Last month a property manager I know spent her entire Monday morning writing forty near-identical emails to tenants about a lift repair. Same message, forty times, swapping the flat number and the appointment slot. It took her just under three hours.
She had tried a chatbot once before and given up, because the drafts came back sounding like a corporate press release. What fixed it wasn’t a better tool. It was a better sequence of steps. Here’s the one she now uses, laid out so you can copy it onto whatever task is eating your week.
Step 1: Choose one task, and put a number on it
Don’t start with “how could we use artificial intelligence?” That question goes nowhere. Start with a specific recurring task and time yourself doing it once, badly, by hand. Good candidates look like this:
- Writing the Friday project update for a client (25 minutes, every week)
- Turning a 60-minute discovery call recording into a proposal outline (40 minutes)
- Answering the same six onboarding questions from new customers (15 minutes each)
Multiply the single run by how often it happens. Forty emails at four and a half minutes each is three hours a week, roughly 150 hours a year. That figure is what makes the rest of this worth doing, and it’s the number you’ll use later to judge whether the fix worked.
Step 2: Write a brief, not a prompt
“Write an email about a lift repair” gets you generic mush. Output is only ever as specific as input, and most people are far vaguer with a model than they’d dare be with a human colleague. Brief a new freelancer with one sentence and you’d expect rubbish back.
The five things every brief needs
- Who it’s for: “a tenant who has already complained twice and is frustrated”
- What the reader should do next: “confirm which of the two time slots works”
- The shape: “under 120 words, one date, one call to action”
- Tone: “plain, warm, no apology tour, no corporate phrases”
- What to avoid: “don’t invent dates, don’t promise compensation”
That last one matters more than people expect. Telling the model what not to do removes half the cleanup work later.
Step 3: Paste in your real material
Models don’t know your lease terms, your pricing or your house style. So give them something to work from: the previous email a human approved, the three bullets from the engineer’s job sheet, the actual clause reference.
This is the biggest single quality jump in the whole process. Two samples of your own writing teach tone better than any adjective you could type. Three sentences of genuine context beat three paragraphs of “please be professional.”
Step 4: Ask for options, then react out loud
Don’t accept the first draft. Ask for three versions with different angles, then say plainly what’s wrong with each. “Version two is closest. Cut the opening sentence, the date should be 14 March, and stop saying ‘we sincerely apologise’.”
Two rounds of that beats ten rounds of vague grumbling. If it keeps missing, the brief is the problem rather than the model. Go back to step 2 and add the constraint you forgot.
Step 5: Verify anything a human would check
Dates, names, numbers, legal references, quotes: check each one against the source before anything leaves your desk. A model that confidently invents a policy number is behaving exactly as designed, because it’s built to produce plausible language rather than truth. Reading up on what artificial intelligence can and can’t do before you trust it with client work is time well spent.
The same scepticism applies to the tools themselves. Plenty of paid AI add-ons now wrap a thin interface around a general-purpose model and charge a monthly fee for it. Learning how to spot a real artificial intelligence product from a reskinned one takes about ten minutes and can protect an entire software budget.
Step 6: Save the version that worked
The conversation where you finally nailed the brief? Keep it, along with the tone samples and the constraints, in a shared note. Next Monday you paste it in and change two facts. The property manager’s three-hour job now takes twenty minutes, most of which is reading forty drafts instead of writing them.
Store these as team assets rather than personal chat history, so they survive someone leaving. It’s also worth tracking who is shipping what. The artificial intelligence companies leading the market in 2025 mostly sell workflow products now rather than raw chat, which helps you decide what’s actually worth a subscription.
Step 7: Add guardrails before you scale
One person quietly saving time is fine. Twenty people doing it without rules is how you end up in a data protection meeting. Four guardrails cover most of it:
- No customer data, contracts or health information into any tool the company hasn’t approved
- Anything client-facing gets a human read before it’s sent
- Log the tasks that went wrong. Those failures are your best training material
- Re-test the workflow each quarter, since last year’s model needed a lot more hand-holding
Teams that skip this stage usually get a quiet backlash: one bad email, one invented statistic, and everyone drifts back to doing it manually. If you’re rolling this out beyond a handful of people, the rollout lessons from artificial intelligence in education transfer surprisingly well. Teachers ran this experiment at scale before most offices did.
The last mile still belongs to a person
The tool writes the fortieth email. You decide that the tenant in flat 12B shouldn’t get the standard one, because her mother is in hospital and the lift is the only way out of the building. You know which sentence will calm a complaint and which one will inflame it. You know when the lease actually permits something the template says it doesn’t.
That judgment is the job, and it’s getting more valuable, not less. Long-term questions about machines that genuinely think are worth following, but the practical skill for the next twelve months is narrower and much more boring than the headlines suggest: briefs, real context, verification, saved templates, human review.
So pick your task this week. Do it by hand once and time it. Write the brief, feed it the material, push back twice. If the first run saves you twenty minutes, you’re already ahead of the two hours most people spend reading about prompt engineering.