What humans own. What agents own. How the loop actually works.
By Chetan Parmar · April 1, 2026 · 7 min read
Every serious analysis of agentic marketing from 2026 - Treasure Data, Talkwalker, Amperity, the Content Marketing Institute - converges on the same operational model. Humans set objectives and guardrails. Agents execute toward those objectives and escalate when they hit boundaries.
The teams that struggle with human-in-the-loop marketing are the ones who either give agents too much autonomy (and get surprised by errors) or too little (and replicate the manual workflow with extra steps). Getting the split right is the framework.
Key Takeaways
Here is what the human-agent split looks like in a real Google Ads workflow. For context on whether this constitutes genuine agentic AI rather than rebranded automation, the three-part test applies.
Objective (human sets)
Reduce CPL by 15% without reducing conversion volume, over the next 30 days. Report weekly progress.
Guardrails (human sets)
Max weekly spend change of 20% per campaign. No campaign pausing without human approval. No changes to the top 3 highest-converting ad groups without review. Flag any single-day spend anomaly above $500.
Execution (agent handles)
Daily monitoring of all campaigns. Weekly analysis report with specific recommendations. Identification of search term wasted spend and negative keyword suggestions. Detection of quality score drops with probable causes.
Human review moments
Weekly report review. Approval of any recommendations before implementation. Review of anomaly flags within 24 hours.
Too vague an objective
The agent optimizes for the wrong metric. "Improve performance" with no definition produces work that looks busy but does not move what matters.
Missing guardrails on high-risk actions
The agent makes changes the human would not have approved. A bid change script running on a live account without a spend cap is a real risk.
Reviewing everything
If humans review every output, the agentic system is not saving time - it is just adding overhead. Define what gets sampled vs what gets full review.
Unclear escalation triggers
The agent does not know when to stop and ask for guidance. Define specific conditions that require human input before the agent can proceed.
A four-component model for deciding what belongs to the human and what belongs to the AI agent in a marketing workflow. Objectives and guardrails are always human. Execution is always agent. Learning is split: the agent surfaces data, the human interprets it and updates strategy. The framework prevents both over-delegation and under-automation.
In Google Ads: the human sets campaign objectives, target ROAS, and brand safety rules (guardrails). Claude Code handles account audits, keyword analysis, RSA generation, and weekly reporting (execution). The human reviews anomalies flagged by Claude Code and decides on budget shifts (learning loop). Smart Bidding handles bid adjustment autonomously.
Four documented failure modes: agents executing without sufficient guardrails (brand safety issues, compliance violations), humans reviewing too much and eliminating the speed advantage, the learning loop breaking when AI surfaces data but no human interprets it, and objective drift when AI optimises for measurable proxies rather than the actual goal.
Automation rules follow fixed if-then logic. The human-agent split framework uses Claude Code's judgment to handle novel situations within defined boundaries. An automation rule pauses campaigns when CPL exceeds a threshold. A Claude Code agent investigates why CPL spiked, identifies the probable cause, and recommends a specific action.
Pick one repetitive marketing analysis you do manually. Define the objective, set the guardrails, and build the agent workflow with Claude Code.
Install Claude Code