The AI marketing stack is not a slide deck. Here's what it actually takes to operate.

The Forrester thesis in 90 seconds

Forrester Research has been making a version of this argument for roughly two years: the CMOs who survive the current transformation are the ones who have operationalized AI across the entire marketing function. Not as a collection of point tools. As an integrated operating model. They call this the AI CMO framework, and the thesis is that marketing leadership is bifurcating. Executives who have built and operate the AI stack at enterprise scale are on one side. Everyone else is on the other.

I agree with the destination. The Forrester framework is describing something real and consequential. The issue is the execution gap.

The framework describes where you need to end up. It does not adequately describe what it costs to get there, how long it takes, or what goes wrong along the way. Most of the marketing leadership teams I have encountered are underestimating that distance by a significant margin. They are treating the AI CMO transformation as a technology procurement project. It is not. It is an organizational capability build, and the two problems have different timelines, different budgets, and different failure modes.

This piece is about the actual build.

What “built the stack” actually means

When I say I built the AI marketing stack at Kyndryl, I am not describing a tool adoption story. I am describing five distinct capability builds that had to work in sequence before any of them produced meaningful results at scale.

The customer data platform. At Kyndryl we stood up a CDP from scratch. No inherited data, no existing taxonomy, no historical behavioral signal to build on. The technology procurement was the easy part. The real work was in the architecture: defining the customer object, building the identity resolution logic, establishing the data governance framework that would let us trust what the platform was telling us downstream. This work takes substantially longer than vendors say it will. Most organizations underestimate it by a factor of three.

The attribution layer. Enterprise B2B buying cycles are long, multi-touch, and poorly served by off-the-shelf attribution models. We built a custom multi-touch attribution framework that could account for the specific complexity of Kyndryl’s sales motion: large deal sizes, sales cycles measured in months, multiple stakeholders per account, and meaningful overlap between marketing-sourced and sales-sourced pipeline. The CFO used this model. That is the test that matters. If the finance organization does not rely on the model, it is not a commercial tool. It is a marketing exercise.

Content production infrastructure. We deployed ChatGPT Enterprise, Jasper, and Claude in production. Not in a pilot. In the actual production workflow, at scale, across multiple global markets. The technology is the easy part. The work that made it produce results was the brief template architecture, the brand guardrail framework, the review and approval process, and the feedback loops that improved output quality over time. We built all of that. It took roughly two years to get it right.

Campaign orchestration. The orchestration platform is only as good as the data flowing into it. This is where most organizations hit the wall. They purchase the platform, connect the obvious data sources, and then discover that the underlying data model is not clean enough to support the personalization logic they wanted to run. The fix is usually several months of unglamorous data work that was not in the original project budget and cannot be put on the transformation roadmap without explaining why the foundation was not built first.

Agentic workflows. We built production agents in Microsoft Copilot Studio to handle specific, bounded marketing operations tasks: campaign brief generation from structured inputs, performance anomaly detection against daily reporting baselines, asset trafficking automation, and account-level reporting summaries. Each agent was designed to solve one problem well. The temptation is to build agents that do many things adequately. The ones that held up in production did one thing well.

These five layers are not a shopping list. They are a dependency graph. The CDP enables attribution. Attribution enables intelligent campaign orchestration. Orchestration enables the agents to make decisions that actually improve outcomes. Building them out of sequence produces systems that look complete on an architecture diagram and do not work when someone tries to use them.

The three failure modes

In watching organizations attempt this build across different industries and different starting points, the failures sort into three consistent patterns.

Tool-first. The organization purchases a set of AI marketing tools in response to competitive pressure or board expectation. The tools are deployed. The transformation is declared. Six months later, the tools are not integrated, the data is siloed in each platform, and the CMO cannot produce a coherent picture of what the investment is generating. This is the most common failure mode, and it is expensive because the tool budget is committed before the architectural questions are answered. The tools are not the problem. The absent architecture is.

