Writing

• Marketing Workflows
• 23 September 2026

How to build an AI marketing workflow: steps, tools and examples

To build an AI marketing workflow, pick one process tied to a clear outcome, map every step before touching a tool, then decide which steps AI does, which a human approves, and what happens when the model is wrong. Only then choose a tool (Zapier, Make, n8n or your CRM’s own AI), build it, test it on messy cases and check whether the business result you named has moved.

The reason to do it in that order is that most firms are doing it the other way round. Masha Shunko of the University of Washington’s Foster School and Serguei Netessine of Wharton, writing in Harvard Business Review in September 2026, point to PwC’s 2026 Global CEO Survey, in which only 12% of CEOs report both revenue and cost benefits from AI, and McKinsey QuantumBlack’s finding that 60% of organisations still see no enterprise-wide EBIT impact (both cited in HBR). Their diagnosis is that companies bolt AI onto old tasks instead of redesigning the workflow around it.

How do you build an AI marketing workflow, step by step?

1. Pick the process, and name the outcome it exists to produce.
2. Map the current workflow, including the ugly bits, and draw the new one.
3. Decide where AI sits and where a human has authority, step by step.
4. Choose the tool that fits your stack and your checkpoints.
5. Build it with the CRM as the system of record.
6. Test it on bad inputs before good ones.
7. Measure the outcome, with guardrail metrics that stop it gaming itself.

Which marketing process should you automate with AI first?

Shunko and Netessine give three tests, refined over three years with dozens of companies. The outcome test asks what the customer actually values (for a lead-gen process, that’s a qualified conversation with sales, which a bigger contact database doesn’t automatically give you). The constraint test asks, for each step, what limitation it was designed around and whether that limitation still applies. The compounding test asks whether the step builds something that makes the next cycle better, such as a correction that improves the scoring rule.

Their description of a good first pilot is work that sits close to the desired outcome, uses up expert time without needing expert judgement, runs on digital inputs, can be checked quickly and creates learning for the next round. In marketing that usually points to content repurposing, lead enrichment and triage, campaign reporting or first-draft email follow-up. It rarely points to brand positioning, which is where the enthusiasm usually wants to start.

In Lean terms, as they put it, automating muda (waste) merely produces cheaper muda. If your monthly report is a 40-slide deck nobody reads, AI will produce it faster, and nobody will read it faster.

How do you map a marketing workflow before adding AI?

Write out every step as it happens today, who does it and what triggers it. Then mark where things go wrong: the missing field, the duplicate record, the lead from a competitor, the blog post that is actually a press release. This is the lead-generation workflow from the second example.

[Web form submitted]
        |
        v
[Create/update contact in CRM] --duplicate?--> [Merge, keep original owner]
        |
        v
[Enrich: company, size, industry] --no match / free email domain?--> [Flag: manual review]
        |
        v
[Score: fit + engagement]
        |
   +----+-----------------+----------------------+
   |                      |                      |
 High (A1-A2)          Medium                  Low
   |                      |                      |
[Route to sales      [Nurture sequence,     [Newsletter only,
 owner + Slack alert] AI-drafted, human      no sales contact]
   |                   approves first send]
   v
[AI drafts follow-up email] --> [Owner approves or edits] --> [Send + log in CRM]

Then check it against the four ways Shunko and Netessine say AI pilots fail when the surrounding workflow isn’t redesigned.

Prompts to generate a flowchart of your workflow

Describe the process in plain English and let an AI tool draw it, then correct it, because the correcting is where you find the steps nobody wrote down. Two prompts I’d use:

Prompt 1 (Mermaid code, for ChatGPT or Claude)
Write a Mermaid flowchart (flowchart TD) of this process. Use rectangles for
actions, diamonds for decisions, and label every arrow. Mark each step as
[AI], [Human] or [System]. Process: when a blog post is published in our CMS,
AI drafts three LinkedIn posts and one newsletter blurb, a marketing manager
approves or edits them in Slack, approved posts are scheduled, rejected ones
are logged with the reason.

Prompt 2 (unhappy paths, for any diagramming tool)
Take this workflow and add the failure branches: missing or invalid data,
duplicate records, low-confidence AI output, an existing customer filling in
the form, and no response from the approver within 24 hours. For each branch,
show who handles it and whether the action can be reversed.

Mermaid is a text-based diagram syntax (decisions go in curly braces, and a labelled arrow looks like -->|yes|), and ChatGPT and Claude both write it well enough that you can paste the result into the free Mermaid Live Editor at mermaid.live; Claude also has a Mermaid Chart connector that validates the syntax and renders the diagram in the chat. If you’d rather drag boxes around afterwards, Miro AI turns a prompt into an editable diagram on every plan including Free, and Lucid AI in Lucidchart has a Process Agent that asks clarifying questions as it builds (Lucid says its AI features are in beta and may become paid).

