What should you actually let a marketing AI agent do?
On this page
Let a marketing AI agent start fully human-gated, then let it earn its way out. Once it has produced good enough output consistently enough, you can take the gate off that job. Anything that goes out into the world keeps a human on it. Guardrails are the floor, not the answer.
The permission layer just became a product feature. PubMatic put a five-step guardrail architecture into AgenticOS on 5 August, down to a one-time passcode for the named human who signs off when an agent exceeds its authority. Meta added ads MCP server rules on 16 July, letting a business portfolio owner govern what an agent may do on their ad account, down to the budget.
I am super happy about these high-level guardrails being implemented into tools. I have been building my harness in a way where guardrails are a big part of it, because we are also building Parvi to be very safe and secure.
But a platform rule only says what an agent may do inside that platform, not what you should let it do at all. That is the question anyone doing agentic marketing actually has.
Agents graduate
My own permission system is that agents can graduate from human-gated to automated if and when they have produced high enough quality output, and consistent enough output, a number of times. I can then remove the human gate once they have proved they can work like that.
The shape is the same every time:
- The human initiates the process and does the plan, with the help of a flagship model.
- The AI does a lot of the work in between.
- The human is there at the end to guard the quality and give the okay for the output.
That is basically how I do it with every single one of my agents, disregarding some small things where the agents have proved that they always create a good-quality output.
To some degree, a lot of the same principles apply to leading people. You give them the freedom to operate and see how the output is, and at some point you do not have to look over their shoulder all that much, because you know that they will provide good quality output. So in a way it is very close to leading people, I would say.
Where I half-agree with Attentive
Attentive calls it the permission paradox: start too restrictive and you get workarounds and shadow agents, so transparency and auditability matter more than tight restrictions.
I think both can be right in a way. If you have too restrictive rules for publishing, or for having agents take action, that is true. But I believe that ultimately, things that go out into the world, whether it is newsletters, social media posts, account edits, outreach messages or whatever, should still have a human gate.
Loosen inside the system. Hold the line at the edge of it.
Some of it is not a permission setting at all. Any time something is pushed into any repository I have a guardrail checking for PII data or secrets, and every client is tenanted in their own system. No ads platform does either for you.
What to do this week
Turn on the platform rules that just shipped, then decide per job, not per tool, whether the agent is gated or graduated. Write down what “consistent enough” means before you need to argue about it, and keep the gate on anything that reaches a customer. Where those gates belong is a good part of what agentic AI consultants in Finland do.
FAQ
What are some effective governance frameworks for agentic AI?
The two worth reading come from different readers. Palo Alto Networks sets out eight implementation steps, from defining scope and authority through to drift monitoring, written for security teams. Attentive's is written for marketers and splits into policy, operational and runtime governance with four risk tiers. Neither answers the question of when to stop gating an agent, which is the decision most teams actually face.
How is agentic AI regulated?
There is no agent-specific regime. You inherit the obligations you already had, data protection and contractual ones included, plus the terms of whichever platforms the agent touches. As Palo Alto Networks puts it, autonomy changes how systems operate, not who is responsible: liability stays with the organisation deploying the agent.
Can you give me an example of AI governance?
Mine is simple. Every agent starts human-gated: I plan the work, the agent does the middle of it, and nothing goes public until I approve the output. An agent graduates out of that gate only once it has produced good enough output consistently enough, and anything that reaches a customer keeps its gate permanently.
What are the four pillars of agentic AI?
There is no agreed set of four. The components that recur across the published frameworks are scope and authority limits, identity and access boundaries, runtime guardrails, audit logs, human oversight thresholds, and a kill switch. If a page gives you exactly four, it is that list trimmed to fit the headline.
