← Research library

The autonomy ladder: five levels of agent independence in marketing

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Agentic marketing programs now assign agents to research, content, lifecycle, paid media, and operations work. The question that follows every new agent is how much it should do without a person in the loop. Debates about whether agents should run marketing on their own tend to stall because they treat autonomy as a single switch.

In practice, autonomy is a dial, and different workflows deserve different settings. A shared vocabulary helps teams decide those settings, write them down, and revisit them as evidence arrives. We use a five-level ladder, shown in our autonomy ladder diagram.

Why a single switch fails

When autonomy is treated as on or off, teams end up in one of two places. Some keep every agent at a draft-only setting long after it has proven reliable, so people spend their days approving routine work. Others give an agent full independence after a good demo, and learn its limits from customers.

Both outcomes come from the same gap: no agreed way to describe the middle. A ladder fills that gap. Each level defines who acts, what control people keep, and what evidence is needed to move up.

The five levels

The levels run from no agent involvement to independent action inside hard limits. Each rung adds independence and requires more evidence than the one below it.

Level 0: manual

People do the work. AI may be used as a general assistant, but no agent is assigned to the workflow. This is the baseline every agent is measured against.

Level 1: assist

The agent produces drafts, research, or suggestions. People decide what to use and do all execution, and nothing reaches a customer or a system of record without a person putting it there. Good fits include first drafts, research briefs, and idea generation.

Level 2: approve to act

The agent prepares a specific action, such as a scheduled post or an audience update, and a person approves or rejects each one. On approval, the system executes exactly what was approved. Good fits include social scheduling, list updates, and routine campaign changes.

Level 3: act and notify

The agent acts on its own within defined limits and tells a person what it did. People can reverse the action but do not approve it in advance. Good fits include bid adjustments within a narrow range, internal reporting, and tagging and enrichment.

Level 4: autonomous within guardrails

The agent acts without per-action notice, inside hard limits. People review samples and exceptions on a schedule. Good fits are high-volume, low-consequence tasks with a strong evaluation history, such as request routing and data hygiene.

Where hard guardrails begin

From Level 2 upward, hard guardrails should be enforced in code, outside the model. At Level 2, a person approves each action, but the tool should still refuse an action that breaks a limit, in case an approval is rushed. At Levels 3 and 4, those limits are the main protection, because no person sees each action before it runs.

A rule in an agent's instructions is a request. A spending cap enforced by the tool connector is a control. Our post on guardrails as code covers how to encode brand, legal, and compliance rules so the agent cannot talk its way past them.

Permissions should follow the level too. An agent at Level 2 needs the right to prepare actions, not to execute them alone. Our post on roles and permissions for a mixed team explains how to match grants to the level an agent holds.

How agents climb the ladder

Moving up should require evidence, not enthusiasm. We recommend promoting one level at a time, and only when all of these conditions hold.

  • Time at level: the agent has run at its current level for a defined period.
  • Error rate: its error rate in review is below a threshold the team agreed in advance.
  • Worst case: the worst mistake it made would have been acceptable at the next level.
  • Written decision: an owner has made the promotion decision and recorded it.

The worst-case check is the one teams skip most often. An agent can have a low error rate and still have made one mistake that would have been serious without a reviewer. That mistake tells you more about readiness than the average.

Running an agent in shadow mode is a useful way to gather evidence before a promotion. The agent proposes actions at the next level while people continue to act, and the team compares the two.

Stepping down

Agents can also move down, and the ladder only works if stepping down is routine rather than a crisis. We suggest two triggers. The first is a spent failure budget: the agent has exceeded the error allowance the team set for its level. Our post on failure budgets covers how to set that allowance.

The second trigger is change. A new model, a new data source, or a redesigned workflow is a good reason to step back a level and re-earn trust, even if nothing has gone wrong yet. The evidence that justified the old level was gathered under conditions that no longer hold.

A worked example

Consider a hypothetical B2B software team with an agent that writes short descriptions for new webinar listings. It starts at Level 1, producing drafts that a content marketer edits and publishes by hand.

After a quarter at that level, the team reviews the record. Most drafts needed only light edits, and the worst error was a wrong speaker title, caught in review. The owner promotes the agent to Level 2: it now prepares the listing in the CMS, and a person approves each one before it goes live.

Two months later, the team switches the agent to a different model. Under the step-down rule, it returns to Level 1 for a few weeks while reviewers confirm the new model's drafts meet the same standard. The decision and its reason go into the run log.

Matching autonomy to consequence

The right level depends less on how capable the agent is and more on what happens when it is wrong. A capable agent writing legal claims should probably stay at Level 1 or 2. A modest agent tagging inbound leads can reasonably sit at Level 3.

The useful question is who would be affected if the agent got it wrong overnight, and how quickly the team could fix it. If the answer is many customers and slowly, the agent belongs lower on the ladder regardless of its record.

We are now studying how levels should work for agents that handle several workflows at once. One agent may deserve Level 3 for internal tagging and Level 1 for customer copy, and we want to know whether teams should split such agents or assign levels per task.

How to apply this: assign every agent in your program a level from 0 to 4 and write it down. For each one above Level 1, note the evidence that justified it, the guardrails enforced in code, and the condition that would move it back down.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.