Pick one recurring task you can check yourself, run it with a worked prompt, and time it honestly before you decide what comes next.
Skip the big strategy question for your first AI task. Pick one recurring task that takes 30 minutes or less, carries little risk and produces something you can check against your own source. Run it three times with a worked prompt, check every line, time the whole job against your usual way, and only then decide whether to keep it, fix it or move on.
Many first serious attempts at AI are too big to judge. People ask for a marketing strategy, a whole deck or an analysis of a market they half know, get back something polished, and have no way to tell whether it’s any good. So they either trust it too much or decide the tool is hype, and neither conclusion teaches them much. A quick win with AI works the other way round: one small, recurring task where you already know what good looks like.
A brand manager asks a chat window for next quarter’s social strategy late on a Friday. Two pages arrive, with content pillars and a calendar. It reads well. It could also have been written for any brand in the category, and there’s no way to check it short of running the quarter.
Turning this morning’s meeting notes into a follow-up email is a different kind of task, because whoever took the notes knows who agreed to what. Within a minute of reading the draft, they can see what it got right, what it made up and what it dropped. That’s what a first win is for: it shows you how the tool behaves on work you can mark, so the trust you build is in your own reading of the output, not in how confident the output sounds.
Start with work you can mark yourself.
So the bar for a first task is narrow on purpose. It’s one recurring task that takes you 30 minutes or less the usual way, with an output you can check against something you already hold. This guide gives you a way to pick that task, run it with a worked prompt, time it honestly and decide what comes next. If you haven’t written a prompt before, our ChatGPT starter guide covers the first steps.
Four tests separate a good first task from a frustrating one, and none of them is about the tool. The task should come round often, so you get several tries in a month. It should be low-risk, so a missed error costs you an awkward correction rather than a client problem. It should be checkable against something you hold, like your notes or the source document. And you should already know what a good version looks like, because that’s what lets you spot a weak one quickly.
Checkable carries the most weight early on, and there’s research behind that. In a preregistered experiment with 758 consultants, run with Boston Consulting Group and published in Organization Science this year, people using AI on a task chosen to sit outside the tool’s capabilities were 19% less likely to produce correct solutions than people without it.[1] The authors call the edge of those capabilities a jagged frontier. What I’d take from it is that the screen gives you no signal about which side of that edge you’re on, so a task you can check is how you learn where it runs for your own work.
| Test | Scores 0 | Scores 1 | Scores 2 |
|---|---|---|---|
| Frequent | Once a quarter or less | About monthly | Weekly or more |
| Low-risk | A miss reaches a client, the budget or a regulator | A miss reaches a colleague | A miss stays with you until you fix it |
| Checkable | You’d need fresh research to check it | You can check most of it against a source | You can check all of it against something you hold |
| You know good | You rarely do this task | You do it, but find quality hard to judge | You’d spot a weak version straight away |
Score each candidate out of 8. Six or more makes a good first win, as long as it takes 30 minutes or less by hand. A 0 on low-risk or checkable means it waits, whatever the total. A Future Factors working tool, not a validated instrument.
Don’t score one idea and run with it. Write down three candidates from the last couple of weeks, because the first idea that comes to mind is usually the most exciting one, and exciting tends to mean big. These are the places I’d look:
Here it is filled in for common marketing work. The strategy question from the opening scores lowest, though it’s the one people tend to try first.
| Role | Candidate task | F / L / C / K | Total | Verdict |
|---|---|---|---|---|
| Marketing manager | Follow-up email from rough meeting notes | 2 / 2 / 2 / 2 | 8 | Pick this first |
| Content marketer | One product description rewritten in three lengths for web, email and social | 2 / 1 / 2 / 2 | 7 | Good second task |
| Marketing ops | Tidying campaign names in an export to match the naming convention | 2 / 2 / 2 / 1 | 7 | Good, if you own the convention |
| Analyst | Summarising open-text survey comments for a board pack | 1 / 0 / 1 / 2 | 4 | Waits: low-risk scores 0 |
| Brand manager | Next quarter’s social strategy | 0 / 1 / 0 / 1 | 2 | Not a first task |
F = frequent, L = low-risk, C = checkable, K = you know good. Example scores to show the method; your own will differ, which is the point.
The same scorecard lands differently depending on the seat. Content marketers usually have rewrites that score well. An analyst’s best first candidates tend to be formatting and tidying jobs rather than interpretation, because interpretation is the part that’s hardest to check. A team lead’s tend to be the recurring updates and follow-ups that eat the end of the day.
The follow-up email makes a good worked example because it scores well for most roles and the check is quick. A marketing manager comes out of a campaign kickoff with a page of half-sentences: who’s taking the landing page, a date someone mentioned for the brief, a budget question nobody answered. Writing that up takes longer than it should.
The prompt below follows Anthropic’s own prompting guidance, which suggests treating the model like a brilliant but new employee who lacks context on your norms and workflows, and testing a prompt by asking whether a colleague with little context could follow it.[2] So it says who the reader is, what the email is for and, most important for a first win, what the model must not add. Our piece on why prompting matters explains why those context lines change the result.
| You’re helping me write a follow-up email after a meeting. Context: [WHAT THE MEETING WAS FOR AND WHO WAS THERE] Reader: [WHO GETS THE EMAIL AND WHAT THEY NEED FROM IT] Tone: [FOR EXAMPLE: FRIENDLY AND DIRECT, THE WAY I USUALLY WRITE TO THIS TEAM] Write an email of no more than 200 words that lists (1) the decisions we made and (2) the actions, each with an owner and a date. Use only what is in my notes. Do not add decisions, owners, dates or names that are not there. If an action has no owner or no date, leave it blank in the email. After the email, give me a separate list headed ‘Unclear in your notes’ with anything you had to guess or could not place. That list is for me, not for the email. My notes: [PASTE YOUR NOTES AS THEY ARE] |
Fill the brackets in. The ‘Unclear in your notes’ list is what turns a draft into something you can check.
