Implement the least autonomy that removes a demonstrated constraint. Use a copilot when a person must still interpret, decide, and act; use an AI workflow when the path is known and can combine rules, models, integrations, and approvals; use an agent only when next steps vary, require dynamic tool use, and fit within explicit boundaries. If a workflow solves the problem, more autonomy is additional cost and risk—not maturity. This prevents buying an agent for a predictable sequence or trapping AI in suggestions when value depends on moving through systems.
Three operating models, not a prestige ladder
- Copilot — recommends, summarizes, classifies, or drafts; a person selects the answer and takes the action.
- AI workflow — follows predefined stages; models handle variable tasks while rules, integrations, and approvals govern the path.
- Agent — receives an objective, decides part of the sequence, and selects tools within defined permissions, budgets, stops, and evaluation criteria.
The distinction is operational. OpenAI describes agents as systems in which a model controls workflow execution. Anthropic distinguishes workflows orchestrated through predefined code paths from agents that dynamically direct their process. Both recommend starting simply and adding complexity when evaluation shows a benefit. Marketing leaders should apply that principle to the work instead of importing an architecture because its label is popular. The question is how much authority improves time, quality, conversion, or revenue—and how much the company can supervise.
Complete the Autonomy Fit Contract
- Outcome — the decision or deliverable to improve and the baseline used for comparison.
- Variability — the number of real exceptions and paths; low frequency does not justify open architecture.
- Judgment — choices that require context, ambiguity resolution, or human preference.
- Tools — data, CRM, media, content, and systems that must be read or changed.
- Authority — what AI may recommend, prepare, execute, or publish.
- Reversibility — how to undo an action, cap impact, and preserve evidence.
- Feedback — signals that reveal correctness, failure, and business effect.
- Operations — owner, review, exception queue, monitoring, cost, and incident response.
Choose based on the weakest part of the contract. High variability with low reversibility does not call for an agent; it calls for decomposition, approval, and a test environment. A stable path with repeated handoffs favors a workflow. Creative or analytical work where an accountable owner compares alternatives favors a copilot. An agent makes sense only when path freedom is necessary and harm is bounded. Google PAIR recommends calibrated expectations, reversible experimentation, feedback, and fallback behavior; failure becomes an operating condition rather than a surprise.
Marketing examples and authority
A copilot can turn approved research into brief alternatives and flag missing evidence; a strategist decides. A workflow can take an approved brief, generate variants, validate fields and claims, route review, localize, and record the released version. A research agent can explore authorized sources, reformulate searches, and prepare an evidence dossier, but it should not publish or change media budgets without authorization. Sending messages, changing eligibility, releasing spend, or writing to a CRM requires authority proportional to impact. Useful systems are often hybrid: agent in a sandbox, deterministic workflow for execution, and a person for brand, money, or customer decisions.
A six-week progressive pilot
- Week 1 — select one decision and baseline current time, quality, rework, volume, cost, and outcome.
- Week 2 — complete the contract, success and failure examples, sources, tools, authority, and stop conditions.
- Week 3 — operate as a copilot; record acceptance, editing, rejection, and time saved.
- Week 4 — turn repeatable stages into a workflow with validation, queues, logs, and approval.
- Week 5 — allow the agent to choose paths in shadow mode only; compare with actual operations.
- Week 6 — release a small, reversible cohort with spend limits, approval, and rollback; retain, reduce, or expand.
In shadow mode, the system sees the same inputs and proposes actions without executing them. This reveals unnecessary calls, loops, weak sources, and exceptions before customers or budgets are affected. Include normal, edge, and adversarial cases. Compare against both a simple rule and the non-AI flow; do not assume more reasoning outperforms a well-designed sequence. The NIST profile reinforces lifecycle evaluation, monitoring, and accountability, but it does not replace business-specific policy.
Measure utility, control, and economics
- Utility — acceptance, correction, completeness, and evidence adherence.
- Flow — time to decision, handoffs, rework, queue time, and used throughput.
- Control — permission violations, tool failures, escalation, rollback, and traceability.
- Economics — model, integration, review, and operating cost per approved deliverable.
- Business — conversion, revenue, margin, retention, or incremental learning tied to the use case.
- Experience — complaints, opt-outs, brand consistency, and burden on employees and customers.
Cost rises with tool count, case diversity, write permissions, latency, volume, context, model calls, environments, evaluations, observability, and support. An inexpensive agent call becomes costly when it loops or demands extensive review. A rigid workflow may cost more to integrate and less to operate. Calculate cost per approved and used outcome, not per generated word. Do not begin with multiple agents: they add coordination, state, conflict, latency, and failure surfaces before the organization proves an open-ended decision is necessary.
Accountability and the U.S. context
Marketing owns outcomes, examples, and brand policy. Marketing operations owns stages, queues, and exceptions. Data owns access, quality, and measurement. Engineering owns identity, integrations, environments, and secrets. Legal, privacy, and security determine permitted uses; finance validates cost and value. U.S. teams should account for state, federal, contractual, and sector requirements plus fragmented martech ownership. Vendor access, consumer data, claims, accessibility, channel policy, and record retention may require different controls. A consultancy can implement; the company retains authority over publishing, spend, and customer contact.
The architecture decision starts with the work
Use the 90-day plan at https://makinai.co/insights/en/where-start-ai-marketing-90-day-first-workflow, diagnose value leakage at https://makinai.co/insights/en/marketing-team-uses-ai-results-unchanged-value-leak, organize knowledge at https://makinai.co/insights/en/ai-marketing-copilots-inconsistent-answers-knowledge-layer, and compare the campaign workflow at https://makinai.co/insights/en/ai-campaign-production-workflow-brief-to-launch. MAKINAI can map and implement the first bounded flow: https://makinai.co/services/en/ai-strategy-transformation-consulting. Bring one decision, its baseline, and the systems involved; the conversation starts with required autonomy, not a software category.
The six-week pilot is a scope frame, not a delivery or ROI guarantee. Ambiguity, data, integrations, and policy may require a simpler or longer design. The sources support architecture and control principles; the Autonomy Fit Contract is a MAKINAI editorial recommendation to validate in the real workflow.