11th Nov 2026
Agentic AI in Marketing: Six Controls for AI Governance
By Len Van Hoogenhuijze
Agentic AI in marketing promises faster campaign planning, asset creation and execution. But as more agents connect to marketing systems, a practical question becomes unavoidable: how do you give them freedom to act without losing control of what goes live?
That was the focus of our session at the recent Agentic AI Marketing Summit in London. In Taming the Agentic Jungle, we explore how a control plane can connect agentic execution to the rules, responsibilities and reporting requirements of your marketing operation.
Watch the session recording here, or read on for the key takeaways.
What an AI control plane actually does
A control plane sits between the AI models that plan and propose actions and the downstream platforms where campaigns are built and executed, such as HubSpot, Marketo or Salesforce. It provides the organisational context and boundaries within which agents operate: business knowledge, taxonomies, policies, templates, quality checks and approval requirements.
Rather than letting every agent invent its own approach, this layer defines which actions are available and what must happen before execution can progress. That helps keep campaign structures consistent, preserves reporting requirements and makes activity visible. The aim is not manual intervention at every turn. It is autonomy within defined boundaries.
When useful AI experiments become an agentic jungle
One team builds an agent to draft emails. Another creates one to generate audience segments. Elsewhere, someone automates campaign IDs or connects several tasks into an end-to-end workflow. Individually, each experiment makes sense. Collectively, they can create overlapping capabilities, inconsistent prompts and uneven controls. Ownership can become unclear, particularly when the person who built an agent moves on.
The answer is not to abandon experimentation. It is to recognise that instructions are not the same as control, and enthusiasm is not a strategy for scaling. Asking an agent to follow your process is different from making sure it cannot bypass that process.
Keep humans accountable, not involved in every task
In the workflow we discuss in the session, a human initiates the campaign and an agent proposes the plan, including targeting and channels. A human approves that plan. The agent then builds assets or drafts copy, with people reviewing and refining the output. Agents can also prepare the campaign in downstream systems, but the decision to launch remains with a human.
After deployment, agents can monitor results and surface risks or learning. This creates a feedback loop without confusing automated activity with decision-making authority. Moving from idea to deployment within a day is an aspiration, not a guaranteed result. The important distinction is where agents add value and where human judgement remains essential.
Six controls for governing agentic AI in marketing
Keeping people accountable is the core principle. The next step is to make that principle part of the workflow, rather than relying on an agent to interpret instructions correctly.
Our approach to AI governance focuses on six complementary controls that connect access, permitted actions and human decisions. Together, they move governance beyond the prompt and into the way execution works:
- Identity and permissions: give every agent and user a defined level of access. This establishes who or what is acting and which parts of the marketing environment they can use. It is the starting point for separating permission to prepare campaign work from permission to execute it.
- Policies and guardrails: establish behavioural rules, transparency expectations and relevant cost or token limits. Once access is defined, these boundaries describe how it should be used. They connect an agent’s work to organisational requirements, rather than leaving each agent to interpret acceptable behaviour independently.
- Action space configuration: restrict agents to approved tools, methods and actions in downstream systems. This translates policy into practical limits on execution. An agent may be allowed to build a campaign draft, for example, without being given the ability to launch it. The distinction is between asking it not to act and withholding the action itself.
- Approval thresholds: specify where human confirmation is required before work can progress. These checkpoints connect automated preparation to accountable decisions, including approval of the campaign plan and the decision to launch. People remain responsible for those decisions without having to perform every task between them.
- Monitoring and audit: retain a reviewable record of activity so issues can be investigated. Permissions and approvals establish the boundaries; monitoring provides visibility into what happened within them. Campaign data and logs give teams a basis for reviewing execution, understanding problems and feeding learning back into the workflow.
- Stop and recovery mechanisms: provide ways to halt execution and, where possible, roll back changes. Visibility alone is not enough if a team cannot intervene when something goes wrong. These mechanisms complete the control framework by connecting detection to a practical response, while recognising that not every action can be reversed.