Run it in four steps:
On the first step, the account matters more than the tool. OpenAI says it doesn’t train its models on an organization’s data by default for business products such as ChatGPT Enterprise, Business and Edu.[3] That statement covers those products. It doesn’t describe a personal account, so if you’re not sure which one you’re signed into, ask before you paste a client’s name.
A draft that reads smoothly is exactly when to slow down. OpenAI’s own help page says ChatGPT can sound confident even when it’s wrong, and tells people to use it as “a first draft, not a final source.”[4] For a follow-up email, wrong rarely looks dramatic. It looks like a plausible deadline nobody said.
| Check | What a miss looks like | How to check it |
|---|---|---|
| Owners | An action assigned to the wrong person, or to “the team” | Match each owner to a name in your notes |
| Dates | A reasonable-sounding deadline that nobody set | Every date in the email should appear in your notes |
| Decisions | “We agreed” for something that was only discussed | Look for agreed, approved and confirmed, and check each one |
| Gaps | An action in your notes that never made it into the email | Scan your notes for verbs and tick each one off |
| Unclear list | Open questions that quietly turn into assumptions | Resolve each item, or move it to your own to-do list |
| Tone | Phrases you’d never use with this person | Read the first and last lines out loud |
In our view the first three rows catch most of what goes wrong in this kind of draft. Check line by line for the first few runs.
It helps to write the split down once.
| What AI does | What you still own | How it gets checked |
|---|---|---|
| Turns rough notes into a structured draft | What was actually decided, and who owns each action | Every owner and date matched against your notes |
| Lists what it couldn’t place | Resolving each open question with the right person | Nothing on the unclear list goes out unresolved |
| Matches the tone you describe | Whether it sounds like you, and whether it goes out at all | You read it in full before you hit send |
A Future Factors working view of the split for this task.
Over a few runs you learn which mistakes this tool makes on your notes, which no demo can show you. For bigger tasks, the anti-hallucination toolkit goes further on catching errors.
Speed is the part people misjudge most. The draft arrives in seconds, so the job feels fast, and the time spent checking and fixing afterwards doesn’t register in the same way.
There’s a striking study on this gap, with caveats worth stating up front. In early 2025, METR, a research nonprofit, ran a controlled trial with experienced open-source developers working in code they knew well: with AI tools they took 19% longer, yet afterwards they believed AI had sped them up by 20%.[5] It was software work with a small group, and METR now marks the result as out of date, with a 2026 follow-up suggesting newer tools probably do speed developers up, on what it calls very weak evidence about the size.[6] The finding that travels is the distance between feeling fast and being fast.
So time the whole job, from opening your notes to a version you’d send, and time your usual way first so you have something to compare against. Don’t quote anyone a number until you have at least three runs.
Time it to sendable, not to first draft.
| Run | Draft time | Checking and fixing | Total to sendable | What I fixed | Send as is? |
|---|---|---|---|---|---|
| Usual way (baseline) | 25 min | None | 25 min | Not applicable | Yes |
| AI run 1 | 5 min | 25 min | 30 min | Invented a date; one action missing | No |
| AI run 2 | 5 min | 20 min | 25 min | Added a tone line to the prompt | After edits |
| AI run 3 |
Placeholder numbers to show the format, not a result or a promise. Notice the example’s first runs are no faster; yours may not be either, and that is fine while you learn the task.
Time is only one reading. The other two are whether you’re still using the prompt a month from now without reminding yourself, and whether the email itself got better or went out sooner. Use, persistence and impact, in that order. If you later want to measure this across a team, our guide to tracking time savings from AI covers baselines and rework properly.
After three runs, the log usually points one of four ways.
Run one task until it’s boring before you add a second.
You can get to this point on your own, and plenty of people do. Structured training shortens the learning curve and builds shared capability across a team, so the habit doesn’t live with one enthusiastic person. In those terms, training builds the capability, a task like this follow-up email is the workflow design that gives it somewhere to be used, and repeating it each week is the reinforcement that turns it into behavior. Adoption is what you see when that sticks. Our guide to what makes someone an AI-powered professional covers the bigger picture, and we’ve compared beginner AI courses if you’d like a structured route.
When you share the win, share the prompt and the check table, not a claim about minutes saved.
A one-week start. Runs two and three can spill into the following week.
Pick the task you do most often and would least mind getting slightly wrong. Write down how long it takes you now. Then run it once this week and check every line before it goes out under your name.
Use the assistant your organization has approved for work. For a first win the task and the check matter more than the brand, and an approved account settles the data question up front.
That’s common on the first run, because you’re learning the task and the tool at the same time. Add one line to the prompt that fixes the main mistake and run it twice more. If fixing still takes longer than writing it yourself after three runs, switch to your next candidate.
Only into an account your organization has approved for that data. Vendor statements about not training on your data generally describe business products, not personal accounts. When in doubt, swap client names for roles first.
When you’d send the draft after a light check, the prompt hasn’t needed changing for a couple of runs, and you’re still using it without reminding yourself. That’s persistence, and it’s a better signal than one fast run.
Share the prompt, the check table and your log, with the caveat that it’s three runs on one task. That’s more credible than a headline number, and it lets your manager judge whether the task would work for others.
I checked every source listed below on October 21, 2026. I didn’t run the worked-example prompt for this article and don’t report results from it; the time log uses placeholder numbers to show the format. I left out time-savings percentages from studies, because a first win shouldn’t be sold on someone else’s number. The scorecard, check table and rules are Future Factors working methods, not research-validated tools, and the scenes are illustrations, not client stories.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.