Your first AI project: pick one task and get it done in two weeks

Here’s a common way small businesses “try AI.” Someone signs up for a chatbot, asks it a few questions, gets an okay answer, and the tab sits open for a week before it gets closed. Nothing changes on Monday morning.

The tool usually isn’t the problem. Nobody gave it a job.

The fix is simple: pick one task, give it two weeks, and measure it honestly before and after. That’s how we plan every first project. One job at a time, checked by a person, kept only if it earns its spot.

Here’s the plan, roughly day by day.

Before day 1: pick the right task

A good first task checks most of these boxes:

• It happens often. Several times a week, at least. A task you do once a quarter won’t give you enough tries in two weeks.
• It has the same shape every time. Similar inputs, similar output. Site notes in, quote letter out.
• It’s easy to check. Someone can look at the result in a minute or two and know if it’s right.
• A mistake gets caught early. The work gets a human look before it reaches a customer, a bank or an inspector.
• Somebody hates doing it. That person will be your best tester.

Good first tasks look like this:

• Writing the quote cover letter from notes you scribbled at the job site
• Sorting the morning inbox and drafting replies to the routine emails
• Turning a voice memo recorded in the truck into a daily job report
• Answering the same five questions customers ask on the phone every week
• Logging photos of receipts into a sheet your bookkeeper can use

Bad first tasks: anything that moves money out the door, legal filings, hiring decisions, patient records, or “fix the whole office.” Those can come later, with more care. Some never get automated at all.

Now write the task down in one sentence. For example: “Turn my site-visit notes into a draft quote letter that I review and send.” If you can’t fit it in one sentence, it’s too big. Cut it down.

Week 1: measure it, set it up, try it on real work

Days 1–2: Time it the old way. This is the step everyone wants to skip. Don’t. Without a “before” number, you’ll never know if the “after” is better.

• Pick five to ten real examples. Not the easy ones. The normal mix.
• Time each one from start to finish. Count the hunting for the file, the phone call in the middle and the reread before you hit send. A phone stopwatch and a notepad are fine.
• Note the quality too. How many needed a redo? How many went out missing something?
• Write the numbers down that day. Memory makes a bad stopwatch.

Day 3: Set the tool up properly. Use a business account, not someone’s free personal login, especially if customer names or prices are involved. Our article on whether it’s safe to put business data into ChatGPT, Claude or Gemini explains what to look for.

Then write your instructions once and save them, so nobody has to retype them. Claude and ChatGPT both call this a Project, with its own saved instructions and files (Claude, ChatGPT). Gemini calls it a Gem. Google says Gems on personal accounts become “skills” starting in November 2026. The name changes, but the idea is the same. Good instructions cover:

• Who you are and who the work is for
• What a finished result looks like, with two or three past examples you were happy with
• Your standard wording, like exclusions, payment terms and sign-off
• What it must never do

Sample instructions: “You write quote cover letters for a masonry and stucco contractor in the Capital Region. Use my site notes below. Keep it under 200 words in plain English. List the scope as bullets. Always include our standard exclusions. Never promise a start date or a price that isn’t in my notes. If something is unclear, ask me instead of guessing.”

Days 4–5: Run it side by side. Do the task the old way and the new way on the same real jobs, then compare. Every time the AI misses, fix the instructions, not just the one result. Keep a running list called “what it got wrong.” That list becomes your training material next week.

One rule starts here and never ends: every result gets a human look before it goes out. AI tools can sound sure and still be wrong. Anthropic, the company that makes Claude, says plainly that Claude can give incorrect or misleading answers and shouldn’t be your only source of truth, especially on anything high-stakes. Treat every AI tool that way.

Week 2: hand it to the team

Days 6–7: Train the person who does the task most. Thirty minutes at their desk, on their own work. Not a slideshow. Show them:

• Where the saved instructions live
• How to feed it the notes, emails or photos
• How to check the result, using the “what it got wrong” list
• What to do when it’s wrong: fix it, and tell whoever owns the instructions

Days 8–9: They run it alone. You stay close. Questions go on a shared list instead of interrupting you each time. Expect a few rough spots. Most get fixed by changing a line in the instructions.

Day 10: Time it again. Same method as week 1, on five to ten new real examples. This time, count the minutes spent fixing the AI’s work. Leaving that out is how people fool themselves.

Then look at the numbers with a straight face. You’ll land in one of three places:

• Clearly faster, same quality or better. Keep it.
• About the same time, but fewer mistakes or less dread. That can still be worth it. Just call it what it is.
• Slower, or it needs heavy fixing. Either the instructions need work or it’s the wrong task. That’s a fine result for two weeks of testing.

No rounding up. If you want to turn your own before-and-after into dollars, our time-savings calculator does the math with your numbers.

Week 3: keep it, fix it or drop it

Keep it. Write a one-page how-to: where it lives, the steps, the check, and who to ask. Name one person who owns the instructions and updates them when something changes, like new prices or a new template. Put a reminder on the calendar to time it again in a month.

Fix it. Give it one more week with better examples or tighter instructions. If it still isn’t working after that, drop it.

Drop it. No shame. Write down why it didn’t work. That note will help you pick a better second task.

Then pick the next one. The second project is easier than the first. The accounts are set up, the team knows the routine, and checking AI work is already a habit. Look for the task right next door to the one that worked. If inbox drafts went well, quotes and proposals might be next (document automation). If answering the same phone questions went well, an AI receptionist might be.

Why first projects stall

• Too big. “Automate the office” isn’t a task. “Draft the quote letter” is.
• No before-number. Without one, the project ends in a shrug.
• Only the boss uses it. If the person who does the task every day never learns it, it fades the week you get busy.
• Skipping the check. One bad email to a customer and the team stops trusting the tool.
• Free personal accounts. Customer data in a free personal login is a problem waiting to happen.

Where we come in

You can run this plan on your own, and we’d honestly be glad if you did. If you’d like help, the AI Workflow Audit is “pick the right task,” done with you at your office. AI Setup and Team Training covers weeks 1 and 2, and the Monthly Care Plan covers week 3 and after. To see where this can lead, our paperwork autopilot is one of the systems I use at my own construction company, shown with sample data.

Not sure which task to start with? Take the AI readiness quiz or book a free visit.