TL;DR: AI agents will take over most marketing execution, including campaign setup, creative testing, and budget allocation. The CMO's role shifts from managing people to managing the system. That means setting intent and guardrails, deciding what escalates to a human, and protecting the long-term brand. Autonomy is earned, not granted.

When I wrote Lean AI (O'Reilly, 2020), the argument was simple. Small teams that use automation well can beat much bigger budgets. Since then, almost every growth engine I've run or advised has hit the same bottleneck. Not budget. Not tools. Human latency.

A campaign waits on a review. A budget shift waits on the monthly meeting. A creative test finishes on Tuesday and nobody reads the results until the following week.

Most of the industry's answer is the co-pilot. Give every marketer an AI assistant and let them move faster. That helps. It also keeps the slowest part of the system sitting at the center of it.

I think a bigger shift is coming. Within the next decade, the marketing function at many companies will run with close to zero human execution. The CMO won't manage a team of specialists. The CMO will manage the system.

Most "zero-human" predictions skip the hard part, though. Nobody should hand an AI system the keys because a vendor demo looked good. Autonomy has to be earned, one level at a time, with guardrails that prove the system deserves more control.

The Autonomy Ladder, Extended

In Lean AI, I adapted the SAE autonomy scale for self-driving cars into a scale for marketing systems. Here it is, with one new rung.

Level 0  No automation. Marketers run everything with basic tools. The CRM
         and dashboards store data and report results, nothing more.
Level 1  Recommendation automation. Systems follow marketer-set rules and
         recommend changes, like shifting spend by channel. A person still
         makes the change.
Level 2  Rules-based automation. The system makes those changes itself,
         with no approval. But the rules are rigid and a person writes them.
Level 3  Computational autonomy. Machine learning observes, learns, and
         improves outcomes. People only set goals and broad parameters.
Level 4  Insightful autonomy. Systems understand context and personalize
         1:1 messages across channels.
Level 5  Fully autonomous. Systems generate their own tests, creative
         variations, and targeting, with no ongoing intervention.
Level 6  Coordinated autonomy (new). Autonomous systems for allocation,
         content, discovery, and attribution share context, hand off
         work, and coordinate directly with each other.

Levels 0 to 5 are adapted from Lean AI (O'Reilly, 2020). Level 6 is new.

The original scale stops at Level 5, where a single system runs its own tests, creative, and targeting without the marketing team stepping in. But Level 5 systems still work mostly alone. The paid acquisition loop optimizes itself, but people still decide how it works with everything else. Level 6 is what happens when those separate systems stop handing work to people and start handing it to each other.

The basic plumbing is already here, even if few companies have connected it yet. The Model Context Protocol (MCP) connects agents to tools and data. The Agent2Agent (A2A) protocol lets agents from different vendors exchange tasks and context. Both run on JSON-RPC 2.0, which means agents can pass structured requests back and forth the way software services already do.

McKinsey has a name for the architecture layer underneath all this. In "The agentic organization," Alexander Sukharevsky, Lari Hämäläinen, and their McKinsey colleagues describe an "agentic AI mesh," where agent-to-agent protocols make integration across systems cheaper and easier. The mesh is the infrastructure. What I care about here is the operating question on top of it. How much should each part of that system be allowed to decide on its own?

What Level 6 Looks Like in Marketing

Here's how a growth initiative runs today versus at Level 6.

TODAY
CMO strategy → VP planning → team meetings → content creation
→ campaign setup → manual review → optimization

LEVEL 6
CMO sets intent and guardrails → agents plan, build, launch,
measure, and reallocate in a continuous loop → humans handle exceptions

Instead of teams organized by specialty, you get four kinds of agents working off the same data.

Allocation agents watch competitors, search visibility, and channel performance, then move budget without waiting for a quarterly review. That includes visibility inside AI answers, not just traditional search rankings.

Content agents generate variations, test them, and retire the losers based on live results instead of a two-week reporting cycle.

Discovery agents handle top-of-funnel conversations, collect what customers tell you directly, and qualify leads in real time.

Attribution agents compare results against baselines and feed that back into bidding and allocation.

The gain isn't just speed. It's that the loop never stops. A human team reviews performance on a schedule. An agent system reviews it continuously.

Where the Zero-Human Argument Breaks

If you stop there, you get the version of this story every vendor is telling. Here's what it leaves out.

Mistakes compound faster. A system that makes a thousand good decisions an hour can also make a thousand bad ones. Speed cuts both ways.

Optimization drifts. Agents chase the metric you gave them. If that metric is clicks, you can end up with a brand nobody recognizes and a funnel full of low-intent traffic.

Agents agree with each other. When one agent's output becomes another agent's input, errors can reinforce themselves. The allocation agent trusts the attribution agent, which is measuring traffic the allocation agent chose.

Accountability gets blurry. When one agent negotiates with another and something goes wrong, someone still has to answer for it. That someone is a person.

Proxies get gamed. Any number an agent is rewarded on is a number it will find a shortcut to. The shortcut rarely matches what you actually wanted.

None of this means Level 6 is a fantasy. It means the path there runs through governance, not around it.

What I Learned Building This at IMVU

At IMVU, starting in 2018, we built an AI-driven user acquisition system that automated bidding and targeting decisions people used to make by hand. We were spending more than $50 million a year on performance marketing, so the stakes were real.

