Forty-nine percent. That was the share of marketing leaders who said they could prove ROI from their AI investments last year. A number that, frankly, already felt generous — I know what "proving ROI" looks like inside a large organization, and it usually means a favorable PowerPoint slide, not an airtight attribution model.
This year, that number dropped to 41%.
Let that register. AI adoption across marketing teams has never been higher. Budgets are up. Headcount around AI initiatives is growing. Vendor pitches have never been more polished. And yet, fewer marketing leaders can connect the money going in to the results coming out. In retail — arguably the vertical with the most mature performance marketing infrastructure — the decline was worse: 54% down to 38%.
I've been turning this over for weeks, and I think the diagnosis is simpler than most people want to admit. We didn't have a measurement problem before AI, and then AI made it harder. We had a measurement problem before AI, and then AI made it invisible.
The Activity Trap
Most marketing teams adopted AI the same way they adopt every new tool: they layered it on top of whatever they were already doing. The AI writes more emails. Generates more ad variants. Segments audiences faster. Produces content at a pace that would have been unthinkable two years ago.
And all of that activity produces metrics. Open rates. Click-through rates. Impressions. Engagement scores. The dashboards have never looked greener.
But here's what I keep seeing when I talk to marketing ops leaders: the activity metrics improved, and the business metrics didn't follow. Or they did for a quarter, and then flatlined. Or they improved on paper, but finance looked at the same period and saw a different story entirely.
I've run into a version of this problem before, well before AI was involved. I once managed a portfolio where several channels looked incredible on engagement metrics — high click-through, strong response rates — but when we built the payback models to track all the way through to actual conversion and retention, some of those "winning" channels were underwater. The cost-per-acquisition looked fine. The cost-per-retained-customer was a disaster.
AI has made this exact dynamic worse, because it excels at optimizing the front of the funnel while being completely blind to what happens afterward.
Why Your Attribution Model Can't Handle AI
The attribution models most marketing teams rely on were built for a world where humans did things in a somewhat predictable sequence. A person saw an ad, visited a site, filled out a form, got nurtured by email, and eventually converted. The models — whether first-touch, last-touch, or multi-touch — all assumed a traceable chain of human actions.
AI breaks this in two ways that aren't getting enough attention.
First, AI-generated content and AI-optimized campaigns create so much surface area that attribution becomes noise. When you're running 200 ad variants instead of 20, and your AI is dynamically adjusting audience segments in real time, your multi-touch model isn't capturing a journey anymore. It's capturing chaos. The model assigns credit to touchpoints that the AI itself created and optimized — a closed loop that tells you nothing about incremental value.
Second — and this is the one that keeps me up at night — your buyer's AI is now involved too. A growing share of consumer and purchase research is happening inside ChatGPT, Perplexity, and Gemini. Those interactions leave no cookie. They generate no pixel fire. They don't show up in your CRM. ChannelEngine reported that AI agent traffic to retail sites jumped 1,300% year-over-year. Amazon's Rufus influenced 66% of Black Friday purchases. Your attribution model has a massive blind spot, and the blind spot is getting bigger every month.
The Measurement Infrastructure Nobody Built
Here's where I'll say something that might be unpopular: most marketing teams that adopted AI over the past 18 months weren't ready for it. Not because they lacked the technical sophistication to use the tools — the tools are easy. They weren't ready because their measurement infrastructure was already held together with spreadsheets and good intentions.
I've seen this pattern over and over again. The stack looks impressive. Salesforce or HubSpot for CRM. An attribution platform. A CDP, maybe. A BI tool. Each one generating reports. But the reports don't agree with each other. Nobody has a single source of truth for what a "conversion" means. And the data team that could build one is either understaffed or answering ad hoc requests from six different stakeholders.
Then someone adds AI to this foundation and expects to measure its impact. Good luck.
The organizations I've seen actually crack AI measurement — and there aren't many — share a trait that has nothing to do with technology. They decided, before deploying AI, what question they were trying to answer. Not "did AI make us more efficient?" but something specific: "Did AI-optimized audience segmentation reduce our cost-per-enrollment in the Southeast region during Q3?" That specificity forced them to build the measurement infrastructure first. Everyone else did it backwards.
What Actually Works
I'm not going to pretend I have this fully figured out. But four patterns keep showing up in the teams that are doing this better than most.
Incrementality testing over attribution modeling. The teams showing real AI ROI have largely abandoned multi-touch attribution for AI-driven campaigns and replaced it with holdout-based incrementality tests. You take a population, exclude a random subset from the AI-optimized treatment, and measure the difference in business outcomes — not clicks, not engagement, but revenue or pipeline or whatever your actual KPI is. It's not new science. It's how direct mail has been measured for decades. But somehow, when AI got involved, everyone forgot the basics.
Leading indicators with payback windows. Earlier in my career, I developed a real-time leading-indicator process that let us optimize mid-campaign without waiting for final conversion numbers. The same principle applies to AI: find the early signals that correlate with eventual business outcomes, validate the correlation rigorously, and then — and only then — use those leading indicators for AI optimization. Most teams skip the validation step and let AI optimize toward signals that have never been proven to predict anything that matters.
Separating "AI made us faster" from "AI made us better." Speed and quality are different things, and conflating them is how most AI ROI claims fall apart under scrutiny. Yes, your team produced 10x more content. Did any of it outperform what you were producing before? Prove it. With holdout data, not anecdotes.
Where This Goes
Gartner is now predicting that over 40% of agentic AI projects will be canceled by the end of 2027, primarily because business cases never solidified. That's not a failure of AI. That's a failure of the organizations deploying it to build the measurement discipline the technology requires.
The marketing teams that figure out AI measurement first won't just be able to justify their spend. They'll be able to compound their advantage — because they'll know what's working, double down on it, and cut what isn't, while their competitors are still arguing about dashboards.
That's not a prediction. It's math.