The best AI project is rarely the most impressive demo. It is usually a repetitive decision, handoff, or task that quietly consumes time every week.
Start with friction, not tools
Most teams begin by asking which AI product they should buy. That reverses the useful order. First list the work that is slow, repetitive, inconsistent, or dependent on one person remembering every detail. Then decide whether AI belongs in that workflow.
Good early candidates include summarising customer conversations, drafting first responses from approved information, classifying enquiries, turning meeting notes into actions, and preparing a first version of recurring reports. These tasks have clear inputs and outputs, and a person can review the result before it affects a customer.
Use three filters
- Frequency: Does the task happen often enough to matter?
- Clarity: Can you explain what a good result looks like?
- Risk: Can a person review the output before anything irreversible happens?
If a workflow scores well on all three, it is a stronger starting point than a large autonomous system. You can measure it quickly and learn without putting the business at risk.
Design the smallest useful system
A practical AI workflow normally needs more than a prompt. It needs a trigger, reliable context, an output format, a review step, and a record of what happened. For example, a new enquiry can trigger a summary and suggested reply, but the reply should use your service information and wait for approval before sending.
Keep the first version narrow. One team, one workflow, one measurable result. Track time saved, turnaround time, error rate, or conversion quality. If the workflow improves the number that matters, expand it. If it only creates novelty, stop.
Protect trust
Do not send sensitive customer data into unapproved tools. Do not let generated claims reach customers without review. Record where AI is used, who owns the workflow, and how a person can correct it. Useful automation should make responsibility clearer, not blur it.
Write down the five recurring tasks your team complains about most. Choose the one with high frequency, a clear definition of done, and a reversible output. That is usually where AI belongs first.
Where should a small business use AI first?
Start with a task that happens often, has a clear definition of done, and produces something a person can check before it reaches a customer. Summarising enquiries, drafting first replies from approved information, and turning meeting notes into actions all qualify. Name the workflow before you buy any tool.
How do I know whether a workflow is worth automating?
Score it on three filters. Frequency: does it happen often enough to matter? Clarity: can you explain what a good result looks like? Risk: can someone review the output before anything irreversible happens? A workflow that passes all three beats a large autonomous system.
Is a prompt enough to run an AI workflow?
No. A practical workflow needs a trigger, reliable context, a defined output format, a review step and a record of what happened. A prompt on its own produces a demo. The structure around it is what makes it dependable enough to leave running.
How do I measure whether AI is actually working?
Pick the number the workflow is meant to move – time saved, turnaround time, error rate or conversion quality – and track it from the start. If the workflow improves that number, expand it. If it only impresses people in demos, stop it.
What should never be automated?
Anything that sends unreviewed generated claims to customers, and anything that requires putting sensitive customer data into unapproved tools. Record where AI is used, who owns the workflow, and how a person can correct it. Automation should make responsibility clearer, not blur it.
Second opinion
Been circling the same problem for months?
Send me what you are working on. If I can see the fix in ten minutes, I will just tell you. If it is bigger than that, we scope it honestly before anyone talks money.
Tell me what is stuck