Who Owns the Brain?

Every major agency holding company now wants to be judged by its artificial intelligence operating system. WPP has WPP Open. Publicis has CoreAI. Omnicom has Omni. Dentsu has dentsu.Connect, relaunched in April 2026 as what the company called the “industry’s first truly agentic AI-powered operating system for modern marketing.” The language varies, but the message is almost identical. The future of marketing will be connected, predictive, automated, personalized, measurable, and increasingly executed through agents.

That is the visible race. The less discussed race sits underneath it.

Once every major holding company has an AI operating system, the strategic question is no longer whether the agency group has an interface, a workflow layer, an agent framework, or a dashboard with impressive verbs attached to it.

The question becomes more basic and more uncomfortable. Who owns the brain?

Not the data. Not the workflow. Not the client relationship. Not the orchestration layer. The actual model intelligence that interprets instructions, generates outputs, reasons across signals, proposes actions, and increasingly participates in execution.

On that question, the largest agency groups have taken different routes, but they appear to share one structural reality. None of them owns a frontier model. They are building data and orchestration moats around intelligence supplied by Google, OpenAI, Anthropic, Meta, Adobe, Amazon, and other major technology providers.

That does not make their AI strategies weak. In many ways, it explains why their strategies look the way they do.

If the cognition layer is rented, then the defensible layer has to sit somewhere else. It has to sit in data, identity, clean rooms, workflows, governance, client integration, vertical expertise, and the ability to turn generic model capacity into marketing-specific execution.

That is the real holdco AI race. The platforms are not only products. They are hedges against dependence.

The operating system became table stakes

The agency groups have spent the last several years trying to convince clients and investors that they are not being disrupted by AI but are becoming AI-native businesses themselves. The operating system has become the preferred proof point.

Dentsu’s latest version of dentsu.Connect makes the ambition explicit. The company describes the platform as a connected AI-powered system across creative, production, media, and experience. Its positioning leans heavily into agentic workflows, interoperability, cloud flexibility, proprietary models and agents, and the promise that data can remain inside the client’s own cloud environment.

The strategic message is not merely that dentsu has tools. It is that dentsu wants to become the connective operating layer through which modern marketing work moves.

WPP’s version of the argument is WPP Open and Open Intelligence. Under Cindy Rose, the company has framed AI, data, simplification, and technology partnerships as central to its turnaround. WPP describes Open Intelligence as a large marketing model connected to client brand data, WPP data, and partner data through a privacy-first approach. The company’s acquisition of InfoSum gave it a stronger data collaboration and clean-room foundation, while its expanded five-year Google partnership gives it privileged access to Google Cloud AI products and Google DeepMind models. WPP has also integrated Anthropic’s Claude models into WPP Open through Amazon Bedrock, which supports its claim that the stack remains model-flexible rather than locked to one provider.

Publicis has taken a different route. It has been the clearest of the large groups in building a data-first strategic narrative. CoreAI sits on top of Publicis’ proprietary data assets, including Epsilon, Marcel, Publicis Sapient’s transformation history, and now Lotame. The Lotame acquisition expanded Publicis’ identity footprint to nearly 4 billion profiles, a scale claim that matters because Publicis wants to compete not only as an agency group but as a data and personalization infrastructure company. But even Publicis, with its unusually strong vertical integration in data, does not own the foundation model layer. Its Adobe partnership integrates Firefly generative AI through CoreAI, and Publicis Sapient has also built a major Google Cloud partnership around generative AI deployment.

Omnicom’s path is scale and consolidation. The acquisition of Interpublic gave the combined company more data, more agency assets, more commerce capability, more production capacity, and more enterprise reach. Omnicom framed the transaction partly around enterprise generative AI capability, saying the combination provided scaled investment resources and access to partnerships with leading frontier model providers. Omni is positioned as the intelligence platform that connects strategy, creativity, media, CRM, commerce, data, and AI.

Different architectures. Same pressure. Each group is trying to build a marketing operating system that can sit between clients and the model providers. That layer may become valuable. It may also become contested terrain.

