Human-Controlled AI Marketing Role SOP / Operating Playbook

Model Role SOP: Human-Controlled Generative and Agentic AI for Marketing

1. Purpose and scope

This operating playbook defines how an AI-Enabled Marketing Workflow Lead moves from a marketing question to a controlled decision. It covers research, segmentation, content operations, campaign planning, testing, measurement, and agent supervision. It is vendor-neutral and should be adapted to applicable law, contracts, organizational policy, sector rules, and system capabilities.

2. Non-negotiable control principle

AI may retrieve from approved sources, structure evidence, generate drafts, propose segments or plans, perform authorized low-risk checks, and assemble reports. Accountable people retain authority over public release, customer contact, audience activation or expansion, paid spend, pricing, sensitive or inferred data, new system access, data export or deletion, destructive actions, purchasing, contracting, legal interpretation, and other external commitments.

The phrase “human in the loop” is not sufficient. Every consequential checkpoint must name the reviewer, evidence required, threshold, available decisions, and escalation route.

3. Intake and use-case selection

Open one workflow record containing:

• business decision and accountable sponsor;
• intended users and affected people;
• geography, channels, and timeframe;
• approved data and source locations;
• expected output and downstream action;
• baseline, success metric, guardrails, and decision rule;
• volume, frequency, and service expectation;
• reversibility and cost of failure;
• governance owners and required approvals.

Score the use case on decision value, repeatability, evidence quality, data readiness, integration complexity, review burden, rights/privacy sensitivity, failure severity, reversibility, and measurement feasibility. Start with high-value, repeatable, reversible internal preparation work. Reject or escalate any request involving political persuasion, children, protected or sensitive inference, unlawful enrichment, deceptive synthetic media, high-impact eligibility decisions, or a requirement for unattended consequential action.

4. Research procedure

1. Convert the decision into answerable questions.
2. Define source tiers, date range, geography, exclusions, and stopping criteria.
3. Retrieve only authorized material; record title, organization, URL or internal location, date, access date, evidence type, and rights status.
4. Separate direct observation, source claim, analyst inference, hypothesis, and unknown.
5. Check material claims against an original or authoritative source.
6. Compare conflicting evidence and record why one source receives more weight.
7. Produce a decision brief with findings, limits, contrary evidence, and next action.
8. Require human review before findings become an audience, public claim, or investment decision.

Synthetic respondents and simulated audiences may suggest questions or scenarios. Label them as simulations and compare them with observed or primary evidence. Never report them as real customer research.

5. Segmentation and audience procedure

For every proposed segment, record:

• business purpose and decision;
• permitted evidence and source date;
• explicit inclusion and exclusion rules;
• consent, contact preference, suppression, and expiry rules;
• sensitive attributes or proxies reviewed and prohibited;
• minimum size and re-identification risk;
• proposition and expected behavior;
• confidence and bias risks;
• channel destination and authorized owner;
• primary and guardrail metrics.

An agent may prepare a segment hypothesis or detect inconsistent rules. A human data or marketing owner approves the rules and a privacy/compliance owner reviews sensitive or ambiguous cases. Activation occurs only through an approved system and identity. Stop if consent, lawful basis, suppression, sensitivity, or destination permissions are unclear.

6. Content operations procedure

Create a governed content brief before generation. It must specify objective, audience, journey stage, proposition, approved evidence, mandatory facts, prohibited claims, tone, accessibility, channel constraints, disclosure, rights, owner, reviewers, and expiry.

Generate only within the brief. Store every material variant with a unique identifier, generation or authoring context, evidence version, owner, review state, intended channel, test cell, and outcome. Reviewers check:

1. factual accuracy and substantiation;
2. brand and message fit;
3. consumer understanding and absence of deception;
4. rights, likeness, voice, and third-party material;
5. disclosure and material-alteration requirements;
6. accessibility and channel rules;
7. audience and offer consistency;
8. tracking and experiment readiness.

Only the named release owner can approve publication. Retire or re-review assets when evidence, product, price, policy, rights, or campaign context changes.

7. Campaign planning procedure

Create a campaign charter with objective, business outcome, audience, insight, proposition, offer, channel roles, journey, budget and frequency guardrails, creative requirements, dependencies, tests, metrics, owners, approvals, and exceptions.

