Poldermarketing
How do I automate repetitive marketing work with AI?
Automate marketing work that repeats on the same trigger, with the same inputs, and a result you can check. Keep schedules and routing in the tools you already run. Use a model to read, classify, or draft. A named person still approves anything a customer sees, anything that spends, and anything you cannot take back.
Rank the task by what a mistake does
A repeated marketing job is one trigger, one shape of input, and one definition of done. Sort by the cost of a wrong output, whether you can take it back, how much judgment it needs, and how often it runs. Customer.io makes the contrast: a broadcast can be junior work you cannot undo, while a strategy sketch is senior work you can throw away.
| Bucket | Output | A person still |
|---|---|---|
| Automate | A read-only scheduled check, or another small reversible action | The exception |
| Augment | A draft, a class, a score, or a segment | Approval before it is public, sent, or spent |
| Keep human | Strategy, a full-list send, the brand line | The decision |
| Email platform | Inbox reputation, suppression, and unsubscribe | You stay responsible for the basis and the opt-out. The platform holds the record. You keep the run log |
Use a rule until the rule breaks
Keep the path as a rule in the tool that holds the record, and use a model to read, classify, summarize, or draft. OpenAI’s guide reserves agent for a model that manages the steps, chooses tools, and can stop and hand control back. A chatbot, a single-turn model, and a sentiment classifier are not agents. Build one only when rules fall short: many exceptions, a brittle rule set, or unstructured text. Otherwise a deterministic solution may suffice. Start with one agent and read access, and split only when conditions nest or tools overlap. See when marketing automation makes sense for a small company before a new system, and write a repeatable process before the model invents the steps.
Write the job as a brief, not a prompt
Write the goal, the inputs, the allowed claims, what done looks like, and the refusals: no new offer, no new number, no invented statistic, quote, or policy line, no email, and no spend change.
OpenAI’s guide says to start from a document you trust, break it into steps with an output each, and name the alternative step when a required fact is missing.
An undescribed field is a guess. Customer.io’s example is an attribute named cname with no description: a coin flip between a company name, a customer name, and a DNS record. The CRM and the email tool should share one described record.
Run one chain beside the old way
Run the automated chain next to the manual job, and keep it only if reviewing its output beats doing the job by hand.
- Map the job as it is actually done, not the tidy version.
- Add one good example, one bad example, and a retry that cannot double a record or a send.
- Run one trigger, one model step, one named reviewer, and one destination beside the old job.
- Compare edits and review time. If review still takes as long as the job, pick another job. Schedule after that, not after a demo.
- Name an owner and a recheck. Do not automate a process you would not hand to a new hire, and do not start by buying a new system.
Google Analytics lets an administrator email up to 50 standard or custom reports, daily, weekly, monthly, or quarterly, as PDF or CSV, to up to 50 people who already have property access. Realtime cannot be scheduled. The date range, filters, and comparisons freeze at scheduling. Events can take 72 hours, so use the fourth of the month, not the first. A schedule lasts 1 to 12 months, 12 by default, then expires, and is creator-invalid if its creator loses access, so use a company admin. The model writes the note and flags an empty or late file.
Zapier can pause for a yes or no, or to collect an edit, notify by email or Slack, and log the decision. Every reviewer needs an account and access to the workflow, and the pause cannot sit inside or after a looping step. On a path that publishes, sends, or spends, set both timeout and rejection to stop.
What marketing workflows are worth automating with AI
Worth automating: one trigger, one input shape, and a wrong result that is a draft, a flag, or a reversible change. A nurture that sends, a model that defines a qualified lead, and anything that can change spend stay off the list. Customer.io calls the unattended bucket smaller than the hype: high volume, low judgment, a cheap mistake. A subtly wrong segment is worse than an obvious one. Twice a year can stay manual. A welcome to the whole base stays human.
| Workflow | Worth it as | Never |
|---|---|---|
| Weekly readout | A note on the scheduled file | A partial export as the week |
| Launch QA | Flags for links, UTMs, claims, exclusions | Publishing the fix |
| Repurpose | Lines from one approved piece | A new offer, number, or proof |
| Feedback themes | An internal list | A reply to the customer |
| New lead | A score on rules you wrote | A model-made definition of qualified |
| CRM hygiene | A duplicate or empty-field flag | A merge or an overwrite |
| Lifecycle draft | Copy for a branch you defined | The send, or a list from an export |
| Paid check | A flag when your limit is crossed | Write access on day one |
Hold send-time picks, a page that rewrites itself, a chatbot that messages or books, and a post that publishes itself until a person has approved that branch. Customer.io’s unattended case is a per-person step nobody reviews one by one, not a weekly send.
Put consent and caps outside the model
A workflow does not create permission to email someone, and it must not be the only component that can send, spend, and change the offer.
