THE SHORT ANSWER
Marketing systems can combine fixed rules, automated workflows, machine-learning predictions, generated material and model-directed actions. These approaches coexist. Choose the least complex arrangement that serves the goal, and define permissions, checks and recovery before increasing autonomy.
The evolution adds options; it does not retire rules
| Layer | Illustrative job | What determines the action? |
|---|---|---|
| Rule | Exclude people who opted out | An explicit condition |
| Automation | Run approved steps after an event | A predefined sequence |
| Machine learning | Estimate which cases need attention | A learned prediction |
| Generative AI | Draft a response from approved evidence | Generation guided by supplied context |
| Agent | Choose which permitted investigation step to take next | A model responding to observations |
This is a conceptual progression, not a claim that every organisation follows the same history. A robust system may use all five. Explicit consent and suppression rules should not become optional just because another part of the process can make flexible decisions.
The Phase 1 agent explanation uses Anthropic's distinction between predefined workflows and model-directed steps. That distinction is more useful than a vendor's maturity label.
Evidence & context: Anthropic
Follow one customer request through the system
Imagine a customer asks whether a service supports their equipment. A rule can route the request. An automated workflow can retrieve the relevant approved specification. A model can draft a clear answer, and an agent might choose to seek a missing model number before proceeding. None of those capabilities authorises an unsupported compatibility promise.
Start by producing a draft for review. Test ambiguous equipment names, outdated specifications, missing documents and a failed lookup. The correct result may be a clarification or escalation rather than a completed answer. Only consider broader permissions after the bounded process is useful and its failures are understood.
The handoffs are part of the product
Who owns a case after escalation? What happens if the same event arrives twice? Can a partially completed action be retried without sending duplicate messages? How does the system behave when an approval expires? A workflow diagram that omits these questions describes a demonstration, not an operating process.
- Separate permission to read from permission to edit, send or spend.
- Keep records of important inputs, approvals and actions with appropriate access and retention.
- Use explicit limits and stop conditions for repeated or unsuccessful steps.
- Design a human fallback and a way to reverse or correct mistakes.
Evidence & context: NIST
Measure the complete process
Time saved on drafting can disappear in verification and rework. Compare the total effort, resolution quality and customer experience with the existing process. Review ordinary cases and difficult cases separately so a good average does not conceal an unacceptable failure.
Use the governance guide to assign responsibility. Detailed implementation of agents and marketing automation belongs in a later curriculum; the first decision is which work needs discretion at all.
Sources & further reading
- Building effective agents
Anthropic. A provider's engineering taxonomy of agents and workflows, not a universal industry definition. We use the conceptual distinction, not its changing product recommendations.
- Generative Artificial Intelligence Profile (NIST AI 600-1)
NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