Start with a task you can describe
Open last week’s calendar, inbox or task list. Look for work someone repeated: rewriting similar replies, turning notes into a follow-up, assembling a weekly summary, or reformatting approved product information. Write down three candidates before looking at tools.
Make each candidate small enough to have a clear beginning and end. “Improve customer service” is a department-level ambition. “Draft a reply to a delivery question using our published policy, then have a colleague check it” is a testable workflow. Record who does it, what they need, and what counts as finished.
OECD research on SMEs identifies both suitability to the work and practical adoption barriers. That supports asking whether a specific use fits before assuming the business needs a particular tool.
OECD: Generative AI and the SME Workforce (2025)Put each candidate through five questions
| Ask | What a useful first candidate looks like |
|---|---|
| Does it repeat? | It happens often enough to collect several comparable examples. |
| Can we check the result? | A reviewer can compare it with a known source or clear quality standard. |
| Is the information usable? | The right facts are available, current and approved for the chosen tool. |
| Can we contain a mistake? | The output can stay in draft and a person can stop or reverse the change. |
| Will the result matter? | Better handling would help an actual bottleneck, customer need or team priority. |
Use the answers to discuss trade-offs, not to generate a magic total. A frequent task with uncheckable outputs is a weak first pilot. A modest task with reliable inputs may teach you more. If nobody can make time to review the output, the workflow is not ready to go live.
Compare real work, not industry labels
The same shop may have a promising catalog task and an unsuitable advice task. The same consultancy may have a useful formatting experiment and a high-stakes judgment it should keep with a qualified person. These are illustrative starting points, not recommendations for every business in that category.
- Retail: turn approved product facts into a draft description. Keep safety and suitability claims out unless an appropriate person verifies them.
- Services: turn your own non-sensitive meeting notes into a draft action list. Check decisions, owners and dates against the original notes.
- Operations: summarize a prepared weekly report. Reconcile every figure with the source and investigate missing information.
A field experiment with consultants found that AI’s usefulness varied across tasks. Treat that as a reason to test the actual workflow, not as a savings estimate for your team.
Harvard Business School: Navigating the Jagged Technological FrontierWhen preparation is the better first move
If prices, policies or instructions contradict one another, first agree which source is current. If the task exposes personal, confidential or regulated information, establish an appropriate data and approval process before trying it. You can use invented, non-sensitive examples to explore the interface while that work is done.
Write a one-paragraph experiment brief
Complete this before choosing a tool
For [one recurring task], we will test [one AI-assisted step] using [approved source information]. [Named person] will check [specific quality criteria] before the output is used. We will compare handling time, review, corrections and quality against our current process. We will stop if [defined failure]. We will decide what to do next on [date].
Share the brief with the person who normally does the work. Ask what it leaves out, especially the exceptions. Choose one workflow, one owner and one review date. Leave integrations and a wider rollout until the small test gives you a reason to proceed.
Original practical guidance by Future of My Business. Linked sources support the adjacent research or risk context; our checklists and examples are editorial tools, not validated benchmarks or endorsements.
How our assessment works