The model layer is not the same as the moat

The industry often talks about AI platforms as if the presence of an operating system solves the strategic problem. It does not. An operating system can coordinate work, normalize data access, manage permissions, orchestrate agents, present outputs, and improve adoption. But it does not automatically answer who controls the model behavior underneath.

That distinction matters because marketing AI is moving from assistance toward delegated action. A chatbot that helps draft copy is one thing. An agentic marketing system that recommends audiences, generates assets, allocates spend, triggers personalization, analyzes performance, and proposes optimizations is a different kind of infrastructure. The more the system acts on behalf of the client, the more model dependence becomes a governance and leverage issue.

A holding company can own the interface while renting the intelligence.

It can own the data environment while depending on an external model to interpret that data. It can own the workflow while a third-party model performs the generative or reasoning step. It can own the client relationship while the most consequential AI capabilities are governed by another company’s roadmap, pricing, safety policies, model weights, availability, indemnity structure, and acceptable-use rules.

That is not necessarily reckless. Very few companies outside the major AI labs can justify the cost of building and maintaining a frontier model.

Even many large enterprises that talk about proprietary AI are really building proprietary layers on top of commercial models, open models, or specialized smaller models. For marketing groups, the strategic question is not whether they should build their own GPT-class system. In most cases, they should not.

The question is whether clients understand the dependency chain.

When an agency says its platform is powered by AI, the client needs to know what part of the intelligence is proprietary, what part is configured, what part is trained or fine-tuned, what part is retrieved from client or agency data, and what part comes from an outside model provider. Those distinctions are not cosmetic. They affect confidentiality, explainability, performance stability, switching costs, auditability, contracting leverage, and operational continuity.

WPP is building around Google, not replacing it

WPP is the clearest case of the rented-intelligence problem because its Google relationship is so prominent.

The company’s five-year expansion with Google is strategically important. WPP says Google Cloud’s AI products, powered by Google DeepMind models, now fuel WPP Media’s Open Intelligence. The partnership is meant to accelerate bespoke audience model development, personalization, production, measurement, and marketing transformation. It gives WPP access to a powerful model and cloud ecosystem at a time when the company is trying to simplify its offer and restore growth.

That is a serious asset. It is also a dependency.

WPP can argue that the value does not sit only in Google’s models. It sits in WPP Open, in WPP’s data and media capabilities, in InfoSum’s privacy-safe collaboration architecture, in partner integrations, in client-specific implementation, and in the marketing expertise required to make the system useful.

That argument is credible. A model by itself does not know a client’s brand, market, margins, media commitments, compliance posture, customer segments, or internal politics. The agency layer still matters.

But the word “powered” carries weight.

If Google Cloud AI products fuel the system, then Google is not a vendor in the background. It is part of the cognitive supply chain. WPP’s model-agnostic posture, including its integration of Anthropic’s Claude through Amazon Bedrock, reduces single-provider exposure, but it does not remove the larger structural issue. The brain is still rented from frontier model providers.

The strategic question for WPP clients is not whether Google’s models are good. They are among the most capable systems in the market. The question is what happens when the economics, policies, or capabilities of the underlying model layer change.

If a model update affects performance, tone, latency, permissible use, image generation, brand-safety behavior, or data-handling terms, who has the right to inspect, challenge, adapt, substitute, or suspend that capability? If the agency has promised a marketing operating system, the client will expect operational accountability. But the agency may not fully control the underlying intelligence.

That is the uncomfortable middle position. WPP may be accountable to the client for the output of a system whose core reasoning capacity comes from outside WPP.

Publicis owns more of the fuel than the engine

Publicis has the strongest data-control story among the major holding companies. Its AI strategy did not begin with a single model partnership or a shiny assistant. It began with a long-running effort to own more of the identity, data, commerce, transformation, and personalization stack. Epsilon gave Publicis a large identity and data foundation. Lotame expanded that footprint to almost 4 billion profiles and more than 90% consumer coverage globally, according to the company’s own announcement. CoreAI then becomes the layer that activates the group’s proprietary data across the business.

