring a UA team is not to hire back the buyers. It is to put someone in charge of the inputs, give them the authority to change event schemas and targets, and measure them on whether the questions the algorithms are answering are still the right ones. That role barely exists in most org charts, and it is now the most consequential job in acquisition.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• The best UA teams spend less time in ad managers
• Payback windows are a cost of capital decision