We didn't turn it on everywhere. We launched it in one geography and gave it a fixed validation window against a real payback target.

It worked. ROAS tripled. CAC fell by a factor of three. We hit 100% payback on UA spend within 60 days.

But faster decisions also meant faster mistakes. So we built a separate, real-time governance layer on top. Every automated decision got a risk score from 0 to 10. Anything scoring 7 or higher automatically tripped a circuit breaker and pulled in a human before the error could spread.

In practice, most decisions ran on their own. The higher-risk ones got flagged and went to a person.

That layer didn't hold the system back. It's the reason we could trust it with more of the business. Full autonomy wasn't where we started. It was what the system earned.

The Autonomy Gates

That experience is where this framework comes from. I use the term "Autonomy Gates" for the checkpoints that decide how much authority an AI system earns at each stage. Before any system moves up a level, it has to clear three of them.

Gate 1. Prove it small. Run the new level in one geography, one segment, or one channel. Set a fixed validation window and a real business metric, like payback period, not a vanity metric like engagement. If it can't win small, it doesn't get bigger.

Gate 2. Score every decision. Give the system a risk score for each action it takes, and set the human-review threshold before launch, not after the first incident. The threshold is the guardrail. Decide where it sits while you're calm.

Gate 3. Promote on evidence, demote on failure. Widen the system's scope, or raise the threshold, only when the data shows it's earned it. And if it trips the breaker too often, move it back down a level. Autonomy should be able to go both ways.

Run every agent in your stack through these gates and Level 6 stops being a leap of faith. It becomes the last step in a sequence you've already proven.

What the CMO Actually Does

If the marketing function has no human executors, what's left for the CMO? More than before, and harder work.

Set intent and constraints. The main input to the system isn't a strategy deck anymore. It's goals, budgets, risk thresholds, and brand rules the agents can actually act on.

Design the system. Choose which agents to run, how they connect, and what data each one can see.

Own the escalation list. Decide in advance what always comes back to a person. Major brand changes. Crisis communications. Spend above a set threshold. Anything with regulatory risk.

Protect the brand. Agents are good at finding what works this week. They're not good at knowing what the brand should stand for in five years. That judgment stays human.

Speed and Guardrails Aren't Opposites

The biggest platforms are starting to say the same thing. A SiliconANGLE recap of Salesforce's Dreamforce 2026 conference described the company's pitch moving away from flashy agent demos and toward measurable, governed outcomes. Salesforce President and Chief Platform Officer Rohan Kumar put it directly: "Without having good guardrails, it's impossible to transition to [the agentic] enterprise."

BCG is making a similar case. In "Harness Engineering: The Operating System for Agentic AI," it argues that as models become interchangeable, the advantage moves to the system around them: specs, rules, audit trails, data, and quality gates.

That matches what I saw at IMVU. The teams that move fastest won't be the ones that remove humans first. They'll be the ones that build the guardrails first, so they can remove human execution safely.

The Timeline

Now. Most marketing teams are working somewhere between Levels 1 and 3. Spend recommendations and ML-driven bidding are common. Coordinated autonomy isn't.

Next. Levels 4 and 5 become normal inside specific functions, like paid acquisition, lifecycle messaging, and content testing, with people supervising.

After that. Level 6 becomes realistic for companies that built their gates early. Humans set direction, define intent, and handle exceptions. Agents handle the rest.

Companies will move through these stages at very different speeds, so the useful question is whether yours can earn Level 6 safely, whenever it arrives.

The CMOs who win the next decade won't be the ones with the biggest teams. They'll be the ones who knew exactly how much autonomy each system had earned, and could prove it.

FAQ

What is agent-to-agent (A2A) collaboration in marketing?

It's when AI agents hand tasks, data, and decisions directly to other AI agents through standard protocols like Agent2Agent, instead of routing every step through a person. For example, an attribution agent can tell an allocation agent to shift budget without waiting for a meeting.

What is Level 6 in the Lean AI autonomy scale?

The Lean AI scale, adapted from the SAE scale for self-driving cars, runs from Level 0 (no automation) to Level 5 (fully autonomous systems that generate their own tests, creative, and targeting). Level 6 is a new extension where autonomous systems for allocation, content, discovery, and attribution coordinate with each other directly, and people handle intent, guardrails, and exceptions.

What are the Autonomy Gates?

A three-step test a system should pass before it gets more control. Prove it small in one geography or segment. Score every decision against a risk threshold set before launch. Promote it on evidence and demote it when it fails.

Will AI replace the CMO?

No. AI will take over more execution, including campaign setup, creative testing, and budget allocation. The CMO's job shifts to setting intent, designing the system, owning what escalates to humans, and protecting the long-term brand.

How is Generative Engine Optimization (GEO) different from SEO in an agentic setup?

SEO is about ranking in search results. GEO is about getting your brand read, understood, and cited by AI answer engines and buyer agents during conversational search. In a Level 6 system, allocation agents track both and move budget against both.

Why not just go straight to full autonomy?

Because mistakes scale as fast as wins. A system that hasn't proven itself in a small, measured test, with a risk threshold that pulls in a human, can do a lot of damage before anyone notices. Guardrails are what make it safe to go fast.