TL;DR
- Don’t start with the smallest use case. Start with one where decisions can be evaluated, mistakes are visible and business impact is measurable.
- Agentic AI will only be as reliable as the data, integrations and processes it operates on.
- Autonomy should be earned. Let agents recommend first, prepare next, make bounded decisions over time and execute independently only when they have demonstrated reliability.
- Keep human oversight where decisions affect customer trust, commercial commitments, regulated interactions or critical business data.
- The real challenge isn’t introducing AI into your marketing stack. It’s making sure your marketing operations are ready to work with it.
There is a lot of conversation around what agentic AI could do for marketing. From building campaigns and optimizing customer journeys to identifying churn signals and making real-time decisions, the possibilities are expanding quickly. But there is a question that doesn’t get nearly enough attention: Are enterprise marketing operations ready to let AI make decisions on their behalf?
Having worked across marketing technology implementations, campaign operations and lifecycle marketing programs, we’ve seen that introducing a new capability is rarely the difficult part. Making it work consistently across systems, teams, data and existing processes is where the real effort lies.
Agentic AI brings this challenge into sharper focus.
The usual recommendation is to start small. That sounds sensible, but we don’t think the size of a use case should determine where an organization starts. A seemingly simple agent that modifies CRM records can introduce problems that remain unnoticed for months. An agent that optimizes send times might make a poor decision, but the impact is usually visible within the next campaign cycle.
The distinction is important. Your first agentic AI initiative should be selected based on how well you can measure its decisions, contain its mistakes and demonstrate business value, not simply how easy it is to build.
This was one of the central themes in our recent discussion with Mary Miller, Account Director at Mavlers, Greg Kihlström, MarTech advisor and host of The Agile Brand, and Matt Kelly, Growth Strategy Partner at Mavlers.
Haven’t watched the discussion yet? Watch the full webinar recording for a deeper conversation on where agentic AI fits into the marketing ecosystem, the operational challenges enterprises need to address and what it takes to move from experimentation to meaningful adoption.
Here are some of the perspectives from that conversation, along with what we believe marketing leaders should consider before introducing agentic AI into their operations.
1. Not every automation problem needs an AI agent
One of the first things we need to get right is understanding the difference between automation and agentic AI. There is a tendency to label almost every AI-enabled workflow as agentic. But adding an AI layer to an existing automation doesn’t necessarily make it an agent. Traditional marketing automation works well when the process is predictable. If a customer abandons a cart, send a reminder. If a subscriber hasn’t engaged in 30 days, move them into a re-engagement journey. These workflows have predefined conditions and execution paths.
An agent becomes more useful when the situation requires interpretation and a decision between multiple possible actions. Greg Kihlström made this distinction during our discussion. If a process can be mapped entirely through a sequence of if-then statements, it is probably better handled through conventional automation.
Consider a customer service scenario where a returned product arrives damaged and is also the wrong size. There may be several possible resolutions – a replacement, a refund or an alternative offer. The appropriate response depends on the circumstances, customer history and business rules.
This is where an agent can add value by evaluating the available information and recommending an appropriate course of action. Our view is straightforward: don’t introduce an agent simply because the technology allows it. Use one when the workflow genuinely benefits from contextual decision-making.
For everything else, well-designed automation is often simpler, more predictable and considerably less expensive to maintain.
2. Your data problems won’t disappear when you introduce AI. They’ll become more visible.
This is probably the least exciting part of an agentic AI conversation, but arguably the most important.
Most enterprise marketing environments have some degree of data fragmentation. Customer records sit across CRM platforms, marketing automation systems, CDPs, analytics environments and other business applications. Different teams maintain different definitions, integrations don’t always synchronize consistently, and duplicate records are more common than organizations would like to admit.
These issues are manageable when people are involved in reviewing and correcting information. They become considerably more challenging when autonomous systems start using that information to make decisions. Mary Miller put it well during the webinar: agents amplify what’s already there. If the underlying data is messy, organizations should expect faster mistakes, not better decisions.
