Start with a repeatable task
Look for recurring work that involves finding, reading or preparing information. For example, a service team may repeatedly search approved procedures before drafting a response. Write down the input, the desired output, the owner and the current effort. Avoid starting with a broad ambition such as “automate customer service”; it hides several different tasks and risks.
Check whether the information is usable
Identify the records and documents the task needs. Confirm their accuracy, ownership, freshness and access restrictions. A retrieval-based assistant can reference selected source material, but it still needs evaluation and a process for contradictory or missing evidence. Decide who updates the content and how changes reach the assistant.
Compare AI with a simpler workflow
Use rules or standard integrations when the task is predictable and inputs are structured. AI may be useful where interpreting language, summarising material or drafting a response adds value. A combined workflow is often practical: deterministic checks handle validation and permissions while the model handles a bounded language task. Choose based on the work rather than the novelty of the technology.
Design the failure path
Specify what the system may retrieve, prepare and execute. Identify actions requiring approval, such as changing a customer record or sending a consequential response. Test unsupported questions, irrelevant documents and instructions embedded in retrieved content. A visible escalation route helps people continue the task when the assistant cannot provide an acceptable answer.
Evaluate before extending the scope
Build a reviewed set of representative tasks and difficult cases. Assess correctness, source support, completion, human rework, latency and cost. Compare the pilot with the current process using the same task definition. Review failures before increasing access or adding workflows; a successful demonstration is not evidence that every case will work.
A USEFUL STARTING POINT