Find work that is worth improving
Do not buy a tool first and then look for somewhere to use it. Speak to the people who know the work. Ask them to show you where they spend too long looking for information, repeatedly write similar text or correct the same types of mistakes.
Choose something that happens often enough for an improvement to matter. It might be finding the right procedure in internal documents, drafting a response to a common customer enquiry or helping a customer choose the correct application form.
Narrow the use case
“Use AI in customer service” is too broad for a first trial. “Draft a response to delivery questions using approved help content” is specific enough to test. Define who will use the system, the type of case it will help with and the sources it may use.
Also decide what the system must not do. It may need to avoid complaints, personal data or sending anything directly to a customer. A clear boundary makes it easier to test quality and identify what must be resolved before further development.
- Choose one user group and one type of work.
- Use a limited set of approved sources.
- Agree who will check the output.
- Describe the situations the system must not handle.
Describe how the work should be done
Write down what happens from the moment a need arises until the work is complete. Who receives the case? Where do they find information? What can AI draft or suggest? What must an employee check before the output is used or sent?
Include difficult situations. If the sources are incomplete, the system should say that it lacks enough information. If a customer asks about something outside its remit, the case should go to an employee. Make these choices before selecting a model or technical approach.
Test with examples from day-to-day work
Collect 10–30 examples that resemble what the system will encounter. Include common cases, incomplete questions, conflicting sources and cases where it should decline to answer. Remove personal data if it is not needed for the test.
Ask subject specialists to assess the results. They should check facts and sources, record what needs correcting and compare the time required with the current process. A polished answer is not useful if checking it takes longer than writing it from scratch.
- Choose day-to-day tasks.Use cases employees already know from their work.
- Define what to assess.For example factual accuracy, relevant sources, time saved and corrections required.
- Record the results.Count what can be used, what needs changing and what should be rejected.
Use the test to decide what happens next
The test may show that the idea is worth developing. It may instead reveal a need for better source material, a narrower use case or a conventional search or form solution rather than AI. The trial has still done its job: it has provided a concrete answer before the investment became large.
If you need help, I can define the use case, build the prototype and run the test with your subject specialists. You will then know what works, what needs to change and what it would take to introduce the system.