Chapter 18: Agentic Marketing: Designing and Setting Up AI Agents

Learning objectives

By the end of this chapter, you will be able to:

• Explain what an AI agent is and how it differs from a chatbot or a traditional automation.
• Choose the right level of autonomy for a marketing task.
• Describe the components of an agent: model, instructions, tools, knowledge, memory, and guardrails.
• Design agentic workflows for SEO, paid media, content, email, social, and analytics.
• Set up an agent step by step using no-code tools or AI assistant platforms.
• Test, evaluate, monitor, and improve agents safely.

From assistants to agents

So far in this series, AI has mostly been a conversational assistant: you ask, it answers, you decide what to do next. An AI agent goes further. Given a goal, an agent can plan the steps, use tools (search the web, read a spreadsheet, query analytics, draft a document, update a record), check its own progress, and continue until the task is done, involving a human at defined points.

	Traditional automation	AI assistant	AI agent
How it works	Fixed “if this, then that” rules	Responds to one request at a time	Pursues a goal across multiple steps, choosing actions
Handles ambiguity	No	Yes, in conversation	Yes, within its instructions and tools
Uses tools	Predefined actions only	Some (search, files, code)	Many, chosen as needed
Human role	Designs the rules	Directs every step	Sets the goal, approves key actions, reviews results
Example	“When a form is submitted, add the contact to the CRM”	“Write five subject lines for this email”	“Every Monday, analyze last week’s campaign data, find anomalies, draft a report with recommendations, and send it to me for approval”

In practice, most useful marketing “agents” today are agentic workflows: structured, multi-step processes where AI handles the steps that need judgment or language, traditional automation handles the predictable steps, and humans approve anything consequential. Fully autonomous agents that spend money or publish content without review are rarely a good idea in marketing, where brand and legal risks are high.

The autonomy ladder

Not every task deserves the same autonomy. Use this ladder to decide how much independence an agent should have.

Level	Description	Marketing examples	Human role
1. Assist	AI drafts or suggests; a human does everything else	Ad copy suggestions, report drafts	Decides and acts
2. Recommend	AI analyzes and recommends specific actions	Negative keyword recommendations, budget shift proposals	Approves each action
3. Act with approval	AI prepares the action and executes it after a human approves	Scheduling approved social posts, updating CRM fields, launching a pre-approved email	Approves in batches or individually
4. Act within limits	AI acts automatically within strict boundaries and reports what it did	Pausing ads with spend above a threshold and zero conversions, tagging content, routing leads	Sets limits, reviews logs, handles exceptions
5. Autonomous	AI pursues goals with broad discretion	Rarely appropriate in marketing today	Oversight and audit

Rule of thumb: the higher the cost of a mistake (money spent, content published to customers, data changed, legal exposure), the lower the autonomy level. Start every new agent one level lower than you think it needs, and raise autonomy only after it earns trust through evidence.

Anatomy of an agent

Every agent, whatever platform it runs on, has six components.

1. Model. The language model that does the reasoning (from OpenAI, Google, Anthropic, or others). Choose based on quality on your tasks, cost, speed, and your company’s approved vendors.
2. Instructions. The agent’s role, goal, step-by-step method, output format, rules, and when to stop and ask a human. This is the most important component, and it’s where your marketing expertise goes.
3. Tools. The actions the agent can take: web search, reading files, querying GA4 or Google Ads, reading and writing to a spreadsheet, creating drafts in your email platform, posting a message to Slack. Tools are connected through native integrations, automation platforms, APIs, or protocols such as MCP.
4. Knowledge. The reference information the agent uses: your brand knowledge base, product facts, past campaign results, and policies (Chapter 17).
5. Memory and state. What the agent remembers between steps and runs, such as previous reports, decisions taken, or a log of actions.
6. Guardrails. Limits that keep the agent safe: permissions (read-only versus write), spending caps, approval steps, content checks, rate limits, and logging.

