AI at work
How to choose your first useful AI task at work
Start with one recurring task you understand well. A short trial can show whether AI helps once preparation, checking and corrections are included.
Direct answer
Choose a recurring task with available source material, a clear result and a reviewer who can spot mistakes. Test it on a few representative examples. Compare the complete effort with your current process, including preparation and corrections, before deciding whether to keep using it.
Choose a small, familiar task
List three things you repeat during an ordinary week. For each, describe what arrives, what you produce and what makes the work frustrating. “Write a reply asking for the missing project details” is specific enough to try. “Improve customer service” needs a smaller starting point.
Good first candidates often involve organizing material you already understand: a draft reply, a checklist from process notes, or a short update from confirmed decisions. You should be able to recognize a useful result without asking the AI whether it is correct.
Set aside tasks where a mistake could be hard to notice or expensive to reverse. An initial writing trial does not need customer-record access, automatic sending or permission to change your website.
Define the input, result and review
Use the AI task pilot worksheet (TXT) to make the trial concrete. Answer these questions before opening a tool:
- Input: Which notes, approved facts or examples does the task require? Who can confirm they are current?
- Result: What should the finished item contain, and where will someone use it?
- Checks: Which details must match the source? Which omissions would make the result unusable?
- Reviewer: Who knows enough to approve it, and what happens when information is missing?
- Boundary: Where does the AI stop? For a first trial, a saved draft is a useful stopping point.
Use fictional or approved, de-identified examples while learning. Confirm which account and information your business permits before using real records. Removing a name alone may leave identifying details in the surrounding text.
Work through a fictional example
A small studio repeatedly turns internal notes into a customer update. Its notes say: two draft layouts are ready; the customer must select one; image permissions remain unconfirmed; no delivery date has been approved.
The task is to organize those notes into three sections: completed work, decisions needed and open questions. The owner checks every statement against the notes before sending anything.
A suitable instruction is: “Prepare a short customer update from these notes. Include only confirmed facts. Put missing decisions in a separate section. Do not add a delivery date, make commitments or send the message.”
Try the same instructions with an incomplete example, too. If the output turns “waiting for image permissions” into “images approved,” record that as a factual error, correct the draft and revise the instructions. A polished tone does not compensate for a changed fact.
Compare the whole process
Record how long the ordinary process takes on several similar tasks. Then trial the AI-assisted approach with comparable examples. Use the same completion standard: a reviewed result that is ready for its intended use.
| Record | What to include |
|---|---|
| Preparation | Finding sources, removing unsuitable details and writing instructions |
| Drafting and waiting | Time spent producing the first result and requesting changes |
| Review and corrections | Checking facts, restoring omissions and editing the final version |
| Quality | Errors, missing details and whether the result meets the original checks |
Track setup time separately from repeat-use time. A fictional comparison of 12 minutes manually against 3 minutes preparing, 2 minutes drafting and 8 minutes reviewing means 13 minutes with AI. That trial has not saved time. The numbers illustrate the calculation; they are not measured Plainview results.
Decide what to keep
Keep the process when the reviewed output meets your standard and the overall effort is worthwhile. Narrow the task if one part works better than the rest. Stop if checking takes too long, essential facts are unavailable or nobody can reliably review the answer.
Save the instructions, an approved example, the source requirements and the checking steps. Ask another intended user to try them without coaching. Their questions will show where the handoff needs work.
For a more detailed practice task, continue with reviewing an AI customer follow-up or preparing a website draft from business notes.
Primary sources and further reading
OpenAI's prompting guide describes defining acceptance criteria and testing answers against known examples. The NIST Generative AI Profile (PDF) explains why inaccurate generated statements and human oversight matter. The small-task worksheet and trial method here are Plainview's practical starting approach, not a certification or guarantee.