Most AI marketing claims are demos. A demo works once, on chosen data, with someone watching. The list below is the short one: things that survive contact with a real business and keep working when nobody is supervising them.
What actually holds up
Lead qualification in messaging. Answering the first few questions on WhatsApp or a website chat, capturing what someone needs, and routing the ones worth a call. It works because the task is narrow, the answers come from your own facts, and a wrong reply is recoverable.
Reporting. Pulling numbers from several platforms, writing the summary, flagging what moved. Genuinely tedious, entirely mechanical, and no judgement is lost by automating it.
First-draft content. Outlines and rough drafts, then rewritten by a person who knows the business. The saving is real. The output is raw material, not publishable work.
Bid and budget optimisation. Advertising platforms have run machine learning on this for years, and it beats manual adjustment for most accounts. The human job moved to what you feed it and what you measure.
Churn and intent prediction. Where you have enough history, spotting who is about to leave or who is close to buying is a genuine edge, because the action it triggers is cheap.
What does not
Publishing at volume. Cheap drafting tempts businesses into mass publication. Google's guidance is explicit that generating pages at scale to manipulate rankings falls under its spam policies, and thin pages dilute the site around them regardless.
Fully automated customer replies on anything complex. A confident wrong answer to a complaint costs more than the salary it saved.
Strategy. A model does not know your margins, which service actually makes money, or why last quarter went the way it did. It will produce a plausible plan anyway, which is the danger.
Anything unreviewed that a customer sees. Not for grammar, for truth.
How to tell a demo from a tool
Three questions settle most of it. What happens when it is wrong, and how would you notice? Does it work on your messiest data or only on a clean sample? Who checks the output, and is that check cheaper than doing the task manually?
If the honest answers are "nobody would notice", "only the sample" and "nobody checks", it is a demo.
Where the advantage actually goes
Not to whoever produces most. To whoever uses the time saved on the parts a machine cannot do: knowing the customer, deciding what deserves to exist, and being willing to publish less.
That is how we work, and every output a client sees passes a person first. If you want to see how your own site reads to search engines and AI assistants, our AI visibility check is free.
What does AI genuinely do well in marketing?
Lead qualification in messaging, cross-platform reporting, first drafts, bid optimisation, and churn or intent prediction where you have enough history. All are narrow tasks with recoverable errors.
Can I let AI write and publish content automatically?
No. Publishing pages at volume to influence rankings falls under Google's spam policies, and thin pages weaken the site around them. Use AI for drafts and have a person rewrite and verify.
Are AI chatbots safe for customer service?
For routine, factual questions, yes. For complaints or anything complex, a confident wrong answer costs more than the time it saved, so route those to a person.
How do I tell a useful AI tool from a demo?
Ask what happens when it is wrong and how you would notice, whether it works on your messiest data, and who checks the output. Weak answers to all three mean it is a demo.
Does AI replace marketing staff?
It replaces tasks, not judgement. Research, drafting and reporting get faster; deciding what to do, and whether the output is true, still needs a person.