Stop plugging AI straight into your ad platforms

Connect Claude to Google Ads or Meta and within a few minutes you can ask about spend and pacing in plain English. For a channel manager checking whether a campaign is on track this afternoon, that's a real step forward, and a connector that can pause a broken ad set without anyone logging in is useful too. But a lot of marketers have taken it further, and the line doing the rounds is that you can connect Claude to Meta and skip the data team. That stops being true the moment you add numbers up across channels or check them against a CRM. Every platform takes credit for the same sale, so an agent wired straight into them starts producing numbers nobody should put in front of finance.

MCP, the standard that lets AI tools plug into other software, is quickly becoming how AI reaches every business system, and it belongs in any sensible marketing stack. What I object to is an agent working on raw platform data with nothing in between to tell it what your metrics mean or what it's allowed to change. Anything that ends up in a board pack or a CAC conversation should come from a governed warehouse with a marketing knowledge layer on top, because every agent you add is another one deciding for itself what your numbers mean.

Where does a pure MCP setup fall apart?

Ask an agent connected to Google Ads, Meta and LinkedIn for last quarter's blended cost per lead. First it has to decide how to match campaigns across three naming conventions, which conversions count as leads, and how to handle currencies and overlapping users. It makes those calls silently, and may make different ones tomorrow, or for a colleague who asks an hour later.

Even perfect choices wouldn't help, because the figures were never meant to be added together. Meta counts a conversion within 7 days of a click or 1 day of a view, Google Ads within 30 days of a click, and LinkedIn within 90 days of a click or a view. Each platform marks its own homework. When all three claim the same customer, the channel total runs well above anything your CRM recognizes, and finance usually notices before marketing does.

Revenue is worse, because the ad platforms have no idea what happened after the lead arrived. A search campaign can log a steady stream of form fills and downloads and close almost nothing. Add the CRM and an agent can spot the gap, but it still has to improvise how clicks match to deals, which touch gets the credit and what counts as revenue. Next month it may decide differently, with no record of last time, and I wouldn't move a line of budget on that.

History is patchier than it looks, with Meta's Insights API stopping at 37 months, its unique-reach and hourly breakdowns going back only 13, and LinkedIn keeping demographic data for two years. Anything older is gone unless someone stored it. Large pulls also hit rate limits and context windows, and the agent won't always say the results were cut short.

Credentials sit with whoever set up the connector, and there's rarely a record of how a figure was produced. Add write access to live ad accounts and your CISO will shut it down the moment they hear about it.

Why fix it once, where the data lands?

A governed warehouse deals with these problems as the data lands, so nobody has to solve them again in every conversation with an agent. Campaign names and currencies are standardized on the way in. Attribution follows one set of rules across channels, with conversions deduplicated before anyone counts them.

Ad, web and CRM data sit side by side, which turns pipeline and CAC from estimates into calculations. History goes back as far as you've kept it, access is managed in one place, and every figure can be traced to where it came from.

When the CMO and the CFO ask the same question they should get the same number, and either of them should be able to see how it was worked out.

What does skipping the warehouse really cost?

The usual pushback is that a warehouse is expensive to run and its data is never fully live. Both points assume the alternative comes free.

Any team running paid media across several channels and markets, with finance checking the numbers, is already paying for not having one, in the days spent every quarter reconciling reports before the QBR and in budget moved on numbers that later turn out to be wrong.

As for freshness, nobody sets quarterly allocation or channel mix on the last hour's data. Daily is plenty for the decisions that move serious money. The work that really does need live data is operational, and that's the job I'd leave to connectors.

What can't a warehouse tell an agent?

Point an agent at raw warehouse tables and it still has to guess what counts as revenue or a qualified lead. It can do the arithmetic perfectly and still reach the wrong conclusion, because it doesn't know that ad platforms over-claim or that you ran this exact test last spring and it flopped. That's a confidently wrong answer, and it's much harder to catch than a wrong number.

A marketing knowledge layer fills that gap by sitting between your AI and your data. It holds one agreed definition for each metric, along with your goals, the marketing know-how to explain why a number moved and a memory of what you've already tried. Ask an agent for blended CAC and you get the company's definition with the working shown, whoever is asking.

In the stack I'd build, MCP sits at both ends. Agents use it to reach the knowledge layer, and their decisions go back out through it to the ad platforms. If the analysis shows a campaign isn't producing pipeline, an agent can cut its spend in seconds, inside limits you've set on what it can change without sign-off, such as a cap on daily budget moves, with anything bigger going to a person to approve.

Over the next year plenty of marketing teams will put agents straight onto their ad accounts. Sooner or later one of those agents will hand someone a confident CAC figure that falls apart the moment finance asks how it was calculated, and from then on the warehouse won't look optional. It's a lot cheaper to build it before that meeting than after.

We built Adverity Atlas for exactly this: a marketing knowledge layer that sits between your AI and the warehouse you already have, so every agent and every report works from the same definitions and context.