Start with a repeated task your team understands. You need to know what goes in, what a good result looks like and who is responsible when something goes wrong.
Look for repeated friction
For one working week, keep a simple record of tasks that create unnecessary repetition. Ask your team to record the task, how often it happens, roughly how long it takes and what makes it frustrating.
Examples might include preparing a research brief, summarising approved meeting notes or gathering information for a weekly update. Do not assume each needs AI. A better template, clearer ownership or a straightforward software integration might solve the problem.
Compare three candidates
Choose three tasks and answer these questions for each:
| Question | What to record |
|---|---|
| What is the task? | A specific action rather than a department or vague goal |
| How often does it happen? | Actual frequency and a representative sample |
| What does it cost today? | Time, rework and delays, not just the time spent typing |
| What would a good output look like? | An example and acceptance criteria |
| What information is involved? | Source, permissions and sensitivity |
| Who will check the result? | A named person with time to review it |
| What happens if it is wrong? | Consequences and a recovery step |
| Can we test it separately? | A small scope with a manual fallback |
Use the answers to choose an initial pilot with visible value and manageable consequences. Avoid starting with a process whose rules nobody can explain.
Decide what AI should do
Separate the task into steps. Some may be fixed rules, some may need drafting or interpretation, and some may require human decisions.
For a meeting workflow, the steps might be: receive approved notes, draft a summary, check the summary, confirm decisions and share the approved record. AI can support the draft without being allowed to declare a decision final.
Measure the whole process
Keep a baseline from several representative tasks. During the pilot, record preparation, review, correction and maintenance time as well as generation time.
A draft produced quickly is not an improvement if checking it takes longer than doing the original task. Review quality too: what was missed, what needed correction and whether the people using it understood the output.
Know when to stop or expand
Agree the review point before building. If the pilot produces usable results with manageable review effort, decide whether to expand it. If it fails, identify whether the issue is the process, the information, the tool or the scope.
Keep the manual route available while you learn. A useful first workflow should improve the way your team works, not create a new dependency nobody owns.
Want help choosing the right first step?
Book a deep consultation with Juan at Triple Pepper. We can explore your goals, review the processes taking up your team's time and identify a sensible direction for a pilot. The consultation's scope and fee are agreed before booking.
---