The AI infrastructure brands actually need

When I was working at MediaCom in the 2010s, something important was happening beyond the agency world. Large global clients were investing heavily in their own cloud data systems. These were not simply better places to put the numbers for an enterprise BI dashboard. Increasingly, they were foundations for marketing too, because understanding the customer and understanding the business were becoming inseparable.

That observation helped lead to Canton Marketing Solutions , the business I founded with Nick King . We saw an opportunity to work alongside clients, connecting specialist marketing knowledge with the technology they were investing in. They needed people who could understand both the business ambition and the practical work required to deliver it. Canton was subsequently acquired by Goodway Group and became part of its consulting business.

There was a memorable Canton pitch where a blue-chip enterprise challenged whether we were big enough to do the job. Our response was that we had more expertise in the specific marketing technologies and disciplines they needed than the giant consultancies we were competing against. We offered to take the Pepsi test: bring the relevant experts to the next pitch, and we would bring ours. The client was buying the people who could solve its problem, not everybody else on the supplier’s payroll. We won the business.

Pepsi is perhaps a cheeky reference here, because my team and I worked on the Coke pitch back around 2013. The world has come a long way since then. We can now put research, modelling and specialist tools within reach of a small team in ways that would have been extraordinary at the time. But access to those capabilities is not the same as knowing how to bring them together to help a brand win.

I think about that distinction a lot in the AI era. How many people who can build an AI agent can also plan a television campaign using GRPs, take meaningful control of Performance Max, or understand what a creative brief needs to change? Equally, how many experienced marketers can turn their expertise into a reliable system? Both sets of skills matter, and neither automatically supplies the other. Being enormous is not the same as having depth in the particular thing a client needs.

Look at where enormous business value has accumulated: Salesforce, Snowflake and Databricks. These are different businesses, but each has become part of how enterprises manage customer relationships, data or intelligence, rather than simply another tool for buying an impression. Databricks alone announced a $134 billion private valuation in February 2026. That is a scale of ambition our industry should pay attention to. My reading is that the prize is not just making a marketing task cheaper; it is becoming part of the capability on which the enterprise depends. Databricks financing announcement

This is not an argument that marketing and advertising are unimportant. Google and Meta would be rather strange businesses to build if that were true. It is an argument about which problem you solve, whom you solve it for, and how much of the resulting value you can capture. Salesforce is particularly instructive here: customer-facing technology can become an enterprise foundation, not merely a departmental add-on. Marketing does not end at the boundary of an ad platform, and its infrastructure should not either. Salesforce’s history

I wish more of the adtech and martech companies launched in the 2010s had built towards that opportunity. There was extraordinary expertise in those businesses, but I think too much of it went into optimising a narrow transaction rather than strengthening the enterprise behind it. What if more had helped brands connect their data, retain their knowledge and make better decisions across the whole business? That would not have guaranteed success, but it would have aimed their expertise at a larger and more durable problem. As we build for the AI era, I do not want us to repeat the mistake of confusing a useful component with the whole opportunity.

That is the lesson I would carry into AI adoption. Stay close to the client, work with the foundations it already has, and make the additional capability transparent and modular. The goal is to unite the CMO, CTO and CFO around a system that helps the business grow. The CMO needs more effective marketing, the CTO needs control and a maintainable architecture, and the CFO needs credible economics. Three separate sales pitches are not going to join those priorities together.

I find it helpful to think about the infrastructure in five layers. Each has a job, and none should be confused with the whole proposition. The point is not to make the diagram look sophisticated. It is to show how a business gets from access to intelligence to better work, with the right controls along the way.

The first layer is access to capable models. Brands do not need to build their own frontier model to benefit from one, and I would avoid designing the whole business around a single provider. Different tasks will have different requirements, and those requirements will change. Buy access intelligently, agree the terms that matter, and retain the ability to change components without rebuilding the operation.

The second is routing that work sensibly. Some jobs justify an expensive reasoning model, while others belong with a smaller model, an appropriately hosted open model, or ordinary code. The measure should be the cost of a usable result, including checking and correction, rather than the cheapest token. A gateway can help manage that routing, but owning the gateway does not mean you understand the work travelling through it. That is where infrastructure discussions too often stop.

The third layer is the client-controlled operating foundation: its evidence, definitions, memory, plans, permissions and audit trail. This should fit the client’s own cloud or on-premises environment, rather than require it to surrender everything to another black box. Client hosting does not automatically mean that nothing leaves the environment; any external model processing still needs explicit controls over what is sent, where and under which terms. The important principle is that the client’s accumulated knowledge is not trapped inside a disposable conversation with a rented model.

The fourth layer is connections, in both directions. The system needs to read relevant research, customer information and performance data, and connect to the platforms where marketing happens. It also needs to understand what the information means, because a television audience, a search query and a CRM segment are not interchangeable objects. Sending an audience definition to a platform is only useful if the translation preserves the intent. Reading results back should then change the next decision, rather than simply fill another dashboard.

The fifth layer is specialist capability, and this is where the Canton story becomes particularly relevant. Generative AI can help interpret briefs, synthesise evidence and develop messages, while deterministic code handles calculations and rules, and statistical or machine-learning models support forecasting where they have been properly validated. These are different tools with different responsibilities, not one chatbot wearing several job titles. A forecasting model can be wrong, and an elegant explanation does not make it right. The expertise belongs in the definitions, methods, tests and judgement that connect those components to the actual marketing problem.

Consider a brand deciding whether to move money between broadcast television, CTV, paid social and search. That decision affects reach, frequency, creative requirements, demand capture and the way results should be measured. It needs commercial context as well as media data, and it may reveal that the creative or customer experience matters more than another budget adjustment. The plan should bring those dependencies together and make the choices inspectable. This is why I see the plan as the control panel: it turns the foundations into coordinated decisions and action.

That is the model we are building towards at TAU across planning, activation, measurement and intelligence. The client controls its environment and its information, while reusable software, connectors and methods continue to improve. Development can compound across engagements without pooling confidential client data. Updates need testing, versioning and an agreed route into the client’s operation, so a shared core does not become an uncontrolled dependency. The ambition is to combine the benefit of a developing product with the fit of a system that understands the individual business.

None of this requires every enterprise system to be perfect before useful work begins. Start with an important question, establish what evidence and permissions it needs, and deliver a bounded improvement that can be evaluated. Then connect that improvement to the next one, instead of accumulating another collection of isolated pilots. Time savings matter, but the larger prize is better decisions, stronger performance and growth that the finance team can recognise.

When a supplier tells you it has the infrastructure for marketing AI, ask it to demonstrate the marketing. Give it an awkward brief, a measurement disagreement or a plan whose numbers do not add up, and see what happens. Ask who understands the answer, how it is checked and how the client stays in control. I still rather like the Pepsi test: bring the people and the system that can do the work, and let the work decide.