AI governance for marketing teams

The autonomy ladder

Autonomy is set per surface and per location, so a single operation runs several of these at once. The useful question is never “what level are we at” but “which surfaces have earned the next one.”

1. Level 0 — Draft only

   The agent produces, a human publishes everything. Correct for a brand-new surface, and a permanent state only if the surface is genuinely high-stakes.

2. Level 1 — Review then publish

   Everything queues, but the queue is ordered by confidence so reviewers spend their attention where it matters. Most surfaces should sit here for the first month.

3. Level 2 — Auto-publish above a threshold

   Outputs scoring above a set number publish unattended; the rest queue. This is where the labour saving actually appears, and where most operators stop moving because raising the threshold feels risky without evidence.

4. Level 3 — Auto-publish with sampled review

   Everything publishes, and a random sample is reviewed after the fact to detect drift. Appropriate once you have months of scored history showing the gate holds.

5. Level 4 — Autonomous with exception handling

   The system publishes and self-corrects, escalating only anomalies. Few marketing surfaces need this, and claiming it before Level 2 has run for a quarter is a sales position rather than an engineering one.

Almost every operation that has deployed anything is stuck between Level 1 and Level 2, because moving requires evidence that the gate holds, and nobody set up the scoring that would produce that evidence. That is the single most common finding, and it is a measurement problem rather than a risk problem.

What you should not govern

• Internal drafts nobody outside the team will see. Governing these adds cost and removes the fast iteration that makes the system good.
• Word choice within an already-approved claim. If the claim passed, arguing about synonyms is a taste dispute wearing a compliance badge.
• Surfaces with no customer visibility and no regulatory exposure — internal summaries, backlog notes, drafts in a private queue.
• Every output equally. Uniform scrutiny means the high-risk output gets the same thirty seconds as the low-risk one, which is how real problems slip through.

Over-governance fails in a way that looks responsible, which is why it survives far longer than under-governance. A team that reviews everything equally is a team that reads nothing carefully.

Common questions

What is AI governance in a marketing context?

Five things, made operational rather than written into a policy document. What may be said — the claims allowed per vertical and jurisdiction, expressed as checks an output can be scored against. Who approves what — routing rules that send an output to auto-publish, review, or escalation based on a score rather than a job title. How much autonomy each surface gets, set per surface and per location rather than globally. What gets recorded — for every published output, the input, model and version, rules evaluated, score, routing decision, and any human touch. And what happens when it is wrong, including the feedback path that turns an override into a change to the system.

What are the levels of AI autonomy for marketing content?

Five, in practice. Level 0 draft only, where a human publishes everything. Level 1 review then publish, with the queue ordered by confidence so attention goes where it matters — most surfaces should sit here for a first month. Level 2 auto-publish above a threshold, which is where labour saving actually appears and where most operators stall because raising the threshold feels risky without evidence. Level 3 auto-publish with sampled post-hoc review to detect drift, appropriate once months of scored history show the gate holds. Level 4 autonomous with exception handling, which few marketing surfaces need and nobody should claim before Level 2 has run for a quarter.

How do you set an AI auto-publish threshold?

As a number in defined units, changeable on purpose by a named owner. Score every output on the dimensions that matter — brand fit, claim compliance, factual grounding against the master record — then set the auto-publish cut where your reviewed sample shows an acceptable error rate, and record why. The common failure is not a threshold set too high or too low; it is a threshold nobody can locate, in units nobody can explain, that therefore never moves. A gate that cannot be tuned becomes a queue, and a queue that grows gets approved in bulk without reading.

What should you not govern?

Internal drafts nobody outside the team will see, since governing them adds cost and removes the fast iteration that makes the system good. Word choice inside an already-approved claim, which is a taste dispute wearing a compliance badge. Surfaces with no customer visibility and no regulatory exposure. And crucially, not every output equally — uniform scrutiny means a high-risk output gets the same thirty seconds as a low-risk one, which is how real problems slip through. Over-governance fails in a way that looks responsible, which is why it survives longer than under-governance.

How do you tell an AI governance consultant from a policy vendor?

Ask to see a governance system running unattended today and its auto-publish rate. Ask what the threshold is, in what units, and who can change it. Ask what exactly is written to the audit log for one published output, and to see a record. Ask how an override becomes a change to the system rather than a one-time correction. Ask what they deliberately chose not to govern. A policy vendor answers these with frameworks and a document; a practice answers with numbers, a record, and a name.