Written from the trenches of Auckland real estate by Amit Sharma — Bayleys agent, 10+ years marketing experience.
The reason AI content sounds like AI is almost never the model. It is that people ask it to invent the substance. When you type "write me a Facebook post about roof maintenance", you have given it nothing of your own, so it returns the statistical average of everything ever written about roof maintenance — which is exactly what average sounds like. The fix is a workflow change, not a prompt trick: you supply the raw material, the opinions and the specifics, and AI does the shaping. I am Amit Sharma, an Auckland marketer with a community of over 214,000 across my channels, and I use AI every single day for content. Here is the actual workflow, and the three places where it genuinely cannot help you.
Step one: capture raw material before you open any AI tool. This is the step everyone skips and it is the entire difference. Before you write anything, record yourself talking for two or three minutes about the thing. Use your phone's voice memo app. Talk like you would to a customer standing in front of you — what happened on the job, what the customer got wrong, what you would do differently, what annoys you about how your industry does it. That messy recording contains your actual opinions, your phrasing, and the specific details no model could ever guess. Transcribe it, and that transcript becomes the input. AI is a very good editor of your material and a very poor inventor of it. Treat it as the second half of the process, never the first.
Step two: give it your voice, not a description of your voice. Asking for a "friendly, professional, conversational tone" produces nothing useful, because those words describe most business writing ever published. Instead, paste in three to five things you have genuinely written or said — old posts, an email to a client, that transcript — and tell it to match the rhythm and vocabulary of those. Then be specific about the mechanics you want: short sentences, no rhetorical questions as openers, no lists of three, plain words. If you have a phrase you always use with customers, say so. Voice is imitable, but only from examples. It is never derivable from adjectives.
Step three: write the constraints down and reuse them. Keep a plain text file with your rules and paste it into every session. Mine bans a specific set of tells: the em-dash-heavy cadence when it is doing the work of a full stop, sentences built on "it is not just X, it is Y", openers like "in today's fast-paced world", closers like "whether you are a small business or a large enterprise", the word "delve", the word "leverage" as a verb, and any sentence that could sit unchanged in a competitor's post. Add a rule that says every paragraph must contain something only you could know. That last rule alone kills most of the generic output, because the model is forced to reach back into your raw material instead of its own averages.
Step four: make it produce more than you need, then cut hard. Ask for eight opening lines, not one. Ask for three versions of the same post at different lengths. The first thing any model produces is its most predictable output; the fifth and sixth are usually where something interesting appears. Then you do the part that cannot be delegated — choosing. Selection is taste, and taste is the thing you have that the tool does not. I typically throw away eighty percent of what I generate. That is not waste, it is the process working correctly. Generating is cheap; judging is the value you add.
Step five: read it out loud before it goes anywhere. This one costs thirty seconds and catches nearly everything. If you would not say a sentence to a customer standing in your shop, delete it. AI writing fails the read-aloud test in a very particular way — it sounds slightly formal, slightly balanced, slightly too symmetrical, like someone being careful. Real business writing is lopsided. It has a long paragraph then a short one. It repeats a word instead of hunting for a synonym. It ends bluntly. Reading aloud is the cheapest quality gate there is, and almost nobody does it.
Step six: add one thing the model could not have written. Before you publish, insert a specific detail — a street, a season, a customer's exact objection, a price bracket, a piece of local knowledge, something that happened on Tuesday. One concrete detail per piece is enough to make the whole thing land as human, because specificity is the signature of someone who was actually there. Generic content is not bad because it is generated. It is bad because it is generic, and it was generic long before AI existed.
Where AI genuinely fails, part one: anything that must be true. This is the important one, and it is non-negotiable. Language models produce text that reads correct, which is a completely different property from being correct. Every number, date, statistic, rule, price, legal requirement or regulation it gives you has to be verified against a real source before it goes near a customer. In a New Zealand context that means checking things like tax rules against official guidance, not against a confident paragraph. I have a simple standard for my own content: if I cannot verify a number, I remove it and describe the principle instead. A post with zero statistics and one honest principle beats a post with a plausible statistic that turns out to be invented. The reputational cost of one wrong figure in front of your customers is far larger than whatever the figure was buying you.
Where AI genuinely fails, part two: knowing what matters to your customers. A model can tell you what people generally ask about your industry. It cannot tell you that in your suburb, the objection you hear at the door is always about parking, or that your customers only ever call after a storm, or that the thing they are really worried about is being ripped off by the last guy. That knowledge lives in your inbox, your quotes, your phone calls and your comment section. The businesses that get the most out of AI are the ones that feed it that knowledge deliberately — keeping a running list of the actual questions customers ask, in the customers' own words, and using that list as the source of every content idea. If you want a structured way to build that, my AI business course covers exactly this, and the workflow above is the backbone of it.
Where AI genuinely fails, part three: judgement and taste. The model has no idea which of eight headlines is good. It will happily tell you all eight are strong. It does not know what your audience is tired of, what you promised last month, what would embarrass you, or when a piece is technically fine but should not be published. It has no sense of timing or of the room. Every workflow I run has a human decision at the end, and every time I have skipped that step to save ten minutes, I have regretted it. Use AI to widen the field of options. Never use it to make the final call.
A practical setup for a small business. Keep three saved documents: your voice file with the writing samples, your rules file with the banned patterns, and your customer-questions file with real phrasing from real people. Start every session by pasting all three. Then record a two-minute voice memo about whatever you did that week, transcribe it, and ask for five posts drawn strictly from it, in your voice, following your rules. Cut to two. Read them aloud. Add one specific detail to each. That entire loop takes maybe twenty minutes and it produces content that does not sound generated, because most of it is not — it is your material, tidied.
My honest take. AI has not made content easier, it has made bad content faster. Feeds are filling with fluent, competent, forgettable writing, and audiences are getting very good at spotting it. That is an opportunity, not a threat. The differentiator is now the same thing it was always supposed to be: having something specific to say and saying it like a person. AI is genuinely brilliant at removing the friction between having a thought and publishing it well — the blank page, the structure, the tidying, the eight variations. It is useless at having the thought. Keep that division clear and it becomes the best assistant a one-person business ever had. Blur it and you will publish a lot of words that nobody remembers. If you want the whole system taught properly, start with the masterclass or the practical checklists in my free guides.