We’ve seen how seemingly minor operational inconsistencies can create larger problems. Two teams may use different definitions of an active customer. A customer may exist as multiple records across platforms. An integration may pass incomplete information without triggering an obvious error.
An AI agent working across these systems may make decisions that appear perfectly reasonable based on the information available to it, but are fundamentally incorrect from a business perspective.
This is why we would prioritize a few foundational areas before expanding agentic capabilities:
- Customer identity and deduplication: Establish a consistent customer identity across systems and resolve duplicate records.
- Data definitions and governance: Ensure teams agree on what critical customer attributes actually mean.
- Integration reliability: Validate that the information exchanged between platforms is accurate, complete and current.
- Ownership and accountability: Identify who is responsible for the data, the workflows and the decisions made using them.
Interestingly, data quality can itself be a good starting point for agentic AI. An agent can identify duplicate records, assess potential matches, recommend merge strategies and flag cases where merging could result in information loss. Instead of asking an administrator to manually review thousands of records, the agent can prioritize the work and prepare recommendations.
Matt Kelly brought up a practical example of administrators working against tight deadlines to clean CRM records and inadvertently deleting valuable information in the process. This is precisely where an agent can help, not by immediately taking over the process, but by making the work more systematic, reducing manual effort and keeping critical decisions subject to review.
The goal should be to make your data foundation more reliable before allowing AI to make decisions at scale.
3. Don’t give an agent more autonomy than it has earned
A question we increasingly hear from marketing leaders is how much control they should give an AI agent.
There isn’t a universal answer. It depends on the nature of the workflow, the consequences of an incorrect decision and how quickly the organization can identify and correct it. During our discussion, Mary emphasized starting with workflows that have clear guardrails and relatively fast feedback loops. Greg highlighted the importance of understanding the cost of a mistake. Matt stressed the need to document and audit agent activity.
We believe these considerations should determine how autonomy is introduced.
A practical way to approach this is through four levels.
Level 1: Recommend
The agent analyzes information, identifies opportunities and recommends actions. A person reviews and executes them.
Level 2: Prepare
The agent builds campaigns, prepares content, configures workflows or queues actions. A person reviews and approves execution.
Level 3: Decide within boundaries
The agent makes decisions within approved rules, templates and operational limits. People monitor outcomes and review exceptions.
Level 4: Execute autonomously
The agent independently executes approved activities within its assigned scope. People retain responsibility for governance, monitoring and intervention.
The important point is that organizations shouldn’t jump directly to Level 4. An agent should demonstrate consistent performance at each stage before receiving greater authority. Consider how this translates into everyday marketing operations:
| Workflow | Recommended autonomy | Why |
| Identifying duplicate CRM records | Level 1 | Every proposed merge needs validation |
| Bulk CRM updates and deletions | Level 1 | Incorrect changes can be difficult to reverse |
| Campaign preparation and staging | Level 2 | Human approval can happen before launch |
| Win-back campaigns involving discounts | Level 2 | Commercial decisions need oversight |
| Send-time and channel optimization | Level 3 | Performance can be evaluated continuously |
| Content variations within approved templates | Level 3 | Outputs can be constrained and measured |
There are also areas where we would be particularly cautious about granting autonomous execution: pricing, discounts, regulated industries such as financial services and healthcare, decisions with significant financial implications, and any operation that writes directly to a system of record.
A mistake in a campaign configuration can often be corrected. A bulk CRM deletion or an incorrectly issued commercial offer may be much harder to reverse.
The principle is simple: let agents demonstrate that they can make good decisions before allowing them to execute those decisions independently.
4. Be careful about what you’re measuring. Fast feedback doesn’t always mean a good outcome.
There is another challenge that becomes particularly relevant when agents start influencing customer behavior.
Let’s take churn prediction and win-back campaigns. An agent identifies a customer as showing signs of potential attrition and triggers a retention intervention. However, the communication itself may influence the customer’s perception of the brand or even accelerate the decision to leave.
Matt Kelly raised this possibility during our discussion, pointing out that engaging too early based on an attrition signal can sometimes become a self-fulfilling prophecy. This creates a measurement challenge. The agent predicts churn, initiates an action and subsequently observes churn. On the surface, the prediction appears accurate. But did the customer leave because they were already planning to, or did the intervention contribute to the decision?