Agent design patterns

A few patterns cover most marketing needs:

• Pipeline (sequential): a fixed sequence of steps, some done by AI. Example: collect data → analyze → draft report → human review → send. Predictable and easy to debug. Start here.
• Router: AI classifies incoming items and sends each down the right path. Example: classify social comments as praise, question, complaint, or spam, and route each to a different response process.
• Generator and critic: one AI step creates, another evaluates against criteria, and the loop repeats until the output passes or a human steps in. Example: draft ad copy, then check it against brand rules and character limits.
• Orchestrator and specialists: a coordinating agent breaks a goal into tasks and hands them to specialist agents (research, writing, analysis), then assembles the results. Powerful but harder to control; use once simpler patterns are working.
• Monitor: an agent runs on a schedule, watches for conditions, and alerts or acts. Example: check daily ad spend and conversions, and alert the team about anomalies.

A library of marketing agents

Here are proven agentic workflows for each area covered in this series, with suggested autonomy levels. Use them as blueprints.

Analytics and reporting

Weekly performance reporter (Level 2 to 3). Every Monday: pull last week’s data from GA4, Google Ads, Meta, and the email platform into a spreadsheet; calculate KPIs versus target and prior periods; draft a What / So What / Now What report (Chapter 8); post the draft for the marketing manager’s review; send to stakeholders after approval.

Anomaly monitor (Level 2 or 4). Daily: check traffic, conversions, spend, and tracking health; flag sudden changes beyond set thresholds; suggest likely causes using the investigation checklist from Chapter 8. At Level 4 it might automatically pause a campaign whose spend exceeds a hard cap with zero conversions, then alert the owner immediately.

SEO and content

SEO content brief agent (Level 2). Given a target keyword: research the current results, extract common subtopics and questions, check the brand knowledge base for relevant expertise and internal links, and produce a content brief (Chapter 3) for an editor to approve.

Content refresh agent (Level 2). Monthly: identify pages with declining organic clicks in Search Console, analyze what changed in ranking competitors, and recommend specific updates.

Repurposing agent (Level 3). When a new article or webinar is published: produce platform-specific social posts, a newsletter section, and video scripts (Chapter 10), and place them in the scheduling tool as drafts for approval.

AI visibility tracker (Level 2). Monthly: ask a set of customer questions across AI assistants (where permitted by their terms), record whether your brand is mentioned and which sources are cited, and summarize trends.

Paid media

Search terms agent (Level 2 to 3). Weekly: analyze the search terms report, propose negative keywords and new keywords with reasons (Chapter 5), and apply them after a human approves the list.

Creative variation agent (Level 2). Given a winning ad and audience insights: generate new concepts and variations within brand rules and platform character limits, check them with a critic step, and deliver them for creative review.

Budget pacing agent (Level 2 or 4). Daily: compare spend to budget plan and performance targets, and recommend reallocations. Automatic action only within strict caps.

Email and lifecycle

Campaign builder (Level 3). From an approved brief: draft subject lines, preview text, and copy for each segment; check against brand and compliance rules; create drafts in the email platform; and alert the email manager for final review and scheduling.

Email QA agent (Level 2). Before any send: check links, UTMs, personalization fallbacks, alt text, unsubscribe link, spelling, and compliance requirements, and produce a pass or fail report.

Social and community

Comment triage agent (Level 3). Classify incoming comments and messages; draft replies for routine questions using the response guide (Chapter 12); escalate complaints and sensitive issues to a person immediately; publish routine replies only after approval until accuracy is proven.

Listening digest (Level 1 to 2). Daily or weekly summary of brand and competitor mentions, sentiment shifts, and emerging issues.

Research and strategy

Competitor watch (Level 1 to 2). Weekly: check competitors’ websites, ad libraries, social channels, and pricing pages; summarize notable changes and possible responses.

