Kana CTO on Scaling Agentic AI in Marketing

1. Marketing activities involve multiple connected steps, from analyzing campaign performance to adjusting targeting and reallocating resources. How can AI agents coordinate these workflows while accounting for changing objectives and constraints?

Agents need one shared model of the goal. At Kana, objectives, budgets, and compliance rules sit in a planning layer above the agents. The planner breaks a goal into tasks and gives each task to a specialist agent: analytics, targeting, budget, or execution. Agents report results back, so each next step uses current data. When an objective or budget changes, the planner rebuilds the plan and the agents follow it. High-impact decisions go to a person for approval. This lets a team move from reacting to campaigns after the fact to continuous optimization. The hard part is state. You must always know what each agent knows and when it learned it.

2. Where do you see the strongest enterprise use case for agentic AI today: campaign execution, audience segmentation, content operations, performance analysis or cross-platform workflow management?

Cross-platform workflow management has the most value. Content generation and segmentation give real productivity gains, but each one works inside one tool. The large gain comes when an insight in one system becomes an action in another without a week of handoffs. For example, an agent sees that a promotion is losing margin in one region. It finds the audience that drives the loss, moves budget to a better segment, and sends the change for approval in the same hour. For that to work, agents must keep context across platforms, learn from past decisions, and keep a person in the loop. Most enterprises should start with one narrow workflow, such as segmentation, and connect outward from there.

3. Marketing decisions often involve competing objectives, such as improving conversion rates while controlling acquisition costs or protecting customer experience. How should AI agents evaluate these trade-offs when operating within enterprise marketing workflows?

The business must state the trade-off, and the agent must apply it. Give agents an explicit objective: a primary metric, hard constraints, and the limits you will accept. For example: “maximize conversions, CAC must stay under $40, frequency must stay under five per week.” Agents cannot trade against hard constraints. People set and own the weights on soft goals. Agents also need history: what the team decided before, why, and what happened. When a trade-off goes past an agent’s authority, the agent escalates with the options and the cost of each. Brand, customer trust, and large spend stay with people. Each human decision becomes feedback that improves the next recommendation.

4.  An AI agent that can execute marketing actions introduces different risks from a system that only generates recommendations. What approval mechanisms and monitoring controls are necessary before agents can operate across enterprise marketing platforms? 

Match oversight to risk. Read-only actions run freely. Reversible, low-cost actions run within spend and policy limits. Irreversible or high-spend actions need human approval. Every action needs four things: an identity (which agent, on whose behalf), a policy check before it runs, a log of what it did and why, and a way to undo it. Monitoring must watch business outcomes as well as errors, because an agent can work correctly and still waste money. Set spend caps, rate limits, and anomaly alerts for each platform, and give people the power to pause any agent. Build these controls into the workflow before the first agent goes live.

5. Will enterprise marketing evolve towards coordinated networks of specialized agents, or will organizations continue to rely on a central AI system to manage most workflows? What are the trade-offs between these architectures?  

I expect coordinated networks of specialized agents that work inside one shared framework. Specialized agents do better work in narrow domains, and each one can use the model that best fits its task on quality, cost, and latency. A central layer holds goals, policy, shared state, and the audit trail. A purely central system is simpler to govern, but it becomes a bottleneck and gives average results across tasks. A purely decentralized network scales well, but agents conflict, repeat work, and lose context, and nobody owns the outcome. The best design puts coordination and governance in the center and keeps specialized skills at the edges. Success depends on how well the agents share context and priorities.

6. What are the biggest technical barriers to deploying agentic AI in existing enterprise market stack? Are integration, data quality, identity management or system interoperability proving most difficult to address? 

The largest barrier is fragmentation. Most enterprises have added platforms over many years, and each platform has its own data model, permissions, and definition of a customer. Protocols like MCP make it easy to connect an agent to a system. A connection gives access, but it does not give the context an agent needs to decide well. An agent that works across five systems with five definitions of “active customer” will make confident errors at scale. Data quality and identity are the most difficult parts to fix. Start with the semantic layer: shared definitions, resolved identity, consent attached to the data, and access to past decisions and the reasons for them.

 7. As organizations introduce multiple agents across marketing functions, how can they prevent duplicated work, conflicting actions and fragmented decision-making between systems?

Three controls prevent most conflicts. First, assign one owner to each resource. Only one agent at a time can change a given budget, audience, or campaign, and other agents must request changes. Second, keep a shared state store that every agent reads before it acts. It holds current campaign activity, work in progress, and the reasons behind past decisions. Third, use one policy layer that every agent calls, so business rules stay consistent across platforms. Most conflicts happen when teams deploy agents inside separate tools, each with its own logic. If agents cannot see each other’s actions, they work as separate automations on the same budget.

8. Looking ahead to 2027, what separates enterprises that operationalize agentic AI from those that remain stuck in experimentation? What should marketing leaders prioritize now to prepare for that transition? 

Leaders will run agents as operational infrastructure. Laggards will have many demos, no shared state, and no owner for agent outcomes. Marketing leaders should do three things now. First, pick two or three workflows with a real bottleneck and clear ROI. Second, build the foundation for those workflows: resolved identity, shared definitions, and written policies that agents must obey. Third, define decision rights: what agents can do alone, what they must escalate, and which person is accountable. Change how you measure success too. Count better decisions, faster time from insight to action, and less friction between teams. The number of agents deployed tells you little.