Where should AI sit in the workflow, and where should humans check?

“Human in the loop” appears in every AI deck and almost never comes with a definition. Shunko and Netessine’s 4A framework gives you one per step:

• Assist: AI drafts or analyses, a human decides and acts. Use it when stakes are high or cases are rare.
• Approve: AI prepares the action, a human signs it off before it executes. Use it when speed matters but some risk remains.
• Audit: AI decides and executes, humans review samples and fix systemic errors. Use it for high-volume work where mistakes are visible and reversible.
• Automate: AI executes within human-set guardrails, with people stepping in only for exceptions. Use it for frequent, low-risk, rule-bound tasks.

They reduce the choice to two questions: who gets hurt if it goes wrong, and can you undo it. Anything irreversible that an agent initiates should default to Approve, and in marketing that covers a sent email, a published post and a deleted CRM record. Adam Seligman, CTO of the integration platform Workato, put the same idea neatly to Thomas McKinlay (founder of the research advisory firm Science Says), Stefano Puntoni (Wharton) and Serkan Saka (San José State University) for their September 2026 HBR piece on trust in agents: companies need to be able to say “You can draft this email, but you can’t send it.”

What I’d watch for is the approval queue that nobody opens, which is Automate mode with extra admin. Give each checkpoint a named owner and a timeout rule, otherwise the Slack approvals pile up while everyone assumes someone else is reading them.

Which tools should you use for AI workflows in marketing?

Native AI inside your CRM keeps everything in one place and inherits your permissions and data. An orchestration tool sits between systems and suits workflows that span a CMS, Slack, a CRM and an email platform.

Can I integrate AI workflows with my existing marketing CRM?

Yes, and the CRM should stay the system of record. Every AI output that matters (an enrichment value, a score, a draft email, an approval decision) should be written back to a named CRM property or activity, so sales sees it and you can audit it later. The alternative is a shadow database, AI results living in a spreadsheet or a Zapier table that nobody in sales ever opens.

Before you connect anything, agree which system owns each field (so the AI doesn’t overwrite a salesperson’s edits), decide the deduplication rule, carry consent and subscription status through every step, and give the integration its own user with only the permissions it needs. That last one sounds like IT fussiness until you read Shunko and Netessine’s account of a Cursor coding agent that deleted a company’s production database and backups in seconds after finding an overly broad API token.

How can I integrate AI into my existing marketing automation system?

Keep your existing nurture streams, lists and triggers, and insert AI at specific steps: an AI step that drafts the email body inside an existing sequence, an enrichment step before your scoring runs, a classification step that tags form free-text by intent. Each insertion can be tested and switched off on its own. For where agents differ from generative steps, see generative vs agentic AI in marketing.

Worked example 1: how do you repurpose blog content to social with AI and human review?

The outcome here is engagement and traffic from people who match your audience, and the review step matters more than it looks: Adam Peruta of Syracuse University found, in a study co-led with the market research firm Ipsos in which 3,000 US consumers watched 20 video ads (half rebuilt entirely with AI from the same briefs), that the AI versions were 14% weaker on short-term sales potential and 17% weaker on long-term brand equity, even though viewers struggled to tell them apart (HBR, September 2026). So the reviewer’s job is to ask whether a post would make anyone do anything, because it will almost always look fine.

1. Trigger: a new post is published in the CMS (via RSS or a CMS trigger in Zapier or Make).
2. AI step (Assist): the article and your brand voice guide go to an AI step that returns three LinkedIn posts, a newsletter paragraph and a list of claims to check.
3. Guardrail check: flag any figure not in the original article, and anything touching regulated claims (for a pharma brand, every product claim goes to medical and legal review, full stop).
4. Approval (Approve): Zapier’s Human in the Loop sends the drafts to the content manager in Slack to edit and approve. Set a 48-hour timeout that ends the run rather than posting anyway.
5. Publish: approved posts go to the scheduler; declined ones are logged with a reason.
6. Learn: once a month, read the edit log. If the manager rewrites every opening line, fix the prompt, not the manager.

The AI step prompt I’d start from:

You are drafting social posts for [brand] from the article below. Audience: [role, sector, UK]. Use British spelling and the tone in the attached voice guide. Write three LinkedIn posts of 120-200 words, each built on a different specific point from the article. Use only facts and figures that appear in the article; if a post needs a figure that isn’t there, write [NEEDS SOURCE]. At most two hashtags per post. Return JSON with fields post_1, post_2, post_3, newsletter_blurb, claims_to_check.

Worked example 2: how do you set up AI workflows for lead generation?

The outcome is a sales conversation with a company that fits your ideal customer profile, reached quickly. The build below is in HubSpot, with the Salesforce equivalent at step 5.