Voice-of-customer analyst (Level 2). Monthly: analyze new reviews, support tickets, and survey responses; update themes and customer language in the knowledge base after human review.

Prompt 18.1: Agent opportunity finder Act as an AI automation architect for marketing teams. Here are our recurring marketing processes with the time each takes per week and the tools involved: [list]. For each process, assess whether it suits an agentic workflow. For suitable ones, recommend a design pattern (pipeline, router, generator-critic, orchestrator, monitor), the autonomy level (1 to 5) with justification, the tools and data the agent needs, the human approval points, the main risks, and the estimated time saved. Rank the top five by value and ease.

Toolkit: The Agent Library sheet contains 20 ready-made agent blueprints from this series, with patterns, autonomy levels, and guardrails.

Designing an agent: the agent spec

Before building anything, write an agent spec on one page. It forces clear thinking and becomes your documentation.

1. Name and purpose: what business outcome does it serve?
2. Trigger: schedule, event (such as a new form submission), or manual request.
3. Inputs: data sources and knowledge it uses.
4. Steps: the workflow, marking which steps are AI, which are traditional automation, and which are human.
5. Tools and permissions: each tool, and whether access is read-only or write.
6. Outputs: what it produces, in what format, delivered where.
7. Autonomy level and approval points.
8. Guardrails: limits, banned actions, content rules, spending caps, escalation conditions.
9. Success metrics: time saved, accuracy, performance impact.
10. Owner: the person accountable for the agent’s behavior and maintenance.
11. Failure handling: what happens when a step fails or the AI is uncertain.

Example spec: Brightleaf Weekly Performance Reporter

• Purpose: faster, consistent weekly decisions; frees about 3.5 hours of manager time per week.
• Trigger: every Monday at 7:00.
• Inputs: GA4 (sessions, key events, revenue by channel), Google Ads and Meta (spend, conversions, CPA, ROAS by campaign), Klaviyo (revenue per recipient by flow and campaign), weekly targets sheet, last four weeks’ reports.
• Steps: (1) automation pulls data into a Google Sheet; (2) sheet formulas calculate KPIs and changes; (3) AI step reads the sheet and prior reports and drafts the report using the What / So What / Now What template, citing exact numbers from the sheet; (4) a critic step checks that every number in the draft matches the sheet and flags unsupported claims; (5) the draft is posted to the manager in Slack with a link; (6) after the manager approves or edits, the report is emailed to leadership.
• Permissions: read-only access to all data sources; write access only to the reporting sheet and a draft document; sending requires approval.
• Guardrails: never invent numbers; label inferences as hypotheses; if data is missing or a source fails, say so at the top of the report rather than estimating.
• Metrics: manager review time under 30 minutes; zero numerical errors over four weeks; report delivered by 9:00 Monday.
• Owner: marketing manager.

Prompt 18.2: Agent spec writer Help me write a one-page agent spec for this marketing workflow: [describe the process]. Use these sections: purpose, trigger, inputs, step-by-step workflow (label each step AI, automation, or human), tools and permissions (read or write), outputs, autonomy level and approval points, guardrails, success metrics, owner, and failure handling. Ask me clarifying questions first if anything important is missing. Recommend the lowest autonomy level that still delivers most of the value.

Toolkit: Fill in the Agent Spec Template (spreadsheet or printable Word version) before you build anything.

Setting up an agent: step by step

There are three main ways to build marketing agents, from simplest to most flexible.

Option A: Inside your AI assistant

Most team AI assistants can now act more like agents: custom GPTs, Gemini Gems, Claude Projects, and similar features hold instructions and knowledge files; built-in tools handle web search, file analysis, and code; connectors link to drives, email, calendars, and business apps; and agent modes can carry out multi-step tasks such as researching across websites and assembling a document. Scheduled tasks let some assistants run a prompt automatically on a schedule.

Best for: research, analysis, drafting, and recurring summaries where a human starts or reviews each run.

