“Where could we use AI?” is a broad question. A more useful one is: “Which recurring task has a clear input, a predictable next step, and an owner who can review the output?” That question gives a team something concrete to examine before choosing a tool.
A first project should be narrow enough to understand and useful enough to matter. It might prepare a sales follow-up list, summarize routine documents, or turn approved updates into a draft. The goal is to improve a specific workflow, then learn from the result.
1. Does the task happen often enough?
Look for work that comes back regularly: every day, every week, or every time a lead arrives. A rare task can still be valuable, but repetition makes it easier to see the current process, test the new one, and notice whether it helps.
Write down what starts the task and what “done” looks like. If the team cannot agree on those two points, map the process before automating it.
2. Are the inputs available and usable?
AI needs the right information at the right moment. Identify the inboxes, forms, files, spreadsheets, or systems involved. Check whether the necessary data can be accessed, whether it is current, and whether it is consistent enough to support the task.
A workflow that depends on missing or contradictory information will need a data cleanup step or a human decision point. That is useful to discover early.
3. Can you describe a good output?
Make the expected result visible. A useful output might be a draft response, a set of extracted fields, a prioritized list with reasons, or a short exception report. Collect a few real examples of good work and define what must be included.
A simple test: Can someone on the team review the output and say what is correct, what needs changing, and what should happen next?
4. Where should a person stay involved?
Decide which steps are preparation and which are decisions. A system might organize requests or draft a message, while a person approves the message, handles unusual cases, or makes a customer-facing judgment. The right checkpoint depends on the process and the risk of a mistake.
The NIST AI Risk Management Framework encourages organizations to consider context, measurement, and risk management throughout an AI system’s life. Its guidance on mapping also calls for identifying where human oversight is needed. That is a useful lens for a small business pilot, even if the pilot itself is modest.
5. What would success look like?
Choose a measure before building. It could be fewer manual handoffs, faster preparation, more complete records, or less time spent collecting information. Include a quality check: saving time is not an improvement if the output creates rework.
Compare a small sample of the current process with the pilot process. Ask the people doing the work whether the change is genuinely useful.
6. Can you start with one contained slice?
Begin with a single trigger, one team, and one review path. For example, prepare a draft follow-up after a web inquiry, then let a team member check it before it goes anywhere. Keep exceptions visible. Expand only after the team understands how it behaves on real cases.
A quick way to pick a candidate
List three recurring tasks, then score each one from 1 to 3 on these factors:
- Frequency: How often does it happen?
- Clarity: Can you define the input and expected output?
- Access: Is the needed information available?
- Review: Is there a clear owner for exceptions and approval?
- Value: Would improving it matter to the team?
The highest score is a conversation starter, not an automatic decision. Use it to choose a process worth mapping in more detail. A focused pilot can reveal whether the workflow is feasible, what needs human attention, and what to improve next.
Have a workflow in mind?
Tell us where work gets stuck. We can talk through the process and identify a sensible first step.
Talk through a process ↗