This is why Publicis looks different from rivals. It has not merely built a platform on top of someone else’s model. It has built a large proprietary data environment that can make model outputs more specific, more measurable, and more commercially useful.

For marketing clients, that matters. Generic AI output is cheap. Client-relevant intelligence tied to identity, behavior, media performance, creative assets, business transformation work, and activation systems is harder to replicate.

But Publicis still does not own the foundation model layer. Its Adobe partnership integrates Adobe Firefly generative AI through CoreAI. Its Publicis Sapient relationship with Google Cloud is built around helping clients plan, deploy, and manage Google AI technologies and generative AI projects. Publicis can own more of the fuel than its competitors. It can own more of the activation layer. It can own more of the data context. The engine still comes from somewhere else.

That may be a smart division of labor. Publicis does not need to build a frontier model to win. It needs to make outside model capacity more valuable by combining it with proprietary data, identity, media, commerce, and execution. In practice, that may be a better business than trying to compete with Google, OpenAI, Anthropic, or Meta at their own game.

The risk is that clients may hear “CoreAI” and assume ownership of the full intelligence stack. They should not. Publicis’ advantage is not that it owns the brain. Its advantage is that it controls a large amount of the commercial memory the brain can use.

Dentsu is selling the hybrid moat

Dentsu’s positioning is more hybrid. The relaunched dentsu.Connect emphasizes agentic workflows, composability, interoperability, custom small language models, proprietary models and agents, and use of frontier models such as Google’s Gemini and Meta’s Llama. It also stresses cloud flexibility and the option for data to remain in the client’s own environment.

That is a different kind of promise from a pure proprietary platform claim.

Dentsu appears to be saying that the moat is not one single model but a system of specialized intelligence, agency expertise, cloud-agnostic integration, and client-specific deployment. That is strategically plausible. Many marketing tasks do not require frontier-model generality. A smaller specialized model, a well-designed agent, a retrieval layer, and a controlled workflow may outperform a generic large model for narrow use cases. In an enterprise environment, smaller and more specific can also mean lower cost, lower latency, easier governance, and more predictable behavior.

This is where the hybrid model becomes interesting. If the agency can combine frontier models for broad generative capability with smaller proprietary or custom models for domain-specific tasks, it may reduce dependence on any single provider while improving control. That does not eliminate rented intelligence, but it can make the architecture less brittle.

The open question is how much of that proprietary layer is truly differentiated. Every holdco can claim agents, workflows, custom models, and domain expertise.

The proof will sit in performance, switching flexibility, auditability, cost control, client adoption, and whether the system improves business outcomes without locking clients into a black box they cannot govern.

For dentsu, the strategic claim is less about owning the whole brain and more about owning the connective tissue between models, people, data, and client environments. That can be valuable. But it still leaves the frontier cognition layer outside the agency’s full control.

Omnicom bought more scale for the same dependency

Omnicom’s acquisition of Interpublic changed the scale of the market. The merged company can combine Omnicom’s Omni platform, IPG’s assets, Acxiom, Flywheel, media scale, production capacity, creative networks, commerce capabilities, and a much larger talent base. Omnicom framed the post-merger strategy around connected capabilities and enterprise generative AI. It said the combination gives the company scaled investment resources to capitalize on existing first-mover partnerships with leading frontier AI model providers.

That language is revealing. Omnicom is not claiming to own the frontier model. It is claiming scale, integration, and partnership leverage.

In one sense, that is the most realistic position. The combined Omnicom does not need to out-research DeepMind or OpenAI. It needs to use its scale to negotiate better access, deploy model capabilities across more clients, integrate AI into existing marketing operations, and make Omni the system through which those capabilities become useful.

But scale does not erase dependency. It changes the bargaining position. A larger Omnicom may have more leverage with model providers than a smaller agency group. It may negotiate better commercial terms, earlier access, stronger support, and deeper integrations. It may spread AI investment across a larger client base. It may also create more systemic exposure if many workflows depend on the same platform architecture and the same external model supply chain.