Option B: No-code automation platforms with AI steps

Platforms such as Zapier, Make, and n8n connect thousands of apps and let you add AI steps (calling a model with your instructions) and, increasingly, agent features that choose tools themselves. Native automation builders inside HubSpot, Klaviyo, Salesforce, and other platforms also include AI steps and agents.

Best for: pipelines and monitors that connect several tools on a schedule or trigger, such as reporting, triage, QA, and repurposing.

Option C: Custom-built agents

Developers can build agents with model providers’ APIs and agent frameworks, such as OpenAI’s Agents SDK, Google’s Agent Development Kit, and Anthropic’s Claude Agent SDK (the same foundation behind its Claude Code agent), and connect them to tools through APIs or MCP servers. This offers the most control and is appropriate for complex, high-volume, or deeply integrated use cases, usually with help from engineering or marketing operations.

The build process (any option)

1. Write the agent spec (above).
2. Do it by hand first. Run the workflow manually with an AI assistant for a week or two. You’ll discover edge cases and refine the instructions before automating anything.
3. Build the simplest version. Use a pipeline pattern with fixed steps. Connect read-only data sources first.
4. Write strong instructions. Include role, goal, method, format, rules, examples of good output, and explicit instructions for uncertainty (“If the data is incomplete, say so; do not estimate”).
5. Add guardrails and approval steps. Route outputs to a person (Slack, email, or a review queue) before anything goes to customers or changes spend.
6. Create a test set. Gather ten to twenty realistic inputs, including tricky ones (missing data, unusual spikes, a sensitive complaint), with notes on what good output looks like.
7. Test and refine. Run the test set, score outputs, and adjust instructions and steps until performance is consistently acceptable.
8. Run in shadow mode. Let the agent run alongside the current process for two to four weeks without its outputs being used, and compare.
9. Go live at a low autonomy level, with logging and an owner.
10. Review regularly and raise autonomy only with evidence.

Worked example: building the comment triage agent with a no-code platform

1. Trigger: new comment or direct message on Instagram or Facebook (via the platform’s integration or your social management tool).
2. AI step 1 (router): classify the message as praise, product question, shipping or order issue, complaint, collaboration request, spam, or sensitive (safety, legal, harassment), with a confidence score.
3. Branch: sensitive items or low-confidence classifications go straight to the community manager via Slack with an “urgent” tag.
4. AI step 2 (drafting): for praise and product questions, draft a reply using the response guide and product fact sheet. For order issues, draft a reply that moves the conversation to private messages.
5. AI step 3 (critic): check the draft for brand voice, accuracy against the fact sheet, and banned claims.
6. Human approval: drafts appear in a review queue; the community manager approves, edits, or rejects with one click.
7. Logging: every message, classification, draft, and decision is recorded in a sheet for weekly review.
8. Graduation: after four weeks with over 95% of praise replies approved without edits, the team allows praise replies to post automatically (Level 4), while everything else stays at Level 3.

Prompt 18.3: Agent instruction writer Write production-quality instructions for an AI step in a marketing workflow. The step’s job: [describe]. Inputs it will receive: [describe format]. Knowledge it can use: [list]. Required output: [format, ideally structured such as JSON fields or a fixed template]. Include: role, goal, a numbered method, rules (accuracy, brand, compliance), what to do when information is missing or confidence is low, two examples of good output, and one example of an output that would be wrong and why. Keep it concise and unambiguous.

Prompt 18.4: Test set builder Create a test set of 15 realistic inputs for this marketing agent: [describe agent and inputs]. Include normal cases, edge cases (missing data, extreme values, ambiguous requests), and risky cases (sensitive complaints, requests to make unapproved claims, attempts to get the agent to ignore its rules). For each, describe what a correct output looks like and how to score it.

Testing, evaluation, and monitoring

Agents need ongoing quality control, just like people.