Slide-first. The CMO develops a compelling AI transformation roadmap and presents it to the board. The board approves the budget. Then execution starts, and within ninety days it becomes clear that the internal team does not have the technical depth to build what the roadmap describes. The tools are purchased, the headcount approvals are in, but the organizational capability to architect and implement the actual stack is not there. The roadmap was correct. The team cannot build it. This is the failure mode that most visibly damages CMO credibility, because the gap between what was committed to the board and what is actually shipping is not subtle.

Hire-first. The organization hires a VP of Marketing Technology, or a Head of AI Marketing, and expects that hire to solve the problem. They cannot, alone. The AI marketing stack is an organizational capability problem, not a staffing problem. One hire addresses a gap in the org chart. It does not address the architectural gap, the data quality gap, the process redesign gap, or the budget alignment gap that comes from years of underinvesting in marketing infrastructure. The new leader becomes the scapegoat for problems that were never theirs to solve.

What operating the stack actually looks like

Building the stack is a multi-year capital project. Operating it is a different discipline, and the distinction matters.

Operating the AI marketing stack at scale means running a function that includes production systems with uptime requirements, data pipelines with latency and quality SLAs, agentic workflows that need monitoring and maintenance, and attribution models that need recalibration as the business changes.

None of that is traditional marketing work. Most marketing leaders built their careers managing campaigns, agencies, and brand strategy. Managing infrastructure is a different skill set, and the organizations that resist staffing it will find the AI stack becoming a liability rather than a competitive advantage. The tools will drift. The data will degrade. The agents will produce results no one trusts.

The executives I have seen operate this well treat the stack the way an engineering organization treats its production systems. There are owners. There are SLAs. There is a release process for changes. When something breaks, there is a postmortem. This is not common in marketing organizations. It needs to become common.

What to do in your first 90 days

If you are inheriting this problem as a new marketing leader, the most useful work you can do in the first ninety days is an honest audit of where the data layer actually is. Not where the roadmap says it should be. Where it actually is.

That means pulling up the CDP and asking specific questions. How complete is the customer object? What percentage of accounts in the named universe have meaningful behavioral data attached? How is identity resolved across channels and systems? What is the data freshness at each point in the pipeline?

It means looking at the attribution model and asking one question: does the CFO use this? If the answer is no, or qualified, you have a trust gap. That trust gap is your problem before it is a data quality problem. Rebuilding finance’s trust in marketing’s commercial numbers is harder and slower than fixing the underlying model, and it starts on day one.

It means walking through every agentic workflow in the existing stack and asking what it actually handles, what the error rate is, and who owns it when it breaks.

What you find in this audit tells you where to spend the first budget. Organizations with a solid data layer and immature tooling can move fast. Organizations with impressive tooling sitting on a broken data foundation cannot, and will remain stuck until the foundation work is done. Ninety days of structured diagnosis is not delay. It is the fastest route to a stack that actually compounds.

Why the talent pool for this work is structurally small

The person you need to build and operate this stack has to hold five things simultaneously: marketing strategy, marketing technology, data architecture, AI deployment, and organizational change management at enterprise scale.

Each of those is a career-long skill set. The executives who have developed all five did not arrive there through a standard CMO career path. They got there because they were consistently in situations that required technical depth that most marketing executives avoid, and they pushed into that depth rather than delegating it away.

This combination is not common, and it is not growing more common at the rate that demand for it is growing. The programs that produce senior marketing leaders are not producing graduates who can architect a CDP or design an agentic workflow. The technology track that builds those skills typically does not produce the strategic and organizational depth that enterprise CMO roles require.

Boards looking for someone who can operationalize the Forrester AI CMO framework should be specific about what they are actually asking for. The description that circulates on a job posting will be answered by candidates with a compelling narrative about AI marketing transformation. The candidates who have actually built the stack know what they built, and they are not the same candidates.

The slide and the operating system are not the same thing. Closing that gap is the whole job.