This is an important distinction for lifecycle marketers. We would recommend using holdout groups wherever possible. If an agent identifies 1,000 customers as being at risk, a controlled group can be excluded from the intervention. Comparing retention outcomes between the two groups helps establish whether the agent-triggered action actually improved results.
The same principle applies to personalization, recommendations, engagement optimization and other customer-facing interventions. We shouldn’t evaluate an agent simply by how many actions it completes or how accurately it identifies a signal.
We need to understand whether those actions create incremental value. For marketing teams, that means looking beyond execution metrics and measuring outcomes such as incremental retention, conversion, customer lifetime value and campaign profitability.
5. You don’t need to rebuild your MarTech stack to get started
A common misconception is that adopting agentic AI requires organizations to rethink their entire marketing technology architecture. We don’t see it that way.
Most enterprises have already invested significantly in platforms such as Braze, Salesforce Marketing Cloud, HubSpot, Klaviyo and their surrounding data ecosystems. Many of these platforms are incorporating AI capabilities that can be applied to existing workflows.
The more practical approach is to understand what your current ecosystem can already support. Before investing in another platform or building a custom agent, evaluate the AI capabilities available within your existing technology, the data they can access, the integrations they require and the operational costs involved.
For example, an organization running lifecycle marketing on Braze may already have customer segmentation, campaign orchestration and engagement data available within its ecosystem. The next step may be to introduce AI-assisted quality assurance, improve audience selection or support more intelligent campaign optimization. That doesn’t necessarily require a new platform.
6. Governance is what makes agentic AI scalable
A successful pilot is one thing. Managing multiple agents across an enterprise marketing environment is an entirely different challenge.
As agents gain access to customer data, campaign configurations and operational systems, organizations need to be increasingly deliberate about permissions, accountability and oversight. We would encourage marketing and technology leaders to establish clear answers to four questions before expanding deployment.
Who owns the agent?
Every agent needs a named business owner. Someone must be accountable for its performance, operating costs, decisions and exceptions. Ownership cannot sit vaguely between marketing, IT and data teams.
What can the agent access and change?
An agent should only have access to the information and systems required for its assigned responsibilities. Permissions need to be clearly defined, with appropriate restrictions on write-back operations and sensitive customer information.
Are your data privacy and consent frameworks adequate?
Organizations need to understand what customer data is being shared with models, how it is processed and whether existing consent, contractual and privacy requirements cover those activities.
Will the economics hold at scale?
A workflow that works at today’s volume may become significantly more expensive as usage increases. Model costs, infrastructure, integration maintenance and ongoing monitoring should all be considered when evaluating the business case.
These are not questions that marketing can answer in isolation. Agentic AI requires closer collaboration between marketing operations, IT, data, security, legal and finance. The sooner these responsibilities are established, the easier it becomes to expand beyond isolated pilots.
The real opportunity is in building more intelligent marketing operations
Agentic AI will undoubtedly change how marketing teams operate. But we believe its greatest value will not come from replacing individual tasks with autonomous agents. It will come from connecting data, decisions and execution more intelligently across the marketing ecosystem.
Imagine a marketing environment where agents can identify customer behavior changes, evaluate journey performance, recommend improvements, prepare campaign variations and optimize execution within approved business parameters.
That is a meaningful shift from automating individual tasks to building more adaptive lifecycle marketing operations. But getting there requires a strong foundation. Organizations need reliable data, well-integrated platforms, clearly defined processes, measurable outcomes and governance that can support increasing levels of autonomy.
The businesses that get this right won’t necessarily be the ones deploying the most agents. They will be the ones that know where to use them, what decisions to trust them with and how to measure the value they create.
Before asking how many agents you can introduce into your marketing stack, ask a more important question: Is your marketing operation ready to support them?
If you’re evaluating where to begin, take the Lifecycle AI Readiness Assessment. It can help identify the operational foundations and opportunities to address before expanding agentic AI across your marketing ecosystem.




