Where are the AI marketing case studies with real numbers? In fewer places than the conference circuit suggests. Most "AI success stories" count outputs: assets generated, hours saved in theory, tools adopted. The few worth learning from name a specific inefficiency, measure it, explain the method, and say honestly whether the result has happened yet.
The freshest example landed this week. Campaign Middle East reported that performance agency TRKKN used Google's open-source Meridian marketing mix model to show Faces, the beauty retailer, where its revenue in Saudi Arabia really comes from. Three years of data, 11 channels, and a projected 8.5% incremental revenue lift plus a 10.8-percentage-point ROI improvement, simply from moving budget Faces was already spending.
What does the Faces and TRKKN case actually prove?
Something specific and useful. TRKKN deployed Meridian inside Faces' existing Google Cloud environment, modelled online and offline marketing and revenue data from 2023 to 2025, and built response and saturation curves for each channel. Performance Max came out as the strongest channel. The model also showed where extra spend would run into diminishing returns.
Harry Lonergan, digital marketing director at Faces, part of Chalhoub Group, put the value in one line: "Meridian has helped us move on from what we see in last click performance."
Two caveats. First, as the report makes clear, the 8.5% and the 10.8 points come from modelled scenarios, not post-implementation results. Second, this is statistical modelling rather than generative AI, so it tells you nothing about content tools. Neither point weakens the case. Both make it credible, because you know exactly what you are looking at.
Why are most AI marketing case studies so thin?
Economists have a name for markets where prices stop sending the right signals: market failure. One classic cause is information asymmetry, when one side of a deal knows far more than the other. Advertising has lived with it for decades. Platforms mark their own homework. Agencies bill hours for work clients never see. Clients judge channels on last-click numbers that flatter whatever happened to sit nearest the sale.
AI case studies inherit the problem. Vendors publish the wins, rarely the method and almost never the baseline. A line like "saved 40% of production time" sounds like evidence until you ask: 40% of what, measured how, and what happened to the work?
The useful cases do the opposite. They attack an inefficiency the market has tolerated, such as last-click measurement or days of manual reporting, and they show the number. Efficiency, in this sense, really is the mother of innovation: you cannot reinvent a process until you can see what it costs.
Which AI case studies have numbers you can check?
Here is what we can point to today, including our own clients, judged by the same test. Where there is no published figure, we say so.
| Case | Inefficiency tackled | The number | Projected or realised? | What is missing |
|---|---|---|---|---|
| Faces and TRKKN (Meridian MMM) | A last-click view of 11 channels | 8.5% incremental revenue lift; 10.8-point ROI improvement | Projected from three years of data | Post-implementation results |
| SAMY with Defy Agents | Manual performance reporting | Three days of work cut to one | Realised in the workflow | The commercial value of the freed time |
| Enchant with Defy Monitor | Manual CRM insight and campaign deployment | No published figure | Adopted for all its clients | A quantified outcome |
| SharkNinja with Defy Monitor | Understanding LATAM markets | No published figure | Used to build strategies to test and learn quickly | The results of those tests |
| Sprint agencies (Beers With Friends, X&O) | Months-long projects and retainers | Five-day sprints; clients return several times a year | Operating model in use | AI plays no part in the story |
What does "three days to one" look like inside an agency?
SAMY, a global influencer agency, adopted Defy Agents to streamline performance reporting, cutting three days of work to one. It is not a glamorous number. It is the kind that matters, because reporting is the chore every account team knows: pulling numbers from platforms, reconciling them, formatting the deck, checking it, sending it, explaining it.
Agentic workflows on the Defy platform take that kind of repetitive delivery off people's desks. The interesting question is what the other two days become. If they vanish into more of the same, the client sees nothing. If they go into analysis, ideas and conversation, the report stops being a cost and becomes a reason to renew.
Campaign US's breakdown of the agency sprint model shows where speed leads commercially. Beers With Friends works in five-day sprints. X&O's co-founder Brett Banker says the agency has "zero client retention", because clients come back several times a year instead of signing agency-of-record contracts. The piece does not mention AI at all. That is the point: clients already pay for speed and focus. Faster delivery simply makes that model available to more agencies.
How to judge an AI marketing case study before you copy it
- Name the inefficiency. What was wasteful or invisible before? "We adopted AI" is not an inefficiency.
- Find the baseline. Three days of reporting, last-click attribution, a project measured in months. No baseline, no case study.
- Check the method. Which data, over what period, across how many channels? The Faces case answers all three.
- Separate projected from realised. A modelled uplift is a forecast. Treat it as a hypothesis to test, not a result to quote in a pitch.
- Ask where the saved time went. Efficiency only counts if it shows up as better work, faster decisions or lower cost to the client.
- Run your own pilot. One client, one workflow, one month, with the baseline written down before you start.
Where is the evidence still thin?
Mostly where the claims are loudest. Measurement and reporting produce clean before-and-after numbers, so they make good case studies. Generative content is harder: producing more assets is easy to count, while proving those assets sold more is not. Be wary of any case study that stops at volume.
Our own table has gaps too. Enchant and SharkNinja are real deployments without published figures, and SAMY's number measures time, not revenue. We would rather show the gaps than paper over them. The agencies that do well from here will publish their method alongside their wins, because that is the only kind of case study a sceptical client finance director will read twice.
