Where AI Helps, and Where It Costs You
AI earns its place in marketing on volume work: generating variants to test, summarising research, drafting from a clear brief, pulling patterns out of campaign data. It costs you on anything a customer reads as your voice, on original strategy, and on the judgement calls about what's worth saying at all. The line isn't about capability. It's about whether being wrong is cheap or expensive.
That distinction is more useful than the usual framing, which treats this as a question of whether AI is good enough yet. It mostly isn't the relevant question. A tool can be good enough and still be the wrong thing to point at a problem.
The Cheap-to-Be-Wrong Test
Here's the rule we use internally.
If being wrong costs a few minutes, use AI. If being wrong costs trust, don't.
Thirty ad headline variants where twenty-eight are discarded: cheap to be wrong. Use AI, test them properly, keep the two that work.
A client's About page, in their voice, that a prospect will read while deciding whether these people are serious: expensive to be wrong. Write it.
The test works because it stops the argument about quality and starts a conversation about risk. Plenty of AI output is perfectly good. The question is what happens in the cases where it isn't, and whether you'll notice before a customer does.
What We Actually Use It For
Specifics, because "we use AI responsibly" is the kind of sentence that means nothing.
Variant generation for paid media. Ad copy, headlines, hooks. We'll generate thirty, cut to six that a human would write, and let the platform decide. This is the clearest win in the whole category. The creative volume paid social demands has always outrun what a small team can produce by hand.
Research synthesis. Reading twenty competitor pages and pulling out what they all say and what none of them say. The gap analysis is where the value is, and it used to take a day.
First drafts from a real brief. Not "write me a blog about SEO". A brief with the angle, the evidence, the structure and the position we're taking. The draft still gets substantially rewritten, but starting from something is faster than starting from nothing.
Data pattern-finding. Pointing a model at six months of campaign data and asking what changed. It surfaces things worth checking. It does not get the final say on what they mean.
Transcription and repurposing. Turning a client interview into usable quotes and content angles.
What We Won't Use It For
Anything published in a client's voice without a human rewrite. Not because it reads badly. Because brand voice is accumulated specificity, and a model regresses toward the average of everything it has read. Average is the one thing a small business competing against larger ones cannot afford to sound like.
Strategy. A model can tell you what agencies typically do in your situation. That's a description of the consensus, which is frequently the thing you need to deviate from.
Fabricating evidence. No invented statistics, no case studies that didn't happen, no reviews written in a customer's voice. This should be obvious. It is not universally observed, and it is the fastest way to destroy the trust the rest of the work is trying to build.
Client communication of any substance. A generated project update reads exactly like what it is.
The Thing That Changed Our Mind
We used to think the risk with AI content was quality — that it would be noticeably worse and readers would bounce.
That's not what happens. The risk is sameness.
This matters more as AI reshapes how Australians search. When we audit a category now, the pages blur. Same structure, same subheadings, same reasonable hedged tone, same three-item lists. Everyone has access to the same tools, so everyone produces the same shape of thing. Nobody's output is bad. The whole category has become indistinguishable, and indistinguishable means a buyer has nothing to choose on except price.
Which flips the strategic conclusion. The scarce asset is no longer production capacity. It's anything specific enough that a model couldn't have generated it: a real number from a real job, an opinion with a cost attached, a case where your approach didn't work and what you learned.
AI makes the generic cheap. That makes the specific more valuable, not less.
What This Means for Google and AI Search
Google doesn't penalise content for being AI-generated. Its published guidance targets content produced primarily to manipulate rankings rather than help people, regardless of how it was made. Plenty of AI-assisted content passes that test and plenty of human-written content fails it.
The practical consequence runs the other direction from what most people expect. AI systems choosing what to cite favour sources with specific, verifiable, corroborated information. Generic content is exactly what they don't need, because they can already generate generic content. What they can't generate is your pricing, your local data, your documented result on a real job.