For clients, the concern is not only whether the merged company has enough AI capability. It is whether the AI capability comes with enough transparency. Which models are being used? Under what terms? Where does client data go? What is retained? What is logged? What can be audited? What happens if a client wants to move? Can the workflow be ported, or is the intelligence tightly bound to a platform the agency controls and a model provider the client may not have chosen?

The agency group that can answer those questions clearly will have an advantage. The one that hides behind platform language will invite distrust.

The data race was a hedge against rented cognition

The earlier data race in marketing now looks less like a separate story and more like the first phase of this one.

For years, holding companies have pursued identity graphs, clean rooms, commerce data, retail media relationships, customer data platforms, proprietary audience tools, measurement systems, and media-performance intelligence. Some of that was driven by privacy changes, cookie deprecation, platform fragmentation, and the need to defend margins against consulting firms and tech platforms. But AI adds another reason.

If the model layer is commoditizing, the value shifts to what the model can know, access, and do.

That is why the holding companies are racing to own or control the surrounding context. A frontier model available to everyone is not enough. The agency group needs proprietary inputs, trusted permissions, business-specific retrieval, workflow integration, performance feedback, and governance rules that turn general intelligence into usable marketing intelligence.

In that sense, the data moat is not separate from the model-dependency problem. It is the answer to it. If the brain belongs to someone else, the agency has to own the memory, the sensory system, the workflow, the permissions, and the commercial environment in which the brain operates.

This is also why clean rooms matter. InfoSum for WPP, Epsilon and Lotame for Publicis, Acxiom and Omni-related assets for Omnicom, and dentsu’s cloud-flexible client-data positioning are not merely data-management claims. They are attempts to make AI useful without handing over full control of sensitive information to the model provider. They are also attempts to preserve agency relevance in a world where clients can buy model access directly.

The holding company’s strategic fear is simple. If the client can buy the same model, use the same cloud, license the same creative tools, and connect its own data, what exactly does the agency own?

The answer has to be more than “we know how to prompt it.” It has to be governance, data architecture, workflow design, creative judgment, media leverage, operational accountability, measurement discipline, and the ability to make AI perform inside a real commercial system. That is much harder than launching a branded AI platform.

Client risk moves into the cognitive supply chain

The governance problem becomes sharper when agency AI systems move from recommendations to action.

In traditional agency work, clients already had to manage confidentiality, conflicts, media transparency, data rights, measurement disputes, and creative approvals. AI adds a new layer: cognitive supply-chain risk. The model that helps decide, generate, classify, summarize, personalize, or optimize may be outside both the client’s organization and the agency’s ownership.

That creates several practical questions.

A client should know whether its work is being processed through a frontier model, a smaller custom model, an open model, a vendor-specific model, or a combination of those systems. It should know whether prompts, outputs, embeddings, metadata, performance feedback, and derived insights are retained. It should know whether the model provider can use any part of the interaction for training, evaluation, safety testing, abuse monitoring, or product improvement. It should know whether the agency can switch providers without disrupting the workflow. It should know who is responsible when the system produces defamatory content, violates brand rules, leaks sensitive claims, misuses audience data, generates infringing assets, or acts outside approved campaign parameters.

Those questions are not hostile. They are basic vendor governance.

The old procurement checklist is inadequate because AI systems are not static software tools. They can change behavior after model updates. Their performance may vary across languages, markets, prompts, modalities, and data environments. Their terms can change. Their safety policies can change. Their pricing can change. Their availability can change. Their tolerance for certain political, medical, financial, sexual, or sensitive-content contexts can change.

A client that delegates more work to an agentic marketing platform is not only buying efficiency. It is accepting a chain of dependency.

The agency may manage the platform, but the model provider may shape the possible behavior of the system. The client may own the brand, but the rented brain may influence how that brand speaks, targets, responds, and adapts.

That is where authority risk enters the picture. When an AI agent acts on behalf of a brand, the client needs to know whose authority is being exercised. Is the agent following the client’s policy, the agency’s workflow, the model provider’s safety layer, the cloud provider’s terms, or some unstable mixture of all four? If those authorities conflict, which one wins?

Model agnosticism is useful but not sufficient

Many agency groups will answer the dependency problem with model agnosticism. That is the obvious response. If a platform can work with Google, OpenAI, Anthropic, Meta, Adobe, Amazon, and open models, then no single provider can control the system.