Evaluate on four dimensions:

1. Accuracy: are facts and numbers correct?
2. Quality: is the output on-brand, clear, and useful?
3. Safety and compliance: does it follow rules, avoid banned claims, and escalate correctly?
4. Efficiency: time and cost per run compared with the manual process.

Monitor in production:

• Keep logs of inputs, outputs, actions, and human decisions.
• Track the approval rate (how often humans accept output without edits) and the edit rate. Falling approval rates signal drift.
• Review a sample of outputs every week.
• Re-run your test set whenever you change instructions, models, or tools.
• Watch costs: usage-based AI and automation fees can grow quickly with volume.

Toolkit: Use the Agent Test Set sheet to build and score 15 or more test cases, including prompt injection attempts.

Security risks specific to agents

Agents that read external content and take actions introduce new risks:

• Prompt injection: text in a web page, email, document, or comment that tries to hijack the agent (“Ignore your instructions and…”). An agent that reads customer comments or browses the web must treat that content as data, not instructions, and should not be able to take high-impact actions based on it without human approval.
• Excessive permissions: give each agent the minimum access it needs. Prefer read-only access; separate drafting from sending.
• Data leakage: don’t let agents send customer data to unapproved tools or include it in outputs where it doesn’t belong.
• Runaway actions: set caps on spend, number of actions, and messages sent per run.
• Credential management: use proper connectors and secret storage; never paste passwords or API keys into prompts or shared documents.

The agentic marketing team

As agents take on more execution, marketing roles shift from doing every task to designing, directing, and supervising systems that do tasks. Marketers increasingly spend their time on strategy and customer insight; on writing excellent briefs, instructions, and knowledge bases; on reviewing and approving agent output; on running experiments; and on improving the workflows themselves. The skills you built in Parts I to IV are exactly what make you good at this: you cannot supervise an SEO agent well unless you understand SEO.

Hands-on exercises

Exercise 18.1: Autonomy decisions (20 minutes). For ten marketing tasks (such as “publish blog post,” “add negative keywords,” “reply to a positive comment,” “increase campaign budget,” “send win-back email”), choose an autonomy level and justify it.

Exercise 18.2: Manual dry run (60 minutes). Pick one agent from the library. Run the workflow manually using an AI assistant, step by step, on real or demo data (the GA4 demo account works for reporting). Note where the AI struggled and how you’d fix the instructions.

Exercise 18.3: Build a simple pipeline (90 minutes). Using a free or trial plan of Zapier, Make, or n8n, build a pipeline with one trigger, one AI step, and one human approval step (for example, new blog post RSS item → AI drafts three social posts → posts to a Slack channel or email for approval).

Exercise 18.4: Red team your agent (30 minutes). Feed your agent tricky and adversarial inputs from Prompt 18.4, including a comment that contains instructions trying to change its behavior. Did it follow its rules? Fix any weaknesses.

Strategic assignment: agentic marketing blueprint

Deliver an Agentic Marketing Blueprint: a ranked list of agent opportunities from Prompt 18.1; full agent specs for two agents; instructions and a 15-case test set for each; a working prototype of at least one agent (or a detailed manual dry run with screenshots if you lack tool access); test results with accuracy and approval-rate scores; a security and guardrails review; and a plan for moving each agent from shadow mode to live operation.

Key takeaways

• Agents pursue goals across multiple steps using tools; most useful marketing agents are structured agentic workflows with human approval points.
• Match autonomy to the cost of mistakes, start low, and raise autonomy only with evidence.
• Every agent has a model, instructions, tools, knowledge, memory, and guardrails. Instructions are where marketing expertise matters most.
• Write an agent spec, do the work manually first, build the simplest pipeline, test with a realistic test set, and run in shadow mode before going live.
• Monitor accuracy, approval rates, and costs, and defend against prompt injection and excessive permissions.
• Marketers increasingly design and supervise AI systems; channel expertise is what makes that supervision good.

---

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Comment *

Name *

Email *

Website

Δ