Marketing
As of August 2026. Phase 6. High agent adoption, high disclosure exposure, and a homogenisation question nobody has settled.
1. The scenario
A marketing function runs content production, campaign operations, lifecycle and email programmes, performance analysis, and brand governance. Content volume is the perennial constraint and agents relieve it immediately, which is exactly why this department needs governance ahead of enthusiasm.
2. Agent team design
- Agent
- Content drafting agent
- What it does
- Produces drafts against brief, brand corpus and prior performance
- A x L position
- A2, L1
- Notes
- Human edit before publication is the control, not a courtesy
- Agent
- Campaign operations agent
- What it does
- Assembles, schedules and QAs campaign mechanics across platforms
- A x L position
- A3, L1
- Notes
- The best-evidenced win: mechanical, bounded, verifiable
- Agent
- Performance analysis agent
- What it does
- Explains campaign performance against history with provenance
- A x L position
- A2, L2
- Notes
- Explains; the marketer decides
- Agent
- Localisation agent
- What it does
- Adapts approved content per market against local rules
- A x L position
- A2 to A3, L1
- Notes
- Local disclosure and labelling rules vary sharply
3. Planes activated
Knowledge (direct: brand and claims corpus with an owner), Human (direct: publication approval), Evidence (direct: synthetic-content labelling), Improvement (direct), Action, Control, Execution (supporting).
4. Controls
- Synthetic content labelling is a live obligation in more than one jurisdiction. China's AIGC labelling took effect in September 2025 and EU Article 50 transparency duties are enforceable from 2 August 2026. Labelling is a pipeline capability, not a policy statement.
- A claims corpus with a named owner. Marketing agents generate claims, and unverifiable claims about a product are a regulatory exposure separate from AI law.
- Publication approval is a human gate. Nothing an agent writes reaches an audience unreviewed.
- No agent presents as a named human in any channel.
5. Economics
Per run. Low per asset, and volume is the risk rather than unit cost: the cheapest thing to do with a content agent is produce more content than anyone can govern.
Per resolved outcome. Cost per published asset including review time, and review time is the term that grows. A programme that reports asset cost falling while review capacity is unchanged is describing a queue, not a saving.
6. Honest limits
- The homogenisation question is unresolved. Studies credibly disagree on whether the homogenisation effect observed at output level aggregates to population level. The guide publishes the contest rather than a verdict, and the practical implication is to measure output diversity rather than assume it.
- Slide and asset generation was faster and markedly worse in observed-task testing. Speed on creative output is not the metric.
- No independent field measurement exists of production content quality effects at enterprise scale.
7. Metrics
Published assets per reviewer per week, which is the real capacity measure. Review rejection rate. Output diversity, measured rather than assumed. Labelling coverage, which should be 100% and is worth alerting on. Claims-corpus freshness.
Sources
research/R11-governance-risk-sovereignty/ (labelling and transparency), research/R08-productivity-and-collaboration/ (task-level results, homogenisation contest), research/R09-experience-and-channels/.
Source: blueprints/departments/marketing.md in the evidence repository behind this site.