1. Form: HubSpot’s data enrichment can shorten forms by hiding fields it can already fill, so you might ask for only name and work email.
2. Enrichment (Automate): turn on automatic enrichment (it is off by default, and it doesn’t consume HubSpot Credits) so new records get firmographic data from HubSpot’s dataset and third-party providers. Unhappy path: no match or a personal email domain routes to a “needs review” list instead of scoring as a zero.
3. Scoring (Audit): HubSpot’s lead scoring tool (Professional or Enterprise) builds fit, engagement or combined scores, the combined version on a grid from A1 to C3 where the letter is fit and the number is engagement. The AI-built score (Marketing Hub Enterprise only) needs at least 50 contacts, 25 converted and 25 not, to train. Audit a sample weekly, because an AI score trained on last year’s customers will happily rank the intern’s cousin with a Gmail address as A1 if last year’s data was messy.
4. Routing (Automate): a workflow assigns an owner by region or segment when the score crosses your threshold (HubSpot’s own example uses a score over 50) and posts a Slack alert.
5. Follow-up email (Approve): an AI step drafts a first email from the enriched record and saves it for the owner to edit and send. HubSpot’s Nurture Agent sends each lead a personalised email by itself, so check its configuration before letting it loose. In Salesforce, Agentforce Engagement handles outreach and follow-ups, and its Require Manual Approval setting lets owners edit each email before it goes.
6. Close the loop: record outcomes (meeting booked, disqualified and why) on the contact so the score and prompt learn from reality.

On compliance, the ICO says the PECR email marketing rule doesn’t apply to corporate subscribers, though you must identify yourself and give an opt-out; sole traders and some partnerships count as individuals and need consent or the soft opt-in, and UK GDPR still covers a named person’s work email. The ICO’s B2B guidance is under review after the Data (Use and Access) Act, so check it before launch.

How do you test and measure an AI marketing workflow?

Test the unhappy paths first: blank fields, duplicate submissions, a competitor filling in the form, an approver on holiday. Shunko and Netessine’s five questions make a good test script: how the system recognises uncertainty, what triggers escalation, who resolves exceptions, how an action is stopped or reversed and how corrections feed the next cycle. Then run the new workflow alongside the old one for a few weeks.

For measurement, pick one outcome metric and at least one guardrail: meetings booked per 100 leads with sales acceptance rate as the guardrail, or LinkedIn engagement from target job titles with unsubscribe rate. More detail on setting baselines is in measuring the ROI of AI in marketing. Before you build anything, try completing Shunko and Netessine’s sentence for your project:

We will improve [outcome] in [end-to-end process] by redesigning [workflow], using AI to [role], with humans retaining authority over [judgment/control], exceptions routed through [escalation path], success measured by [primary KPI], and performance protected by [guardrails].

If you can’t fill it in, they say, you have a technology experiment rather than a transformation project. At ACTAsia, where I lead AI adoption in marketing, this is the test I’d put any new workflow through first (the wider approach is in the ACTAsia case study).

How do AI chatbots help lead generation for online businesses?

A chatbot on your site is the form that talks back: it answers the question that would have made someone leave, asks qualifying questions, writes the answers to the CRM and books a meeting or hands over to a person. Salesforce’s Trailhead module on Agentforce describes agents acting as the first point of contact for inbound leads and booking meetings for sellers, and HubSpot’s Customer Agent is built to answer common questions around the clock using contact history. The lead-gen value depends entirely on the handover, so treat the chatbot as the first step of the workflow in example 2.

What are the best practices for integrating AI chatbots into a marketing website?

• Ground it only in approved content, and keep pricing, refunds and policy answers on fixed, reviewed text. In Moffatt v Air Canada (February 2024), British Columbia’s Civil Resolution Tribunal held the airline liable for its chatbot inventing a bereavement fare refund policy, after Air Canada argued the chatbot was a separate legal entity, a defence we’d all like to try on our own bad campaigns.
• Test it adversarially after every update. In January 2024 the UK parcel firm DPD switched off the AI part of its chatbot after a customer got it to swear and call DPD the “worst delivery firm in the world”, which DPD blamed on a system update.
• Tell visitors what it can’t do. McKinlay, Puntoni and Saka’s experiments were with employees, not customers, but people told about an AI’s known weaknesses trusted it more and worked with it up to 7.2% more effectively, with the AI itself unchanged.
• Offer a human handover at any point, pass the conversation to the CRM so sales doesn’t ask the same questions again, and cover the bot in your privacy notice.

Where can you find tutorials on building AI-powered marketing workflows?

