You can spot AI-written copy in about three seconds, and so can your customers. It has a smell: tidy, confident, and completely interchangeable with the next company's. The usual reaction is to blame the tool and go back to writing everything by hand at 11pm. That is the wrong lesson. The model didn't fail you. You asked a stranger to write in your voice without ever telling it what your voice is. Fix that one thing and the same tool that produced beige filler starts producing drafts you'd actually sign your name to.
Why The Draft Sounds Like Everyone Else
A language model, prompted cold, does the safest possible thing: it writes the statistical average of every business blog post ever published. That average is polite, vague, and allergic to a strong opinion. It reaches for the words that show up everywhere because those words are, by definition, the most probable ones. That is why cold AI copy always lands in the same place regardless of whether you sell snow removal or bookkeeping.
Your actual voice is the opposite of an average. It is the specific way you explain a quote to a nervous homeowner, the joke you make about January in Edmonton, the fact that you never over-promise a timeline because you'd rather under-book than eat a bad review. None of that is in the model's default. It is in your sent folder, your voicemails, and your head. The whole job is getting it out of there and in front of the model before you ask for a single sentence.
Build A Voice File Before You Prompt
Spend 30 minutes once and save yourself every future rewrite. Open a plain document and paste in real examples of you talking to customers: three or four of your best emails, a couple of texts you've sent explaining a price, the About section of your site if it actually sounds like you. Aim for 600 to 1,000 words of genuine samples. Not marketing copy someone else wrote for you. The real stuff, typos trimmed but the rhythm left alone.
Then add a short set of rules the model can obey, because vague adjectives like 'professional' mean nothing to it. Write constraints it can act on.
- Words we use: plain-English trade terms, first names, 'quote' not 'investment', prices in real dollars.
- Words we never use: 'solutions', 'passionate', 'unlock', 'seamless', 'game-changer', anything you'd be embarrassed to say out loud on a job site.
- Sentence rules: short sentences, one idea each, no three-item lists stacked on top of each other.
- Tone by situation: warmer in a thank-you, blunter in a policy update, no exclamation marks in a price change.
- The line we never cross: no promises about timelines, results, or weather we can't control.
A Real Prompt, Start To Finish
Say you run a landscaping and snow crew and you need a spring email to last year's clients about booking aeration and seasonal contracts. Cold, you'd type 'write a marketing email about spring lawn services' and get something that could belong to any lawn company in North America. Instead, paste your voice file first, then give it the assignment: audience is existing residential clients, goal is to fill the May aeration schedule, mention that early-bird bookings before April 15 lock in last year's rate, keep it to 120 words, and match the samples above.
Now the draft comes back sounding like you wrote it on a Tuesday. You still read every line, cut the one sentence that rings false, and check the date and the price yourself. That editing pass takes five minutes instead of the 40 it would take to write from a blank screen. Run the same voice file against your Google Business posts, your quote follow-ups, and your off-season check-ins, and a week of marketing that used to never get done gets done before lunch.
This is the same principle behind the systems we build at AltaPro. For Zebra Landscaping we built a tool that pulls a job's real numbers into a finished quote, cutting quoting from about four hours to under 20 minutes. For an Alberta contractor we cut bid assembly from two or three hours down to minutes. In both cases the win came from feeding the system real, structured inputs first. The AI assembles fast; the judgment and the numbers stay yours.
The Test That Catches Fakes
Before anything goes out, run three quick checks. Read it out loud: if you'd never say it to a customer's face, delete it. Look for specifics: a real draft names the neighbourhood, the season, the actual service, the actual price. Filler names nothing. And ask whether your closest competitor could paste the exact same text under their own logo. If they could, it isn't yours yet.
Most bad AI copy dies on the first test. You read 'we're passionate about delivering exceptional experiences' aloud and physically wince. Good. That wince is your voice rejecting a transplant. Trust it, cut the line, and move on.
What To Never Hand The AI
AI is a strong drafter and a terrible authority. Keep a few categories away from it entirely, because the cost of a confident, wrong sentence is higher than the time it saved you.
Never let it invent facts about your business: numbers, guarantees, years in operation, certifications, or the specifics of your WCB coverage. It will produce something plausible and false. Never let it write testimonials or reviews; in Canada, fabricated reviews are a Competition Bureau problem, not a style problem. Never send an apology, an incident notice, or anything with legal weight without writing the core yourself. And never let it announce something genuinely new about your company from a cold prompt, because it has no idea what actually happened and will guess.
- Prices, guarantees, and timelines: you set these, every time.
- Reviews and testimonials: real customers only, no exceptions.
- Apologies and incident notices: your words, maybe AI to tighten grammar after.
- Legal, insurance, or regulated claims: out of scope for a chatbot.
When Not To Bother At All
If you send one email a month, building a voice file is overkill; just write the email. The method earns its keep when you're producing marketing regularly and the volume is what's killing you: weekly posts, quote follow-ups, seasonal campaigns, a backlog of pages you keep meaning to write. Below that volume, the honest answer is that the manual, human option is faster and better, and you should take it.
The other case where you shouldn't reach for AI is when the message is the relationship. A note to a client whose project went sideways, a real thank-you after a big job, a reference for a crew member. Those few minutes of your own writing are worth more than any draft, and everyone on the receiving end can tell the difference.