By Tobias Cummins, Chief Operating Officer, Pencil
The word “agentic” is now attached to every marketing product launched in the last six months. Ad platforms, creative tools, CDPs, email systems, analytics dashboards. Everything is agentic, which is another way of saying nothing is.
The failure nobody puts in the slideware
MIT’s 2025 State of AI in Business report found that 95% of enterprise generative AI pilots fail to deliver measurable P&L impact. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027. S&P Global reported that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. The average sunk cost per abandoned initiative was $7.2 million.
This is not a “they don’t understand AI” problem. These are Fortune 500 companies with good people, good vendors, and real budgets. They run a pilot. The pilot works. They try to scale it. It falls apart.
Not because the model fails. The models are fine. Most of them are brilliant at the narrow task you show them in the demo. They fail because the model was never the thing that needed to scale. The thing that needed to scale was the system around the model. How work enters it, and routes to the right capability. How the results are reviewed and measured. How the model is governed when twelve brand managers in nine markets all want slightly different things.
That’s the thing nobody sold you. These problems aren’t addressed in the pamphlets. But if marketers are to be convinced by these tools, they need to see how they work in practice. That’s what an agentic marketing OS is supposed to be.
You’ve been looking at the wrong layer
For the last decade, enterprise marketing has been buying apps. Creative apps. Personalisation apps. Testing apps. Analytics apps. Measurement apps. Each one with its own login, its own data model, its own billing, its own vendor relationship, its own quarterly roadmap.
The average enterprise marketing stack now contains more than 14,000 martech solutions to choose from and most enterprises run 90 to 120 of them in parallel. Most of them don’t talk to each other. The “integration” is a spreadsheet on a marketing ops person’s laptop.
Bolting generative AI onto that stack does not produce agentic marketing. It produces faster chaos.
Marketers don’t need another model; they need a coherent system to put them in.
Think of it like your laptop’s Operating System. An OS is the layer that makes every application on top of it work together, follow the same rules, and draw from the same memory. Your computer doesn’t have fourteen different clipboards. It has one. That’s the OS.
The model may perform exceptionally at its given task, but if there is not a clear way to integrate it with your existing marketing systems, it could end up sitting on the shelf. To allow agentic AI to streamline operations, don’t just look at what the model does, think about how it fits.
The three questions to ask any vendor
If you are a CMO or a head of marketing operations reading this, you don’t need another demo. You need three questions you can ask any vendor selling an agentic model, to test if it will pass real use cases.
1. Show me your evals. Not “our AI is accurate.” Show me the pass/fail rate on a standard set of brand-safe marketing tasks. If they don’t have one, they likely haven’t considered how it integrates with standard marketing processes.
2. Which models do you run? If the answer is one, you’re buying a closed-off system tied to one provider’s AI model. That’s a defensible choice if you know it’s what you’re choosing. Most CMOs don’t realise they’re choosing it.
3. What’s your memory model? Ask how the system remembers what your brand team approved six months ago in Spain. If the answer is “the user can upload guidelines again,” there is no memory. That’s just a very fast intern with amnesia.
These three questions can help you evaluate whether a tool will fit smoothly within your marketing team, and whether it has flexibility built-in to evolve with your business.
Why this matters now
Marketing is being asked to produce roughly five times more content by 2026 than it did five years ago, on roughly flat budgets. CMOs cannot afford to spend money on models that end up collecting dust because teams cannot cleanly integrate them into their workflows.
Buyers need to look past the promises attached to the latest ‘agentic’ marketing products. The word agentic describes the difference between a system that executes instructions, and a system that reasons about a goal, routes it, evaluates the output, and improves.
Most of what’s being sold as agentic right now is the first thing with the second thing’s marketing.
The CMOs that figure out the difference and pick models to fit their system will spend the rest of the decade compounding. The teams that don’t will become the next batch of abandoned-AI stats in next year’s BCG report. All it takes is asking the right questions.
