The opportunity canvas
Look closely.
Choose a small test.
Use this guide to turn a broad AI idea into a workflow question that your team can inspect. Begin with one repeated piece of work and the people closest to it.
A useful first pilot has a clear trigger, approved inputs, an observable output, a reviewer and an exception route. If one is unclear, make that the next question to resolve.
1. Describe the work as it happens.
Choose a recent ordinary example. Write down what triggered the work, which information was available, what someone had to do and where the result went. Then inspect an incomplete example. A difficult case often exposes a hidden decision or dependency.
Keep the description concrete: “the coordinator searches three documents before answering a request” is more useful than “we need better productivity”. Do not begin by assuming which tool will solve it.
2. Find the constraint.
Ask the process owner what makes this task hard to finish correctly. Is it missing information, repeated data entry, unclear responsibility, conflicting sources or a review queue? Some problems need a process or information change before automation is sensible.
3. Make the hypothesis explicit.
Try this structure: “If we change this step, using these approved inputs, this person should be able to inspect this output.” Add the condition that would make the hypothesis wrong.
4. Agree what good work looks like.
List what a reviewer should check: required fields, factual accuracy, source references, completeness and the absence of unapproved commitments. Define what happens when information is missing or inconsistent. A plausible answer should never hide an unresolved question.
5. Compare a bounded pilot with today.
Use an agreed set of representative examples. Record the current effort and common corrections before the trial. During the pilot, note what still needs manual work, what fails and how the exception route behaves. A small test is for learning; it does not establish a return across the whole business.
6. Decide together.
Bring the observations to the process owner. Continue only if the result is useful under the agreed conditions. You may need to change the scope, improve the information, use a simpler method or stop. Record the decision and the next review point.
Check the information before the tool.
Use fictional or redacted examples while access and data handling are being assessed. Tool terms, retention, permissions and who can see the information all matter. Refer to the primary guidance below and involve your organisation's accountable people where needed.
Bring one real constraint
Make the next step a useful one.
A practical conversation starts with the work.