Your team can run its own campaigns now. Here's the system that makes it work.

Every e-commerce team I work with reaches the same question eventually: how much of this could we do ourselves?

Until recently the honest answer was less than you’d hope. Running Google Ads and Meta well needed someone who lived in those interfaces daily, and the two tasks that ate most of the hours — assembling the data and interpreting it — were genuinely hard to hand to software.

That part has changed. Not because the platforms got simpler — they didn’t — but because reporting and routine interpretation are now things a well-configured AI system does reliably. What’s left is judgement, and judgement is needed far less often than reporting is.

This is what I set up inside companies, and what I train their teams to run.

What the system actually is

It is not a chatbot bolted onto your ad account. It’s four things wired together:

A data layer. Google Ads, Meta, GA4 and your commerce backend pulled into one place, on a schedule, with the metrics that matter to your business defined once and centrally. Contribution margin, not revenue. New customers, not conversions.

This is the part that decides whether the whole thing works, and it’s the part most people underestimate. Benjamin Wenner made the argument well in Search Engine Land: an agent optimising on platform-native data alone forms a closed loop that improves metrics while the business gets worse, because “Google Ads doesn’t know your average deal size, sales cycle length, or cash position this month”. His conclusion matches mine exactly — building the agent is the easy part; connecting it to reality is the work.

An AI interface on top of it. Your team asks questions in plain language and gets answers grounded in that data: Which campaigns lost money last week after returns? Why did CPA rise in the Netherlands? Show me every product group where ad spend exceeds gross margin. No SQL, no pivot tables, no waiting on someone else’s Tuesday report.

Guardrails. Defined thresholds for what counts as normal, what’s worth flagging, and what must never change without a human deciding. Most people skip this step, and it’s the one that determines whether the system is useful or dangerous.

A routine. A weekly rhythm the team actually follows: the system prepares, a human reviews, decisions get logged. Without a routine you have a very expensive search box.

What it genuinely handles

Reporting, completely. Any report, any cut, any time, in the format the person asking needs. This alone usually justifies the setup — most companies’ reporting cost is measured in person-days per month, and it largely disappears.

Anomaly detection. The system watches every campaign, every product group, every geography, every day. A person watches the ten things they remember to look at. When one product group quietly starts eating 30% of budget at a loss, the system sees it on day two.

Routine analysis. Search term review, placement review, audience overlap, creative fatigue, feed error detection. These are pattern-matching tasks over large tables — exactly what the technology is good at, and exactly what humans do irregularly because they’re tedious.

Explanation. Not just what moved, but what plausibly moved it, with the supporting numbers attached. Your team stops guessing and starts checking.

Where it fails — know this before you start

Anything requiring knowledge the data doesn’t contain. A competitor launched a price war. A supplier delay means your bestseller now ships in fourteen days. Legal wants a claim removed from your ad copy. The system sees the effect and will confidently attribute it to something else.

Strategic restructuring. Rebuilding an account around margin instead of revenue, splitting brand from generic demand, changing what a campaign is for — these are judgement calls with second-order consequences, and AI is a poor judge of consequences it hasn’t been shown.

Novel situations. A new platform feature, an unusual auction dynamic, a tracking break that mimics a performance drop. The system is trained on the ordinary; the expensive problems are extraordinary.

Anything where being confidently wrong is costly. Which is why the guardrails matter more than the capability.

Every one of these is a case where a specialist earns their fee. The point isn’t to remove expertise from the picture — it’s to aim it at the questions that actually need it.

What you need in place first

Be honest about these before investing anything:

Tracking that matches your accounting. If platform-reported revenue and actual revenue diverge by more than a few percent, fix that first. A system built on broken tracking produces fluent, well-formatted nonsense.
Margin data at product level. Optimising to revenue is why many accounts look successful and lose money. Without margin in the data layer, you get a faster version of the same mistake.
One person who owns it. Not a committee. Someone whose job includes the weekly review and who can act on it.
Realistic expectations. This runs business-as-usual. It does not replace strategy, and the first month is about building trust in the numbers, not about savings.

What actually changes

Three things, consistently.

Reporting stops being a task. Nobody assembles a monthly deck; anyone who wants a number asks for it. That time goes back into the business.

Problems surface days earlier — not because anyone got smarter, but because everything is watched instead of the ten things someone remembered to check.

And the conversation with outside specialists changes character. The hour you buy goes into deciding what to do rather than into establishing what happened, which is a better use of it on both sides.

This is also, broadly, where the industry expects things to go: the 2026 trend discussions in Search Engine Land land on the same combination of rapid AI advancement paired with strong fundamentals and human oversight, with measurement quality — not model quality — as the limiting factor.

I set these systems up and train teams to run them, as remote workshops or as a programme spread over several weeks. If you want to know whether your account and your data are in a state where this would work, a short call will tell you quickly.