Running an Agent-Assisted Marketing Team

Draft, execute, or keep human

Anthropic's engineering guidance Building Effective Agents (December 2024) distinguishes workflows — systems where models and tools are orchestrated through predefined code paths — from agents, where the model dynamically directs its own process and tool use. It recommends finding the simplest solution possible and increasing complexity only when needed, and warns that autonomous agents bring higher costs and the potential for compounding errors, recommending extensive testing and appropriate guardrails.

For marketing, that suggests three bands. Let the agent draft freely: research summaries, content outlines, task breakdowns, variant suggestions, first drafts of emails. Let it execute where the action is reversible, internal and low-stakes: creating tasks, tagging contacts, scheduling reminders, compiling reports. Keep a human approval on anything that leaves the company in the company's name — outbound email, published content, social posts, anything a customer or prospect will read — at least until you have a long record of the agent's drafts being approved unchanged. And keep some things human entirely: positioning choices, pricing decisions, anything presented as a person's experience or opinion.

The 2026 rules for AI-generated marketing

Two sets of rules bear directly on agent-assisted marketing in 2026.

In the EU, Article 50 of the AI Act sets transparency obligations. Among them: deployers of an AI system that generates or manipulates text published with the purpose of informing the public on matters of public interest must disclose that the text was artificially generated or manipulated — but that obligation does not apply where the AI-generated content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication. Deepfake image, audio and video content carries its own disclosure duty. Under Article 113, the Regulation generally applies from 2 August 2026. Most routine marketing copy is not 'public-interest' text, but the editorial-responsibility exception describes exactly the operating model this lesson recommends: a named person reviews and owns what is published.

In the US, the FTC's final rule on consumer reviews and testimonials (announced August 2024) prohibits creating or buying fake reviews and testimonials, explicitly including AI-generated ones and reviews attributed to people without real experience of the business, and allows civil penalties for knowing violations. An agent must never write a review, testimonial or quote in a customer's voice. Google's search guidance (Lesson 3.2) adds the content-side rule: AI assistance is fine, but generating many pages without adding value may violate its scaled content abuse policy.

Governing the agent

NIST's AI Risk Management Framework (AI RMF 1.0, January 2023) organizes AI risk management into four functions — Govern, Map, Measure and Manage — and is explicit that its actions are not a checklist. For a marketing team the translation is small and concrete. Govern: decide who may change the agent's permissions and policies, and write the boundaries down. Map: list which marketing actions the agent can take and what could go wrong with each (wrong recipient, unsupported claim, opt-out ignored, off-brand tone). Measure: review what it actually did, and how often its drafts are approved unchanged. Manage: tighten or loosen autonomy based on that record.

In Kavanah those functions map onto real controls. The workspace sets an autonomy level for the agent (off, assess, safe, full); safe mode removes send tools entirely. In full mode, actions in the 'send messages and email' category — sends and cadence enrolment — follow the workspace's action policy, and the default when no policy is set is to park the action for human approval rather than run it. Parked actions are reviewed in Settings. Automation rules that invoke the agent carry their own send mode: approval parks every send, and the auto mode is reserved for admins. The charter's Negations are enforced: the agent refuses to create tasks that contradict them. AI employees (personas) can be limited to tool scopes — a content persona need not have the email-and-calendar scope at all — and letting AI employees work tasks assigned to them is an opt-in setting, off by default.

One more governance input is cost. Agent work consumes credits; research on a contact, for example, is billed per contact. Point it at work that passes your ICP, not at whole lists.

Running the team

Treat the agent like a capable new hire on probation. Give it the context a person would need — the ICP, the message hierarchy, the brand standard, the research — in Knowledge and the charter, because it drafts from what it can read. Review its work on a schedule rather than only when something goes wrong. Expand its autonomy one action type at a time, when the approval record justifies it, and record the decision. And keep a human name on everything that goes out: the editorial-responsibility principle is good law in the EU and good practice everywhere, because it keeps someone accountable for whether what the company says is true.