Find work that repeats
A useful first scenario is a recurring task involving text, documents or enquiries: collecting initial details, drafting a response or finding information in an approved knowledge base.
Describe the input, the expected output and who will use it. “We need AI for the business” gives you no quality criterion. “Draft a response using approved terms” creates a task you can discuss and test.
Decide where answers come from
The assistant needs current reference material: service descriptions, operating rules and relevant documents. If employees use conflicting versions of the terms, AI will not resolve that inconsistency automatically.
Assign an information owner and an update process. Decide which data the task actually requires. Connecting every document and conversation simply because it is technically possible is not a useful starting point.
Separate answers from actions
Drafting a message, sending it to a customer and changing a CRM record carry different responsibilities. Decide which actions can happen automatically, which require approval and which remain with a person.
Pay particular attention to unusual terms, disputes and actions with external consequences. Limits should be built into the system's design rather than expressed only as a request in a long model instruction.
Design the handoff to a person
The assistant should stop when a reliable answer is unavailable, key information is missing or someone asks for an employee. Decide who accepts that work and what context they receive.
A handoff does not help if a notification disappears into an unowned group chat. The event must be visible, the owner clear, and the context sufficient to continue without collecting all the details again.
Test before expanding
Prepare anonymised examples of ordinary and difficult situations. Check behaviour when information is missing, conditions conflict or the request falls outside the task. A fluent answer alone does not establish that the workflow works.
Ready for a pilot?
- The task has a clear input and expected output.
- Knowledge sources are approved and have an owner.
- Drafting, sending and changing data are separate permissions.
- Stopping and human handoff rules are defined.
- The team can identify and correct mistakes.
Start with a limited workflow whose behaviour can be observed and reviewed. Expand after assessing real operation, not just a successful demonstration. Some processes may not need AI at all.