What we automate in paid media, and where we keep human judgement

Most paid media teams already rely on automation every day: automated bidding, broad match, product feeds, platform recommendations and scripts. Generative AI adds another layer, able to read reports, summarise changes and draft commentary in seconds. The useful question is not whether to use it, but which parts of the work it should do, and where a person needs to stay responsible for the outcome.

Our answer is a boundary based on consequence. Work that prepares, organises or checks can be automated heavily. Work that commits budget, changes what customers see, or depends on context that is not in the data stays with a person.

• Gathering data from several platforms into one view, with the basis of each number labelled.
• Checking for anomalies: spend that has stopped, pacing that has drifted, conversions that have disappeared.
• Classifying at volume, such as grouping search terms by likely intent or tagging creative by format and message.
• Drafting a first version of a report or summary for a person to correct.
• Preparing a proposed change in a form someone can review quickly, with the evidence attached.

In each case the output is an input to a decision rather than the decision itself, and a mistake is cheap to catch because someone looks at it before anything happens.

Some decisions depend on information the data does not contain, or carry consequences that are hard to reverse:

• Approving spend changes and budget moves, especially large ones.
• Publishing or changing what customers see: ads, claims, offers and landing pages.
• Interpreting performance in context. A report may show lower revenue without knowing that a bestseller was out of stock. Cheaper leads may be worse leads, and only the sales team can say.
• Deciding what matters. A model can list every movement in a dashboard; choosing the two that change the recommendation is the consultant's job.
• Anything that touches a client's commercial position, legal claims or reputation.

A useful AI-assisted process lets the reviewer trace each statement to its source. If a draft says revenue fell, the reviewer should be able to see the figure, the period and the comparison without rebuilding the analysis. It should also separate what is known from what is inferred. Revenue declined is an observation. Revenue declined because the creative is tired is a hypothesis that needs evidence.

Good analysis ends with a therefore: a proposed action, who owns it and when it will be reviewed. Automation can help structure that, but the recommendation belongs to someone who can be asked why.

A tool that produces a convincing draft quickly is not automatically a time saving. The measure is the total time to a reliable outcome, including checking and correcting. We start with small, bounded outputs such as a draft analysis, compare them with the source data, record factual errors and omissions, and decide what review is still needed. Autonomy expands only where the error rate is known and the cost of a mistake is small.

• Can the reviewer inspect the evidence behind each statement?
• Can the output be corrected without restarting the whole task?
• Is there a named owner for the decision that follows?
• Can the next step be stopped before it affects spend or customers?
• Is the error rate being recorded, rather than assumed?

If any of those answers is unclear, more autonomy is premature.

We are developing and testing AI-assisted workflows for preparation, monitoring and first drafts. They are tools that support our consultants. They do not run accounts unattended, and we do not treat a workflow as finished until it has been checked against real work.