So the content most likely to get cited by an AI system is the content least likely to have been written by one. We go into how that citation decision works in our guide to generative engine optimisation.
A Workable Split
If you're running marketing for an Australian SME and want a position rather than a philosophy:
Use AI for the first 70% of anything high-volume and low-stakes. Variants, drafts, research, summaries, data exploration.
Spend the saved time on the 30% that can't be generated. Go and get the real number. Interview the client. Photograph the actual job. Write the paragraph that says the thing your competitors won't.
Most teams do the opposite. They use AI to produce more generic output and fill the saved time with still more generic output. The volume goes up and the distinctiveness goes down, which is the worst available trade in a market where everyone's volume is going up.
How to Tell If an Agency Is Over-Relying on It
Four signals, if you're evaluating someone.
Their own content has no specifics — no numbers, no named examples, no positions anyone could disagree with.
They can't tell you what they use AI for. Either defensive vagueness or enthusiastic vagueness; both mean the same thing.
Volume is the headline. "Twelve blog posts a month" is a production commitment, not a marketing strategy.
Nothing they publish takes a side. If you can't find a single sentence someone in the industry would argue with, no human was really involved in deciding what to say.
Our content marketing work and AI marketing services run on the split above, and the part we spend most time on is the 30%. It's also the part clients notice.
For the mechanics of how AI systems decide what to cite, Google's own documentation on AI features in Search is a better starting point than most agency blogs.
Frequently Asked Questions
Can AI replace human marketers?
Not for the parts that matter most. It replaces production capacity for high-volume, low-stakes work like ad variants, drafts and research summaries. It doesn't replace strategic judgement, brand voice, or the first-hand knowledge that makes content worth citing. The roles that shrink are the ones that were mostly production.
How do marketing agencies actually use AI day to day?
Common uses are generating ad copy variants to test, synthesising competitor research, producing first drafts from a detailed brief, finding patterns in campaign data, and transcribing interviews. The useful question to ask an agency is which of these they do and what they deliberately keep human.
Does Google penalise AI-generated content?
Google's guidance targets content made primarily to manipulate rankings rather than help readers, regardless of how it was produced. AI-assisted content that is accurate, specific and useful is not penalised for its origin. Generic content produced at volume performs poorly whether a human or a model wrote it.
What parts of marketing should never be automated?
Anything a customer will read as your voice, original strategy, decisions about what's worth saying, and any claim of fact or result. The rule that works: if being wrong costs minutes, automate it. If being wrong costs trust, don't.
Which AI marketing tools are genuinely worth using?
The category changes too quickly for a list to stay accurate, and most teams get more value from using one general-purpose model well than from assembling a stack of specialist tools. Judge any tool by whether it removes production work without removing judgement.
Does AI-written content hurt brand trust?
Unedited, published in your voice, yes — not because it reads badly but because it reads like everyone else. Brand voice is built from specific detail, and models regress toward the average of what they've read. The damage is sameness rather than obvious error.
How do I keep brand voice consistent when using AI?
Write a voice guide with real examples of your own published work, including phrasings you use and ones you avoid. Use AI for structure and first drafts, then rewrite the sentences a reader will judge you on. Keep the opening, the positions and the closing human.
Can AI do creative strategy?
It can summarise what's been done in your category, which is useful input and a poor output. Strategy means deciding what to do differently from the consensus, and a model trained on the consensus will tend to reproduce it. Use it to map the field, not to choose your position in it.
How much time does AI actually save a marketing team?
Meaningfully on drafting, research synthesis and variant production. Much less on anything requiring first-hand information, client input or a judgement call. The honest answer for most small teams is that it compresses the first 70% of a task and leaves the last 30% untouched, which is still substantial.
How do I tell if an agency is over-relying on AI?
Their own published content has no specific numbers, named examples or arguable positions. They can't say clearly what they use AI for. They pitch volume as the deliverable. And nothing they publish takes a side anyone in the industry could disagree with.