That is partly true. Model flexibility matters. It can improve resilience, pricing leverage, capability matching, and compliance options. It can also let clients choose models that fit their risk posture. A conservative financial-services client may not want the same model setup as a consumer brand running high-volume social content. A pharmaceutical client may require stricter validation, traceability, and review controls. A global brand may need language-specific model evaluation before allowing AI-generated content into local markets.

But model agnosticism is not the same as model governance.

A platform can be compatible with many models while still failing to disclose which model made a particular decision. It can support multiple providers while making switching difficult in practice. It can claim flexibility while binding clients into proprietary workflows, data schemas, agent definitions, and evaluation methods. It can route tasks across models in ways clients cannot inspect. It can produce outputs that appear unified while the underlying cognitive path is fragmented across vendors.

For serious clients, the question is not only “Can you use different models?” It is “Can you prove which model was used, why it was selected, what data it accessed, what it produced, who approved it, what policy governed it, and whether the result can be reproduced or challenged?”

That is where marketing AI platforms will either mature or remain promotional theater.

The agency advantage will come from accountable orchestration

The holding companies do not need to own frontier models to create value. In fact, owning one might be a distraction. The strategic opportunity is accountable orchestration.

That means taking outside model capacity and making it usable inside the messy, regulated, brand-sensitive, budget-constrained reality of enterprise marketing. It means mapping which tasks can be automated, which require human approval, which require legal review, which can use synthetic audiences, which can use client data, which must stay inside a client cloud, and which should not be handled by generative AI at all.

It also means designing AI systems that do not collapse into opaque convenience.

The agency platform should be able to show the client what happened. It should preserve decision records. It should distinguish suggestions from approvals. It should separate content generation from media activation. It should prevent agents from crossing authority boundaries just because the next step is technically available. It should allow model substitution without destroying the operating model. It should make governance part of execution rather than a policy document sitting somewhere outside the workflow.

That is where the holding companies can still win. Not by pretending to own the brain, but by proving they can govern the brain they rent.

The market will likely reward the groups that can combine three things: proprietary data, flexible model access, and defensible execution controls. Data without model access is inert. Model access without proprietary data is commoditized. Both without governance become liability.

What clients should watch now

The next phase of the holdco AI race will not be decided by platform names. It will be decided by proof.

Clients should watch how clearly each agency group explains the boundary between proprietary intelligence and third-party model capability.

They should ask whether the agency can provide model-level audit trails. They should examine whether client data is used only inside agreed environments. They should demand clarity on model updates, fallback procedures, indemnity, acceptable use, human approval, and portability. They should test whether the agency’s AI platform improves actual business outcomes or merely accelerates production volume.

They should also watch pricing. If the intelligence layer is rented, cost structures can shift quickly. A platform that looks efficient at pilot scale may become more expensive when deployed across markets, assets, languages, audiences, and always-on optimization loops. Model costs, cloud costs, licensing costs, data-clean-room costs, and human review costs all matter. AI does not eliminate marketing operations. It changes where the expense and risk accumulate.

Most importantly, clients should not confuse the ownership of an AI platform with ownership of AI capability.

A holding company may own the interface, the process, the data environment, and the service model. The underlying cognition may still belong to someone else. That does not make the platform useless. It makes diligence more important.

The real question is leverage

“Who owns the brain?” is not a philosophical question for the marketing industry. It is a leverage question.

If the brain belongs to Google, OpenAI, Anthropic, Meta, Adobe, or another model provider, then the agency group has to build leverage elsewhere. It has to own the data relationships, the workflow layer, the client integration, the governance framework, the domain expertise, and the measurable business outcome. That is exactly what the major holding companies are trying to do.

The winners will not be the ones with the loudest agentic language. They will be the ones that can turn rented intelligence into controlled execution without pretending the dependency does not exist.

That is the next honest version of the AI platform story. The operating system is only the visible layer. The strategic fight is underneath it.

The holdcos are racing to own everything around the brain because the brain itself is rented.