Last Tuesday, a freelance copywriter I know spent four hours on a job that AI could have handled in about eleven minutes. She wasn’t avoiding the technology. She had seven tabs open, three subscriptions she’d forgotten buying, and no idea which one to point at the work in front of her.
That’s the actual bottleneck. There are hundreds of decent AI tools now, and picking the right one for a specific task is worth far more than knowing about all of them. What follows is the five-step process I run with small teams and solo operators, along with the timings and gotchas that come up every single time.
Step 1: Write down the job before you open a single tab
Tools are easy to browse and hard to choose, so flip the order. Start with three tasks you did last week that felt repetitive. Be specific.
- Writing descriptions for 12 new products in a Shopify store
- Summarising 40 customer support tickets into a themes report
- Turning a 45-minute client interview into show notes and eight social posts
Each of those is a candidate. A goal like “use AI more” is not.
The three-question filter
Before you spend an evening on setup, run the task through this:
- Can I check the output faster than I could create it? If reviewing takes as long as writing, you’ve saved nothing.
- Is the raw material already digital? Phone photos of handwriting and creased scans are where most “AI didn’t work” stories start.
- Would a 70% draft still be useful? Outlines, first drafts and summaries: yes. Final legal or medical advice: no.
If a task fails any of those, park it. Plenty do. Knowing which is which early saves months of aimless tinkering, and the tools that genuinely last tend to sit in the narrow band of AI work that actually pays off, where the output is text, verifiable and disposable.
Step 2: Audition tools on your own work, not a demo
Twenty minutes is enough to judge a tool. Not twenty minutes of watching someone else’s screen recording, but twenty minutes with your brief, your data and your tone of voice.
Here’s the routine I run:
- Minutes 0-3: paste a real task, unedited, including the messy context you’d normally give a colleague.
- Minutes 4-10: run it twice more without changing anything. Consistency matters more than a lucky first hit.
- Minutes 11-17: fix the output by hand and time yourself doing it. That repair time is the true cost.
- Minutes 18-20: write one line. Saved X minutes, needed Y minutes of cleanup.
The same logic works for images, except you’re comparing prompt adherence and how badly hands, logos and lettering come out. Start from a shortlist rather than the top search result, because the best free AI image generators vary enormously in how much rework they create. A few free tiers also quietly throttle you after a handful of generations, which is exactly the kind of thing a 20-minute test reveals before you build a habit around it.
Step 3: Chain two or three tools instead of hunting for one perfect one
No single AI tool handles a full workflow well. The people getting real leverage are chaining them, and the good chains rarely run longer than three links.
Concrete example: a 45-minute podcast episode turned into a newsletter and five social posts.
- Transcribe. A speech-to-text model turns the audio into a messy transcript. Ten minutes, mostly waiting.
- Extract. A long-context chat model reads the transcript and pulls out the three strongest arguments with timestamps.
- Draft. The same model writes a 600-word newsletter from those three points in your voice, using two previous issues as style examples.
- Illustrate. An image generator produces a header from a prompt describing the episode’s central metaphor.
Hands-on time: around 35 minutes, plus 20 minutes of your own editing to put the sentences back the way you’d say them. That last part isn’t optional, and anyone claiming otherwise is selling something.
If your work already lives in Google’s ecosystem you can shorten the chain, since Google’s own AI tools now summarise transcripts and draft documents inside the apps you already have open, which removes a copy-paste step or two.
Step 4: Write prompts like a brief you’d hand a freelancer
Most disappointing output is a brief problem, not a model problem. Compare these two.
Weak: “Write a blog post about our new coffee subscription.”
Workable: “You’re writing for a specialty coffee brand whose readers already own a grinder. Write a 900-word post for people deciding between a weekly and a fortnightly subscription. Include a section on storage and one on cost per cup at £14 a bag. Tone: direct, no exclamation marks. Open with the storage problem, not a greeting.”
The second version names the reader, the decision, the structure, the constraints and the opening move. It also gives the model somewhere useful to be wrong, so you can correct one thing instead of rewriting everything.
Keep a prompt file
Every prompt that produces something usable goes into one plain text file, sorted by task. After a month you’ll have ten or twelve you reach for constantly and the setup cost drops close to zero. This is the step people skip, and it’s the one that compounds.
Step 5: Set a kill switch before you get attached
Subscriptions are sticky and tool enthusiasm fades slowly, so decide the exit criteria up front. If a tool hasn’t saved you at least 30 minutes in two weeks, cancel it.
Track it roughly. A tally of minutes saved in your phone notes is plenty. You’ll be surprised how many tools look impressive in week one and contribute nothing by week three. Free tiers make this test cheap to run, and free AI tools worth using tend to survive on their own merits rather than on a launch-week discount.
A worked example, with numbers
A solo consultant I worked with published a weekly newsletter. Her old process: two hours reading and note-taking, three hours drafting, forty minutes finding an image, twenty minutes formatting. Just under six hours, every Sunday.
After running the five steps, it became: 35 minutes feeding source material to a research assistant and asking for the five strongest points, 50 minutes drafting with a saved prompt and a folder of past issues as examples, 8 minutes on imagery with a generator she’d tested properly, and 20 minutes editing and formatting. Roughly 1 hour 55 minutes, plus a 15-minute review on Monday when her eyes were fresh.
Close to four hours back each week. Not one of those hours came from a single magic tool. They came from keeping the chain short and testing every link.
Where this falls apart
Some tasks will resist every version of this process, and it’s worth naming them.
- Anything confidential. Client contracts, medical details, unpublished financials. Read the data policy or don’t paste it.
- Anything needing verified citations. Models produce confident, plausible references that occasionally don’t exist. Check every one.
- Anything where your voice is the product. A ghostwritten op-ed or a wedding speech still needs you writing the sentences.
- Anything about people. Performance feedback, delicate emails, apologies. Tools flatten nuance precisely where nuance is the whole job.
What to do on Monday morning
Pick one task from your list of three. Run the 20-minute audition on two tools for it. Write down the minutes saved and the minutes spent cleaning up. If the maths works, chain a second tool onto it next week. If it doesn’t, drop it and move to the next task on the list.
Five tasks in, you’ll have a small, unglamorous stack that reliably hands back an afternoon a week. That’s the version of AI tools that changes how your week feels, and it looks nothing like the demos.