Start with vendor documentation, because product names move fast (Salesforce renaming Agentforce Lead Nurturing to Agentforce Engagement will have broken a few YouTube tutorials). The docs linked in the table above are all written as how-tos, and Salesforce Trailhead’s “Agents for Customer Engagement at Scale” module is free. For the thinking before the build, the Shunko and Netessine article is the best short read I’ve found, and for background there’s the pillar on what AI in marketing is.

If you want one next step, take the process your team complains about most, draw it with prompt 2 above, and count how many of its failure branches currently have an owner.

FAQs

What is an AI workflow in marketing?

An AI workflow in marketing is a sequence of connected steps, such as capturing a lead, enriching it, scoring it and emailing it, where some steps are done by an AI model and others by rules or people. The difference from ordinary marketing automation is that AI steps handle judgement-like tasks, such as drafting, classifying or prioritising, and so need explicit review points.

Do I need a developer to build an AI marketing workflow?

Not for most first projects. Zapier and Make let marketers add AI steps and approval pauses without code, and HubSpot and Salesforce include AI features inside the CRM. n8n offers more control but expects more technical skill. A developer becomes useful when you need custom integrations, strict data controls or high volumes where per-task pricing gets expensive.

Is AI lead scoring better than rules-based scoring?

It can be, if you have enough clean history. HubSpot’s AI score (Marketing Hub Enterprise) needs at least 50 contacts, 25 converted and 25 not, to train, and it learns from whatever past conversions you had, including bad ones. I’d run a rules-based fit score alongside it, audit a sample of AI-scored leads weekly and compare both against sales acceptance.

How much human review should an AI marketing workflow have?

Match review to risk and reversibility. Using Shunko and Netessine’s 4A framework, anything irreversible or public, such as sent emails and published posts, should need human approval. High-volume, reversible steps like enrichment can run automatically with sampled audits. Every checkpoint needs a named owner and a timeout rule, or the queue fills up unread.

Is B2B email follow-up from an AI workflow legal in the UK?

Generally yes for corporate subscribers, provided you identify yourself and include an opt-out, according to ICO guidance on PECR. Sole traders and some partnerships are treated as individuals and need consent or the soft opt-in, and UK GDPR still applies to named business contacts. The ICO says its B2B guidance is under review after the Data (Use and Access) Act.

Sources

1. Harvard Business Review, Masha Shunko and Serguei Netessine, “Stop Automating Old Processes. Design New Ones Instead.”, 14 September 2026. hbr.org
2. Harvard Business Review, Adam Peruta, “Research: AI-Generated Ads Perform Worse Than Human-Made Ones, Even When Customers Can’t Tell Them Apart”, 3 September 2026. hbr.org
3. Harvard Business Review, Thomas McKinlay, Stefano Puntoni and Serkan Saka, “To Adopt AI at Scale, Employees Need to Trust Agents”, 9 September 2026. hbr.org
4. HubSpot Knowledge Base, “Understand the lead scoring tool”, accessed 23 September 2026. knowledge.hubspot.com
5. HubSpot Knowledge Base, “Get started with data enrichment”, accessed 23 September 2026. knowledge.hubspot.com
6. HubSpot, AI and agents product page, accessed 23 September 2026. hubspot.com
7. Salesforce Help, “Agentforce Engagement Overview”, accessed 23 September 2026. help.salesforce.com
8. Salesforce Trailhead, “Agents for Customer Engagement at Scale”, accessed 23 September 2026. trailhead.salesforce.com
9. Zapier Help Center, “Request approval to keep your workflow running with Human in the Loop”, accessed 23 September 2026. help.zapier.com
10. Zapier, “Zapier’s AI tools: Get to know our governed AI products and features”, updated August 2026. zapier.com
11. Make, “Make AI Agents”, accessed 23 September 2026. make.com
12. n8n Docs, “Human-in-the-loop for tools”, accessed 23 September 2026. docs.n8n.io
13. n8n Docs, “AI Agent node”, accessed 24 September 2026. docs.n8n.io
14. Zapier, “Plans and pricing”, accessed 24 September 2026. zapier.com
15. Claude, “Mermaid Chart connector”, accessed 24 September 2026. claude.com
16. Mermaid, “Flowcharts Syntax”, accessed 23 September 2026. mermaid.js.org
17. Miro Help Center, “Miro AI with Diagrams and mindmaps”, accessed 23 September 2026. help.miro.com
18. Lucid Help Center, “Generate a diagram with AI in Lucidchart”, accessed 23 September 2026. help.lucid.co
19. Information Commissioner’s Office, “Business-to-business marketing”, accessed 23 September 2026. ico.org.uk
20. American Bar Association, Business Law Today, “BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot”, February 2024. americanbar.org
21. TIME, “Delivery Firm’s AI Chatbot Goes Rogue, Curses at Customer and Criticizes Company”, 20 January 2024. time.com