How to adopt AI agents in the CRM operations of marketing: it’s not which agent to use, but how to orchestrate them
Every day a new AI agent appears that promises to transform some aspect of marketing. Marketing platforms have their own agents, CRM systems do too, agencies are developing their own, and brands themselves are starting to consider creating in-house agents. This is generating an important strategic decision: Should we trust the agents built into the platforms or build our own agents?
There is a conversation going on in the Marketing and CRM teams at B2C—should we use the AI agents that our platform already incorporates, or build our own? It’s a reasonable question. And it’s, to a large extent, the wrong question.
The wrong question isn’t because it doesn’t have an answer—it does, and we’ll see it—but because it places the focus on technology instead of the outcome. The real dilemma faced by CRM teams isn’t technical: it’s strategic. And it has to do with how to get different agents, platforms, data, and teams to work together coherently to produce relevant customer experiences on a scale.
That ability has a name: Orchestration of AI Agents. And it is, in our view, where the real competitive advantage in CRM and Marketing Automation will be played out in the coming years.
The dilemma facing marketing CRM teams today
The options on the table for any brand that wants to incorporate AI agents into its CRM operations are essentially two.
The first is trusting the agents already integrated into the platforms. Braze, Salesforce, Klaviyo, Bloomreach have —or are actively developing— native agent capabilities: autonomous shipment optimization, predictive segmentation, content generation, automated decisioning. The advantage is obvious: they are easy to adopt, they are integrated by default into the stack that the client already uses and scale with relative ease.
The problem is also evident: the standard agents of the platforms do not know the specific context of each company. They do not know what the communication tone of this brand is, what business rules they apply in each market, which segments have a differential historical behavior, which campaigns have worked and which have eroded the relationship with the customer. They operate on generic data and produce generic outputs. And they generate dependence on a closed ecosystem that can be difficult to reverse.
The platform agent is powerful but generic. The native agent is specific but expensive and slow to build. The real question isn’t which one to choose: it’s how to get either—or both—to produce results that are consistent with the brand’s strategy, connected to their own data, and coordinated with the rest of the stack. That’s the missing layer in most current implementations. |
Why orchestration is the answer to both options
Faced with the dilemma between platform agents and in-house agents, there is a third way that is emerging as the most relevant one: orchestration. Not as an alternative to any of the previous two options, but as the layer that makes any of them work well.
The orchestration of AI agents in CRM consists of combining three capabilities that, separately, none of the previous options fully resolves:
- The platforms provide the infrastructure: Computational power, trained models, scalability, and security that no medium-sized brand could build on its own at a reasonable cost. This layer exists and works. The challenge isn’t building it: it’s leveraging it.
- The brands provide the context: first-party data, knowledge of the real behavior of their customers, specific business rules, brand strategy, tone of communication, and their own use cases. This context is the most valuable asset a brand has and one that generic platform agents cannot replicate.
- The orchestration connects the two planes: It designs the ecosystem of agents, decides what each one does when, connects the context data with the platform's models, defines the guardrails within which the agents can act autonomously, and establishes the points where human intervention remains necessary.
The result of combining these three elements well is something that neither of the initial options alone achieves: scale customization with brand context, operational efficiency without loss of strategic coherence, and production speed without erosion of identity.
What this means in practice for a CRM team
Orchestration is not an abstract concept. It materializes in concrete decisions about how the workflow of the CRM team is designed when agents start operating within it.
Instead of starting with a large AI architecture, the most effective starting point is to experiment with the agents that existing platforms already incorporate—Braze, Klaviyo, Salesforce, Bloomreach—and to precisely analyze three questions: what works well without additional intervention, where are the limitations, and most importantly, where do they lack context about the customer or the brand to produce truly useful outputs.
That last question—where they lack context—is exactly where the orchestration work lies. Because the value isn’t in activating an agent: it’s in redesigning how a given task is performed when the agent can take on part of it. Identifying the specific friction point in a workflow, mapping out the entire process, determining which tasks the agent takes on and which still require human judgment: that’s the design work that makes the difference between an AI implementation that generates value and one that generates confusion.
The data that changes the analysis: agents need context, not just data
There is a common mistake in AI implementation projects in marketing: confusing «connecting data» with «giving context.» They are not the same thing, and the difference matters a lot in the performance of agents.
An agent who has access to the entire customer database but doesn’t know what it means for this brand that a customer buys at a discount versus at full price, or what it strategically means that a customer from a specific market has a lower purchase frequency than the average, will produce segmentations and messages that are technically correct but strategically neutral.
The context that agents need is not data volume: it is data specifically prepared so that they can make decisions aligned with the brand strategy. This includes codified business rules so that the agent can apply them, a history of what has worked and what has not for each segment, and explicit restrictions on what the agent can and cannot do on behalf of the brand.
In the future, it’s likely not the brand with the best individual agent that will win, but the one that manages to build the best-orchestrated agent system. The question worth asking is not «which agent should we use?» but «how do we get all the agents, platforms, data, and teams to work together coherently?» That’s the question that separates brands that scale with AI from those that get stuck in pilots that never mature. |
The human-in-the-loop model: why human supervision is not a step backward
One of the risks of enthusiasm for agential AI is assuming that the ultimate goal is to maximize the agent's autonomy and minimize human intervention. That logic is understandable—it comes directly from the promise of efficiency that AI makes—but in the context of CRM, specifically in sectors like retail, fashion, and DTC, it is more nuanced than it seems.
Self-driving agents optimize towards metrics. When those metrics are perfectly aligned with the brand’s strategic goal, full autonomy makes sense. But in brand marketing, the alignment between metrics and strategy is rarely perfect: an agent that aggressively optimizes the conversion rate may be simultaneously eroding the exclusivity perception of an aspirational fashion brand, and that doesn’t show up in any dashboard until the damage is already done.
The model that makes the most sense for most retail CRM teams at the moment is the human-in-the-loop: agents that operate autonomously within well-defined guardrails, with human oversight and validation of decisions that have the most strategic implications. It is not a model of transition towards full autonomy: it is a model of collaboration that recognizes that brand-based criteria and strategic business understanding are human contributions that are not easily automated.
Open architecture as a prerequisite to not getting stuck
There is a technical consideration that has very direct strategic implications: the orchestration architecture of agents must be open, not closed. The goal is to avoid replacing a closed ecosystem with another one.
Open standards are emerging—such as the Model Context Protocol (MCP)—that allow different models, platforms, and data sources to communicate with each other without relying on a single provider. An orchestration architecture built on open standards enables connecting and replacing components as technology evolves: if a platform launches an agent with better capabilities, it can be incorporated without rebuilding the entire system. If the context data evolves, it can be updated without affecting the rest of the architecture.
The long-term goal is an ecosystem where platform agents, more specific agents, and human teams work together seamlessly, with each component doing what it does best and with orchestration ensuring that the whole produces something coherent with the brand’s strategy.