Taken together, these controls describe a connected sequence: establish who can act, define the boundaries, constrain the available actions, require approval where needed, and make execution visible and interruptible. Turning that sequence into everyday practice is an operating-model question as much as a technical one.
Build around the operating model, not just the technology
Marketing operations is a natural home for this approach because it already connects the foundations agents depend on: data, technology, process and measurement. Those foundations give the controls their organisational context.
Approval thresholds need clear ownership, permitted actions need to reflect campaign processes, and monitoring needs to capture the information required for reporting. Faster execution cannot compensate for unreliable data, disconnected tools or unclear approval paths. Without those foundations, agents can simply accelerate inconsistency.
The technology then needs to support that operating model. We focus on four building blocks: an agent runtime, orchestration, a data repository and collaboration. The runtime provides somewhere for agents to run; orchestration sequences the work; the repository holds campaign data and logs; and collaboration supports human review and approvals. Together, they connect the work agents perform with the information and decisions needed to govern it. These are four capabilities, not necessarily four separate products.
The real outcome is trust
Bringing the controls and operating model together starts with the outcomes you need, not the tools you want to deploy. Agree those outcomes, including reporting requirements and the data you must capture. Then clarify ownership and map the campaign workflow so that agent autonomy is explicit: what can happen independently, at which stage, and when approval is mandatory. This connects the ambition for faster execution to a shared understanding of how work will be governed.
A control plane is not a promise that nothing will fail. It is a way to make execution more accountable, observable and recoverable. Trust comes from knowing what agents are allowed to do, where people retain authority, and how the team can investigate or interrupt activity when needed. That is the basis for greater confidence in AI-powered marketing: not unrestricted autonomy, but a structure that lets you expand it responsibly.
FAQs
A control plane is a governance layer between AI agents and the systems where campaigns are built and executed. It provides the organisational context agents need, including business knowledge, taxonomies, policies, templates and quality checks. It also defines permissions, available actions and the approvals required before work can progress. For example, an agent may be able to prepare a campaign in a downstream platform without having permission to launch it. The aim is autonomy within defined boundaries, supported by visibility into campaign activity and clear human accountability.
Prompts describe how an agent should behave, but instructions are not the same as enforced controls. Asking an agent not to launch a campaign is different from withholding its ability to launch one. Permissions, restricted actions and mandatory approval steps establish boundaries beyond the prompt, limiting what an agent can actually do in connected systems. Monitoring and audit records then make activity reviewable, while stop and recovery mechanisms provide ways to intervene if something goes wrong. Prompts still guide the work, but they sit within a wider framework of control.
No. The aim is to keep humans accountable without requiring them to perform every task. In the workflow described here, a human initiates the campaign and an agent proposes a plan, including targeting and channels, for human approval. Agents can then draft copy, build assets and prepare the campaign in downstream systems, with people reviewing and refining the output. The decision to launch remains with a human. After deployment, agents can monitor results and surface risks or learning, supporting a feedback loop without taking over decision-making authority.
Not necessarily. A control plane connects agents to downstream marketing platforms through governed orchestration, rather than assuming those platforms must be replaced. The architecture should reflect your existing stack, campaign processes and operating model. The four building blocks described in this article are an agent runtime, orchestration, a data repository and collaboration to support human review and approvals. These are capabilities, not necessarily four separate products. The key is how they work together to preserve permissions, approval requirements, campaign data and reporting as agents prepare and execute work.
Start by agreeing the outcomes you need, including reporting requirements and the data you must capture. Clarify who owns the campaign workflow and its key decisions, then map the stages from planning through to deployment and monitoring. For each stage, define what an agent can do independently, which tools and actions it can access, and where human approval is mandatory. Establish how activity will be recorded, how issues will be investigated and how execution can be stopped or changes reversed where possible. This gives you a defined basis for expanding autonomy, rather than adding agents before the operating model is clear.