The question is rarely whether to use AI in marketing any more. It is where, on what, and how to avoid the version of it that produces a great deal of content nobody reads.
The useful frame is not tools. It is jobs. Some marketing jobs are pattern work with a clear right answer, and those are where a model earns its keep immediately. Others depend on judgement, taste or facts only you hold, and those are where it quietly makes things worse.
What AI is genuinely good at
These are the jobs where the gain is immediate and the risk is low.
- Turning one finished piece into other formats: post to newsletter, newsletter to social, webinar to article
- First drafts of structured, repeatable copy: product descriptions, meta descriptions, alt text, subject line variants
- Summarising and clustering: reading a hundred survey responses or reviews and telling you the five themes
- Editing: tightening your writing, catching the paragraph that says nothing, flagging jargon
- Getting past the blank page, which is a real and underrated use
What it is bad at, and will stay bad at
The failures cluster in one place: anything that requires knowing something the model has no access to.
It does not know your customers. It has never sat on a sales call, read your support inbox, or watched somebody fail to find the pricing page. Everything it says about your audience is an average of everyone's audience.
It does not know what is true about your product today, and it will produce a confident, plausible sentence about a feature you do not have. In marketing copy this is not a quirk, it is a claim you are now publicly making.
It has no taste and no stake. It cannot tell you which of two campaigns is the braver one, or which sentence is embarrassing, or when the honest answer is that the idea is weak.
The rule that prevents most of the damage
Use AI on things you can verify, and be the source of everything you cannot.
You can verify structure, grammar, format conversions and whether a summary matches its input. You cannot verify a statistic it produced, a quote it attributed, or a claim about your own product. Those must come from you, and if you find yourself accepting them because they sound right, that is the moment the work started getting worse.
Context is the whole difference
The gap between AI output that sounds like everyone and output that sounds like you is almost entirely context.
Write down, once, the things a new hire would need: who you sell to and what they actually worry about, the words your customers use, the words you refuse to use, what you do and do not do, three examples of writing you would be happy to publish. Feed that in every time.
This is dull and it is the highest leverage half hour available. Every prompt after it starts from your business rather than from the internet's average business.
A first week that produces something real
Rather than evaluating tools, run one week of actual work and judge from that.
- Monday: write the context document above, honestly, in your own words
- Tuesday: take your single best existing piece and turn it into four other formats
- Wednesday: draft twenty subject lines or headlines, keep the two you would send
- Thursday: paste in your last fifty support messages or reviews and ask what the recurring themes are
- Friday: have it edit something you wrote, then decide which of its cuts you agree with
The trap that catches everyone
Volume. The moment publishing becomes nearly free, the obvious move is to publish far more, and it is almost always the wrong one.
Search engines are explicit that content produced at scale primarily to rank is a problem. Readers are less explicit and reach the same conclusion faster. Ten pieces a week that nobody finishes is not ten times one piece that people share, it is a slower version of nothing.
The better use of the same capacity is depth: use the time you saved drafting to add the example, run the experiment, get the quote, or go back and make last quarter's best post genuinely the best answer to its question.
Where it goes from helpful to actually useful
Drafting is the shallow end. The step change is when the same system also knows what happened after you published.
A tool that writes is useful. A tool that writes, publishes, and then reads the opens, clicks and rankings, and lets the next thing be informed by that, changes the loop rather than one step in it. Judge anything you are considering on whether it closes that loop or just fills a text box faster.
How to tell if it is working
Not by output. Counting posts published is the metric that rewards exactly the behaviour that fails.
Measure the things that would have been true anyway if the work were good: time from idea to published, whether people finish what you publish, whether it earns links or replies, and whether the pipeline moved. If those are flat and your output tripled, the AI made you faster at doing something that was not working.
Common questions
- What should I never use AI for in marketing?
- Anything you cannot verify. Statistics it produced, quotes it attributed, and claims about your own product are the three that turn into published falsehoods. Also never let it answer an unhappy customer unread.
- Why does AI content sound generic?
- Because without context it returns the average of everyone's answer. Writing down your audience, your vocabulary, the words you refuse to use and three examples of writing you would publish is the half hour that fixes most of it.
- Should I publish more content now that drafting is cheap?
- Usually no. Search engines treat content produced at scale mainly to rank as a problem, and readers reach the same conclusion faster. The better use of the saved time is depth on fewer pieces.
Keep reading