I’ve spent fifteen years in marketing measurement. Attribution models, incrementality tests, MMM builds, the lot. And if there’s one pattern that’s repeated at every single client I’ve worked with, from big four banks to FMCG giants, it’s this: the reporting stack gets more sophisticated every year, and the data underneath it barely improves at all.
Right now every marketing and analytics team in the country is having the same conversation. “How do we get AI into our reporting?” Boards are asking for it. CMOs are demanding a roadmap. Someone’s already trialled a co-pilot on the dashboard. And almost nobody in that conversation is asking the question that actually matters: is the data we’re feeding this thing any good?
It isn’t, by the way. In most organisations I’ve worked with, it isn’t even close.
The bit nobody wants to hear
Here’s the uncomfortable truth. An AI layer sitting on top of a broken measurement foundation doesn’t fix the foundation. It just gives you broken answers faster, with more confidence, and in a nicer chart.
I’ve sat in rooms where a business has three different attribution models running across three different platforms, none of them talking to each other, none of them agreed on by finance and marketing, and the answer to “which channel is actually working” changes depending on who’s presenting. Layer AI-generated commentary on top of that and you haven’t solved anything. You’ve just automated the confusion and made it sound authoritative.
This is the bit that gets skipped in every AI readiness deck I’ve seen. Everyone wants to talk about model selection, about which LLM, about the slick natural-language interface that lets an exec ask a chatbot “how did paid social perform last month.” Nobody wants to talk about the fact that paid social’s spend data hasn’t been reconciled with the platform’s own reporting in eighteen months, that the UTM taxonomy is inconsistent across three agencies, or that “last month” means something different in the CRM than it does in the ad platform.
AI doesn’t fix data governance. It exposes the lack of it, at speed.
What “fixing the data” actually means
This isn’t a call for another twelve-month data warehouse project before anyone’s allowed to touch AI. That’s the other extreme, and it’s just as unhelpful, because it gives the perfectionists in the room an excuse to stall everything indefinitely. What it means, practically, is:
- A single source of truth for spend and outcomes, agreed between marketing and finance before it ever hits a dashboard. If your MMM inputs and your board reporting don’t reconcile, nothing built on top of them will either.
- Consistent taxonomy across channels and platforms, so an AI layer isn’t guessing whether “Meta – Prospecting – Q3” and “FB_Prospect_Jul-Sep” are the same thing.
- Clear separation between correlation and causation in what’s actually feeding the model. Most reporting stacks I audit are still leaning on last-click or multi-touch attribution as gospel, with no incrementality testing to check whether any of it holds up. AI trained on a flawed attribution model just gets better at confidently repeating the flaw.
- Someone accountable for data quality, not as a side project bolted onto a BI analyst’s existing job, but as an actual owned function with the authority to say no to a launch until the tracking’s right.
None of that is glamorous. None of it makes a good slide for the board. But it’s the difference between an AI layer that genuinely accelerates decision-making and one that just accelerates bad decisions.
Why this is happening
Vendors love this moment. Every martech and analytics platform on the market has an AI feature ready to demo, and demos always look brilliant on clean sample data. Nobody’s demo runs on your actual, messy, half-reconciled dataset. So the buying decision gets made on the promise of the interface, not the integrity of what’s underneath it, and six months later someone’s asking why the AI-generated insights don’t match what finance is seeing.
There’s also a simpler, more human reason. Fixing the data is slow, unglamorous, cross-functional work that requires marketing, finance and IT to actually agree on things. Buying an AI tool is a much faster way to look like you’re moving.
Where I’d start if I were you
Before you sign off on any AI-enabled reporting layer, ask three questions. Can you trace every number in your current dashboard back to a source you’d trust in front of your CFO? Do marketing and finance agree on what “performance” means for each channel? And has anything in your current attribution approach actually been tested for incrementality, or is it just assumed?
If the honest answer to any of those is no, that’s your starting point. Not the AI layer. The data underneath it.
Get that right first and the AI part becomes genuinely useful, because it’s finally reasoning over something solid. Skip it, and you’ve just bought yourself a very expensive, very confident way of being wrong.