AI marketing workflows: nine recipes with trigger, steps and checks
The most useful AI marketing workflows are small: one trigger, a few steps, one model call and a person who checks the result. Below are nine recipes a small team can build with an email or CRM platform, a connector such as n8n, Make or Zapier, and a language model. Each lists what to check by hand.
How to read the recipes
Every recipe has the same four parts. The trigger is the event that starts it. The steps run in order. The tools column names tool types, because the recipe works with any product that does that job. The check by hand column is the part you keep doing yourself. If a recipe has no check, it is not finished.
Lead and sales recipes
| Recipe | Trigger | Steps | Tools | Check by hand |
|---|---|---|---|---|
| 1. Enquiry triage and reply draft | New contact form submission | Classify enquiry type and language. Look up the contact in the CRM. Draft a reply from approved text blocks. Send the draft to an approver. | Form, connector, language model, CRM | Every draft before it is sent, with extra care for prices and dates |
| 2. Lead enrichment and fit score | New business contact in the CRM | Add company size and sector. Ask the model for a fit score from 1 to 5 with a one-line reason. Write both to the CRM. Alert sales for scores of 4 and 5. | CRM, enrichment service, connector, language model | Compare scores with won and lost deals each month. Confirm you have a lawful basis for the added data |
| 3. Meeting follow-up | Meeting notes are saved | Summarise the notes. Draft a follow-up email with the agreed next steps. Create the tasks in the CRM. | Notes or transcript tool, language model, CRM | The summary against what was really agreed. Consent of participants if the call was recorded |
Content and campaign recipes
| Recipe | Trigger | Steps | Tools | Check by hand |
|---|---|---|---|---|
| 4. Content reuse | A new blog post is published | Draft three social posts and a newsletter blurb in your house style. Save them to a review board. | CMS feed or webhook, connector, language model, scheduling tool | Facts, tone and links. Publish by hand, never automatically |
| 5. Weekly performance digest | Schedule, Monday morning | Pull figures from analytics, ad accounts and the email platform. Ask the model for a five-line summary of the biggest changes. Post it to the team channel. | Analytics, ad platforms, connector, language model, chat | Two figures against the source dashboards every week |
| 6. Subject line options | A campaign draft gets a certain tag | Generate five subject lines and two preview texts. Store them as options next to the draft. | Email platform with built-in AI, or connector plus language model | Choose two by hand for an A/B test. Reject any claim the email does not support |
Customer and retention recipes
| Recipe | Trigger | Steps | Tools | Check by hand |
|---|---|---|---|---|
| 7. Review and feedback routing | New review or survey response | Classify sentiment and topic. Draft a reply. Send negative cases to a named person. Log the topic in a sheet. | Review or survey tool, connector, language model, spreadsheet | Every public reply. Every negative case, personally |
| 8. Shared inbox tagging | New email in the marketing inbox | Classify as lead, partner, supplier, press or other. Apply the label. Create a CRM contact for leads only. Forward press requests to their owner. | Email, connector, language model, CRM | The other label once a week, to catch misfiled leads |
| 9. Re-engagement shortlist | Schedule, once a month | Find contacts with valid consent and no opens or orders for a set period. Draft a short win-back message per segment. Queue the campaign for approval. | Email platform or CRM, language model | Segment size, consent status and the offer before anything is sent |
Rules that apply to every recipe
- Give the model your approved wording, product facts and tone examples in the prompt. It should rearrange your words, not invent new claims.
- Ask for structured output, such as one category from a fixed list, so the next step can rely on it.
- Send only the fields the step needs. A classifier rarely needs a phone number or a full address.
- Keep a log of input, output and the approver for each run.
- Tell people when they are talking to a machine. The EU AI Act contains transparency obligations for AI systems that interact with people directly.
- Build one recipe, run it for a few weeks, then decide on the next.
Where these workflows go wrong
| Problem | How you notice | Fix |
|---|---|---|
| The model invents a detail | A draft quotes a price or date you never gave it | Supply the facts in the prompt and instruct the model to leave gaps it cannot fill |
| Categories drift | The other bucket keeps growing | Review misfiled items monthly and adjust the category list |
| Approvals pile up | Drafts wait for days | Name one approver and a stand-in, and send reminders after a few hours |
| Duplicate contacts | The same person gets two emails | Search by email address before every create step |
| Silent failures | Nobody notices for a week | Add an error branch that posts to a shared channel |
Common questions
Which AI marketing workflow should I build first?
Build enquiry triage with a reply draft first. It runs on every new lead, the saving is visible within days and a person still approves each message, so the risk is low. If you receive few enquiries, the weekly performance digest is the safer starting point because nothing it produces reaches a customer.
Do I need to know how to code?
No, not for these recipes. Connectors use visual builders, and calling a language model is one configured step. What you do need is a clear written description of the workflow, clean contact data and patience for testing every branch. Code becomes useful when a tool has no ready-made integration.