How to use AI as a copilot for budget and channel decisions: what it reads well, where its advice goes wrong, and the checks to run before you act on it.
Media buying is a series of decisions made with incomplete information: where next month's budget should go, whether to cut a channel showing mixed results, whether a weak campaign needs patience or a rebuild. AI tools now offer answers to all of these, from the ad platforms' own recommendations to optimisation software and general AI assistants you can hand a spreadsheet.
AI helps most where the problem is the sheer volume of data: spotting patterns across hundreds of ads, flagging unusual changes early and summarising performance. It misleads when its advice rests on assumptions it does not show you, above all how credit for a sale is shared between channels, and when the right answer depends on things outside the data, such as margins, stock or your sales team's capacity. Use it to inform the decision, not to make it.
The cost of getting this wrong is real: a confident recommendation can move thousands of euros in the wrong direction, and the damage often shows up weeks later. This article sets out what to hand to AI with confidence, what to question every time, and a routine for using it as a copilot rather than an autopilot.
Pattern recognition across large numbers of ads is AI's clearest advantage. Working out which elements go with a lower cost per sale (the hook in the first three seconds, the format, the length, the offer, a person on screen or the product alone) across hundreds of variations is beyond what anyone can do reliably in a spreadsheet. AI can sort it in minutes and give your next creative brief a factual starting point. Treat what it finds as ideas to test rather than conclusions, because a pattern across 300 ads can still be chance or a side effect of where the ads happened to run.
Anomaly detection is the second strength. A model that knows each campaign's normal range can flag a departure within hours rather than at the next weekly review. It is a high-value, low-risk use that almost every account should have running, because what it catches is expensive when missed.
Summaries are the third. Turning exports from four platforms into one readable page is tedious work that AI does well, provided someone checks the totals against the source before anyone acts on them.
AI recommendations are only as good as the data behind them, and the most important assumption in that data is attribution: the rule that decides which ad gets the credit for a sale. Each platform sees only its own part of the customer's journey and counts conversions by its own rules, so the totals across platforms often add up to more sales than you actually made.
A model that moves budget based on last-click results, where the final click before a sale takes all the credit, will steadily undervalue anything that starts journeys rather than finishes them. Video, social media ads aimed at new audiences and, for business-to-business firms, LinkedIn tend to lose out. Searches for your own brand name, which often follow exposure to those ads, tend to win.
Say Google Search reports a cost per sale of €40 and your Meta campaigns aimed at new audiences report €90. A tool trained on last-click data will shift money from Meta to Search. But if many of those Search clicks come from people typing your brand name after seeing a Meta ad, Search sales fall a few weeks after Meta is cut, and the saving turns into a loss.
The recommendation in that example was internally consistent and commercially wrong. Before acting on any AI budget advice, ask which attribution rule it uses and over what time window. Then ask whether it separates new customers from returning ones, and what it leaves out entirely.
Media buying judgement depends on facts that sit outside the ad accounts, and no model takes them into account unless someone feeds them in. A recommendation can be perfectly logical within the data and still wrong for the business, because the deciding fact was never in the data. These are the facts it most often misses.
The practical answer is a one-page context sheet: margins by product or service, stock and capacity limits, known events and seasonal patterns. Check every AI budget recommendation against it before you act. It takes ten minutes and catches the mistakes that cost most.
The ad platforms' own recommendations deserve particular care. Some are plainly useful, such as fixing disapproved ads, repairing tracking or adding missing ad formats. Others mainly raise spend or loosen targeting, such as increasing budgets, broadening keyword match types or adding placements.
These suggestions are written by the company that sells the advertising, which does not make them wrong, but it does mean you should judge each one against your own numbers. Apply the ones you can justify, dismiss the rest with a reason, and do not treat the account score shown next to them as a measure of your results.
Third-party optimisation tools need different questions. What data can the tool actually see, which goal is it optimising for, and can it show why it recommends a change? If a tool cannot explain its reasoning in terms you can check, keep it in an advisory role.
General AI assistants are useful for exploring an export, but their mistakes look convincing: they misread columns, mix up totals and occasionally invent figures. Ask for the calculation behind any number, check totals against the platform, and never upload files containing customers' personal data, which brings obligations under the GDPR, the EU's data protection regulation.
The question every owner wants answered is a scenario: if we move €5,000 a month from Meta to LinkedIn, what happens to enquiries over the next three months? Most AI tools cannot answer that reliably, because each platform's data describes what happened inside that platform, not what would happen if money moved between them.
Marketing mix modelling (MMM) is the established method for this. It estimates each channel's contribution from your history of spend and sales, including effects that appear weeks later. Google and Meta have both released free, open-source MMM software, Meridian and Robyn respectively. AI assistants built on such models let you ask questions in plain language, but the answers are only as good as the model underneath.
A trustworthy model needs a long and varied history, typically two years or more of weekly data with real changes in spend, which many growing businesses do not yet have. A simpler test often answers the immediate question better: halve or pause one channel in one region or for a fixed period, keep everything else the same, and compare the results with the regions or weeks where nothing changed.
Whatever the method, insist on a range rather than a single figure. 'Between 20 and 45 extra enquiries a month' is honest, while 'exactly 32' is false precision that makes a decision feel more certain than it is.
AI changes what you know when you make a decision, but the decision is still yours.