Collect problems before discussing solutions
Ask employees to show you where work gets stuck. Look for information that is difficult to find, questions answered repeatedly, documents that need to be read and summarised, or forms that customers often complete incorrectly.
Describe the situation without deciding that the solution must be a chatbot, an agent or a particular model. “A caseworker spends 20 minutes finding the current procedure across three systems” is a problem you can investigate. “We need an AI assistant” is only a proposed solution.
Assess each use case using the same questions
Create a simple overview and give each use case a provisional assessment. The numbers do not need to be precise. The purpose is to ask the same questions about each use case and make disagreements visible.
- How often does this happen, and how much time or frustration does it cause?
- Are the sources digital, and are you permitted to use them?
- Can subject specialists describe a good result?
- Can an employee check the output before it has significant consequences?
- What would actually happen if the system were wrong?
Start with work that happens often, uses known sources and allows an employee to correct the answer before it is used. Defer use cases where one wrong answer could mislead a customer, cause an application to be handled incorrectly or affect how an employee is assessed.
Look for work where AI has a clear role
AI can help when employees need to find relevant content across many documents, combine information, sort large volumes of text or create a draft that will be checked anyway. Examples include a source-linked answer from internal procedures, a summary of consultation responses or draft product copy based on approved facts.
AI can also form part of a customer-facing service, for example by explaining complex terms or asking questions that direct a customer to the right form. The team must then decide what advice the system may give, which sources it uses and when an employee should take over.
Defer use cases you cannot check
Do not start with work where you lack reliable sources, cannot describe a correct result or have nobody qualified to assess the output. Take particular care if the system would use extensive personal data or make decisions with serious consequences for individuals.
Some problems are better solved by updating documents, simplifying rules, improving a form or collecting the information in one system. Do that first if this is where the work gets stuck. If you later add AI support, the system can use up-to-date sources and employees know who must check its answers.
Choose one use case you can try within a few weeks
Describe the use case on one page: who does the work today, what they need help with, which sources the system may use, what it should produce, what an employee must check, and how you will compare time spent, errors and manual corrections with the current way of working.
If it is difficult to choose between several use cases, I can lead the review with your team, recommend where to begin and build the first prototype.