Map handoffs among Marketing, Sales/RevOps, Product, Customer Success, Data/Engineering, Creative, Finance, Legal/Privacy, Security, and agencies. State the input, output, owner, service expectation, and escalation for each handoff.

An AI system may assemble a draft plan or recommend a change. A person checks evidence, audience logic, claims, brand, privacy, cost, expected value, test design, and operational capacity. Budget, audience, outbound, and release actions require explicit approval at the point of action.

8. Testing and AI evaluation procedure

Before deployment, define:

• hypothesis and expected mechanism;
• unit of assignment and comparator;
• primary metric and practical decision threshold;
• guardrail metrics;
• sample, duration, seasonality, contamination, and stopping constraints;
• segment checks and minimum reporting cells;
• tracking and version identifiers;
• scale, revise, stop, or investigate rule.

For AI outputs or agents, add a test set that includes normal, edge, failure, and adversarial cases. Score factuality, completeness, policy compliance, brand fit, stability, escalation, latency, and cost. Record false positives, false negatives, reviewer disagreement, and unacceptable failure classes. A passing average cannot compensate for a critical safety failure.

9. Agentic workflow specification

No agent connects to operational systems until the following fields are approved:

• bounded purpose and named accountable owner;
• trigger and termination event;
• context sources and prohibited data;
• tool allowlist and per-tool permission;
• actions the agent may prepare, simulate, or execute;
• actions that are always prohibited;
• validation steps and evidence requirements;
• human checkpoints and authorized decisions;
• log schema and retention;
• cost, volume, error, drift, and complaint thresholds;
• stop conditions and exception routing;
• manual fallback and rollback;
• test evidence and release version.

Use least privilege. Separate read, draft, propose, approve, publish, send, spend, export, delete, price, and commit permissions. The default production pattern is proposal-only for consequential marketing actions.

10. Measurement and decision procedure

Maintain a metric dictionary covering name, business meaning, formula, numerator, denominator, time zone, source, owner, update cadence, exclusions, quality checks, and known limitations. Build a KPI tree that connects workflow activity to customer and business outcomes.

Reproduce material calculations. Check filters, joins, campaign taxonomy, attribution window, duplicate events, missing consent, currency, margin, returns, and cohort timing. Separate attributed credit from incremental effect. Use a holdout, baseline, or credible comparator where feasible.

Issue a Measurement Decision Memo containing:

• decision requested;
• treatment or workflow version;
• comparator and timeframe;
• primary and guardrail results;
• uncertainty and data-quality notes;
• segment and failure analysis;
• operational cost and review burden;
• recommendation: scale, revise, stop, or investigate;
• owner, approval, next review, and rollback state.

11. Operating cadence

Daily

Review exceptions, failed validations, cost or spend anomalies, queued approvals, data freshness, tracking failures, complaints, and rollback readiness. Do not normalize repeated exceptions; investigate their root cause.

Weekly

Review research gaps, segment and suppression health, content queues, campaign pacing, experiment integrity, agent evaluations, cross-functional handoffs, and decisions due. Record owners and dates for every action.

Monthly

Review business outcomes, incrementality evidence, model and workflow drift, permissions, vendor changes, content retirement, complaint patterns, data quality, and portfolio priorities. Retire low-value automation rather than preserving it for sunk-cost reasons.

Quarterly

Reassess use-case value, policy, risk, access, contracts, measurement validity, skills, operating ownership, and the 2027-readiness roadmap. Test emergency stop, manual fallback, and rollback.

12. Exception and escalation matrix

13. Evidence retention

Retain the approved brief, source ledger, data and permission record, workflow version, prompts or instructions where policy permits, tool and model identifiers, input/output hashes where appropriate, evaluation results, human approvals, execution record, metric definitions, decision memo, exceptions, and rollback evidence. Apply organizational retention and deletion policy; do not keep personal or confidential data merely because an AI system used it.

15. Closing standard

A marketing AI workflow is ready only when the team can answer: what decision it supports; what evidence it used; what the system was allowed to do; what a person checked; how success and harm are measured; when it stops; how it rolls back; and who owns the consequence. Speed without those answers is not 2027 readiness.