Direct answer: first allow the agent to recommend a campaign and assemble drafts from approved sources. Do not let it publish or change audience, channel, frequency, timing, or spend until each decision has a commercial objective, current inputs, a boundary, an approver, telemetry, and rollback. If the team cannot explain why a change improved revenue or customer experience—and who owns the failure—the agent should remain in recommendation or shadow mode.
The capability shift is current. HubSpot uses goals, audience, briefs, historical performance, brand context, and CRM data to plan campaigns and create draft assets. Salesforce documents agentic campaigns that select campaign, channel, content, and timing per customer with configurable autonomy. Microsoft offers a preview agent that scaffolds journeys from existing segments, triggers, and content while requiring explicit review and publication. The license arrived before most operating models.
Write the Campaign Agent Operating Envelope
- Outcome — one commercial decision and one countermetric: conversion, margin, pipeline, or retention alongside opt-out, complaint, waste, or saturation.
- Context — brief, audience, offer, proof, brand, content, history, and CRM; source, version, expiry, and owner.
- Authority — what it may recommend, configure, publish, and optimize; prohibited fields, channels, markets, and verbs.
- Boundaries — spend, bid, volume, frequency, audience size, schedule, claims, exclusions, and duration.
- Approval — who releases concept, facts, privacy, audience, spend, and exceptions; when approval expires.
- Evidence — baseline, comparator, outcome event, logs, cost, and reading window.
- Stop — pause conditions, rollback, manual fallback, communication, and on-call owner.
This envelope is narrower than a general action contract because it binds the interdependent decisions of one campaign and defines which combinations may change live. A technical contract is still required for each CRM or media write, as explained in https://makinai.co/insights/en/ai-agent-crm-action-contract-permissions. The envelope decides whether the write is commercially appropriate for the objective and the customer.
Use four release levels
- Level 1 — recommendation: the agent analyzes and proposes; the team configures.
- Level 2 — draft: the agent assembles campaign, assets, and journey in an unpublished environment; accountable owners approve.
- Level 3 — bounded launch: it publishes only a predefined cohort, channel, and budget; material changes return for approval.
- Level 4 — bounded optimization: it chooses among approved options inside ranges, with monitoring, automatic pause, and periodic review.
Do not promote the agent because it has been available for a month. Promote it when the previous level clears release gates: complete and current inputs; low material-correction rate; no claim, audience, or channel outside scope; reconciled logs; readable outcome and countermetric; and tested rollback. Strong draft quality does not establish readiness to reallocate spend.
Separate prelaunch from live decisions
- Prelaunch — objective, eligible audience, offer, proof, proposition, channels, content, experiment, budget, and dates. These normally require accountable approval because they define strategy, promise, and exposure.
- Live — selection among approved variants, timing inside a range, frequency, saturation pause, and small reallocation under a cap. These can be automated when events, limits, and comparison are reliable.
- Always blocked — inventing a claim, removing suppression, expanding data purpose, creating a sensitive audience, breaching a cap, changing market, or declaring success without the specified comparator.
Illustrative US example
A hypothetical midsize B2B company with CRM, an agency, and three channels wants faster demand-generation launches. Its current cycle is 14 business days, and 24% of assets return because the offer or evidence is wrong. The figures are illustrative, not a benchmark or MAKINAI client result. The agent receives a structured brief and current assets; during the first cycle it only drafts. In the second, it may launch a 10% cohort under a spend cap without changing audience rules. Only later may it choose timing and variant from approved options.
The project tests two separate hypotheses: whether the agent reduces cycle time and rework, and whether the bounded campaign improves qualified conversion or cost per opportunity against a comparator. Faster production with unchanged quality or business performance is an operating efficiency—not automatic evidence for broader autonomy.
Run a six-week pilot
- Week 1 — select one campaign family, reconstruct three runs, and baseline time, cost, rework, outcome, and harm.
- Week 2 — write the envelope, sources, owners, approved options, limits, and stop cases.
- Week 3 — configure access, events, logs, environment, tests, and rollback; operate in recommendation mode.
- Week 4 — create drafts and compare completeness, corrections, and elapsed time with current operations.
- Week 5 — launch one bounded cohort with approval, comparator, and daily monitoring.
- Week 6 — read operations, experience, and business evidence; hold the level, reduce authority, expand one boundary, or stop.
Cost grows with platforms, channels, brands, markets, journey states, variant volume, CRM quality, content libraries, integrations, permissions, credits, environments, evaluation, observability, review, and support. Marketing owns objective, budget, and promise; RevOps owns identity, eligibility, and suppression; creative owns the system and quality; media and martech own configuration; analytics owns events and comparison; engineering owns integration; privacy and legal specialists own their domains. The agent cannot be the implied owner between teams.
Measure three layers and state attribution limits
- Operations — cycle, hours, rework, cost per released campaign, error, and recovery time.
- Experience — frequency, saturation, complaint, opt-out, suppression breach, and inappropriate contact.
- Business — qualified conversion, pipeline, contribution, retention, or lift against a comparator.
- Control — proposed, approved, rejected, reversed, blocked, and out-of-envelope changes.
Attribution remains limited. Before-and-after analysis mixes seasonality, offer, media, and selection. Platform metrics describe what the platform attributed, not what the agent caused. Use holdouts, geographic tests, or audience splits when feasible. When volume, platform support, or operations cannot support them, disclose uncertainty and do not relabel correlation as incremental return.
Use https://makinai.co/insights/en/copilot-workflow-ai-agent-marketing-what-build-first to choose the architecture and https://makinai.co/insights/en/ai-campaign-production-workflow-brief-to-launch to design brief-to-launch production. This article starts after the platform already offers an agent: the decision is how much of the campaign it may operate. To scope a real pilot, bring MAKINAI at https://makinai.co/services/en/ai-strategy-transformation-consulting one brief, campaign history, approval map, and the business decision that must improve.