UK ICO guidance bars marketing email or text to an individual without specific consent. Soft opt-in covers only your own previous customer and a similar product they bought or negotiated to buy, with an opt-out at collection and in every message. Prospects and bought-in lists are outside it. Texts and social direct messages count. Screen a do-not-contact list. A UK corporate subscriber can be emailed without consent. You still show who sent it and give a working opt-out. Sole traders and some partnerships count as individuals. The brief page is under review and has not absorbed PECR regulation 22(3A). Guidance updated on 28 April 2026 allows a charitable soft opt-in only for a charity, and only for details that charity collected on or after 5 February 2026 from someone who showed interest in, or gave support for, its charitable purposes. The message’s only purpose must be those purposes. A shop purchase or a wifi login is not that support. The opt-out is still required at collection and in every message.
Article 13 of Directive 2002/58/EC requires prior consent for direct-marketing email to subscribers and to users. A sale lets the same party reuse those details for its own similar products, with a free and easy objection at collection and in each message. It bans a hidden sender, no stop address, a breach of Article 6 of Directive 2000/31/EC, and a push to a site that breaches it. Paragraphs 1 and 3 cover natural persons only, so a company domain is not a Europe-wide yes.
In the Netherlands, Article 11.7 of the Telecommunications Act requires prior consent for email, text, and app messages to people and to companies. One exception is your own equivalent product to someone who bought or took a paid service, with an easy free objection when you collected the address and in every message. Consent is still the rule for a company. Two further exceptions sit beside that one: the company consciously published that address for this mail, or the company is outside the EEA and you follow that country’s rules. A general address published for something else does not qualify. A newsletter, a survey, a contest, or an account does not make someone a customer. Keep consent, suppression, and unsubscribe in the email platform, not in an export. Article 11.7 does not replace the GDPR. That use still needs its own legal basis, and the person is told what happens to the address. The model does not receive a copied list.
Caps and exclusions sit in the ad account. A later version may rotate an approved creative or change a bid only inside a cap the workflow cannot raise, with a person on the exception. Until that log is almost clean, it only flags. OpenAI’s guide scores each tool on read versus write, reversibility, permission, and financial impact, and says to pause before a high-risk tool runs.
Move the line only after the edits shrink
Move a task only after the log says the edits shrank, and only for information or an easy undo.
Mark each run accept, light edit, or rewrite. Record an invented fact, a duplicate, mail to someone who opted out, a spend change nobody intended, and whether anyone used the output. If they accept it and ignore it, stop. If the edits shrink but the saved hours go to more drafts, the job did not get smaller. Customer.io moves a task after dozens of reviews you actually read, with almost no rejections. Recheck a sample when the offer, the field names, or the model changes. A clean log from last month does not cover a new offer. If you still rewrite most outputs, fix the brief and the field names, or keep the task human. Renew the analytics schedule before it expires.
An experiment becomes a workflow after it has a result. Skip the build when the offer is not one sentence, nobody can review that week, or the tool already sends the notification.
I start with an internal job you can throw away
I am Piet Baudoin, one person, Poldermarketing, fully remote, in Dutch and English, for freshly funded startups. Hire me freelance for a bounded build or a few days a week. I am an AI-native growth marketer: I build and run marketing, AI, and automation, not only the advice, without a separate specialist for every part. I am strong in AI, content, automation, and workflows. Google Ads and Meta Ads are relatively new to me: I set them up and review them, but do not scale them.
Use the free growth scan, see how I work, and start at positioning and messaging when the words are the constraint.
Questions people ask
What marketing task should I automate first?
Pick a weekly job whose output you can check, and that cannot email anyone or change spend. A weekly internal readout is the usual first build. The analytics tool sends the file on a schedule, a model writes a short note in your headings, and you read it before anyone else does. Run that note once beside the old process. Put it on a timer only after you would have accepted the draft with light edits.
When is a chat window enough?
A chat window is enough when you do the task a few times a week, the inputs are already open, and you are sitting there as the reviewer. Use a workflow when the same paste runs between two systems on a timer, or when a missed step would send or spend: one trigger, one model step, one named person, one destination. An agent, a model that also chooses tools, waits until plain rules keep failing.
Can the workflow send email or change ads by itself?
Not in the first version, and not because a draft looks polished. Mail to a person needs a basis you already hold, a suppression list inside the email platform, and a named approver on anything you cannot recall. Ad changes sit behind caps and exclusions in the ad account. A first workflow may read results and flag what moved. It should not be able to raise spend, launch a send, and rewrite the offer in one run.
How do I know it is safe to review less?
Widen unattended work only where the output is information or easy to undo, and only after runs you actually read. Stop if a clean log is never opened. That is the same failure as accepting every draft unread. A send, a spend change, or a merge does not get that wider pass. If the log stays unread, end the job or name someone who will read it.