What to hand to an AI marketing agent, and what to keep

The argument about AI in marketing is usually conducted at the wrong altitude — "can it replace a marketer" — where it cannot be resolved and nothing follows from either answer. The useful version is task by task. Some marketing tasks are production: known inputs, known output shape, repeated often. Those delegate well. Others are judgement: choosing, weighing, committing. Those do not, and the cost of pretending otherwise is a lot of confident, generic, unusable output.

The line: production versus judgement

A task delegates well when three things are true: the input exists in a form the agent can read, the output has a shape you could describe to a new hire, and a mistake is cheap to catch and fix. Where all three hold, delegation is nearly free. Where any fails, you are buying review work rather than removing it.

Delegates well

• First drafts against a specific brief — articles, product copy, captions, ad variants, email sequences.
• Reformatting: one idea into five platform-shaped versions, one article into a summary and a set of posts.
• Bulk production from structured data you already have — descriptions from attributes, alt text, meta titles.
• Repetitive analysis: crawling for broken links, missing metadata, thin pages, index changes, rank movement.
• Scheduling and publishing mechanics once something is approved.
• Translation drafts, as a starting point for someone who reads the language.
• Research summarisation: reading a lot of material and returning the shape of it.

Delegates badly

• Deciding what to say. Positioning, angle, which claim to lead with — these depend on things not written down anywhere.
• Choosing which queries to target, because the answer depends on what you can credibly claim and where you can realistically win.
• Anything with a legal, financial or medical surface. Draft, yes; publish unreviewed, no.
• Pricing, refund terms, guarantees, availability promises — anything a customer can hold you to.
• Relationship work: replies to real customers, negotiation, apology, anything where being wrong costs trust.
• Deciding when the strategy is not working. That requires wanting to know, which is not a capability you can delegate.

The brief is the whole job

The single largest determinant of output quality is not the model — it is what you gave it. A vague brief produces work that is grammatical, plausible, and interchangeable with everyone else's. A brief that carries real constraints produces work that could only be yours.

A brief worth writing contains:

1. The exact query or job, not the topic. "How do I know which size to order" beats "sizing".
2. Who it is for, in enough detail to exclude people.
3. The angle — what this piece argues that the existing results do not.
4. Your specifics: numbers, prices, timelines, real examples, the failure cases you know about.
5. What not to say. Claims you cannot support, comparisons you will not make, words you do not use.
6. The shape: page type, rough length, required sections, where it links.

Six lines, most of them reusable across a quarter. This is the work that does not go away, and teams that skip it conclude the tools do not work — which is true, given the input.

Feed it your data, not the internet

An agent working only from general knowledge can only produce a remix of what is already published, which is precisely the category that does not rank. An agent working from your site, your catalogue, your search data and your past performance can produce things nobody else could write.

• Your site and its existing pages, so new work does not contradict or duplicate old work.
• Your catalogue with real attributes, so descriptions come from facts rather than invention.
• Your search data — the queries you already appear for, and your position on them.
• What has worked before: your top pages, your best-performing posts, the emails that got replies.
• Your constraints: markets, languages, compliance limits, the things you do not sell.

Keep the irreversible steps human

The rule that prevents almost every serious failure: automation runs up to the point where an action becomes hard to undo, and a human crosses that line.

• Publishing to a live site — a wrong fact that gets crawled outlives the correction.
• Sending to a list — you cannot unsend, and complaints cost future deliverability.
• Spending money — budget changes and bid strategies compound before anyone notices.
• Anything a customer can rely on — a price, a delivery date, a guarantee.
• Deleting or redirecting URLs, which is fast to do and slow to unpick.

This is the same boundary that governs automating a WordPress content pipeline: humans decide and approve, machines produce and monitor.

How to hand a task over

1. Do it manually three times and write down what you actually did, including the judgement calls.
2. Turn that into a brief with constraints and examples of good and bad output.
3. Run it delegated with full review — every item checked, and every correction recorded.
4. Read the corrections. Patterns in them are missing constraints; add them to the brief rather than fixing each item forever.
5. Reduce review to spot-checks once corrections become rare. Keep the approval step for anything irreversible.
6. Re-check quarterly. Briefs drift out of date as the product and the market change.

The test for whether it worked

Compare total time before and after, including review. If reviewing the output takes longer than doing the task did, the delegation failed — either the brief is too thin or the task belongs in the judgement column. That is a real outcome, not a defeat: knowing which tasks do not delegate is most of the value of trying.

And measure the outcome, not the volume. Publishing four times as much is only progress if the additional pages do something. Volume without a query worth targeting is the most expensive way to look productive.

Frequently asked questions