When a team starts exploring AI, the list of possible uses can quickly become larger than the capacity to assess them. Start smaller. Look for one recurring piece of work where people spend time gathering, sorting or reshaping information—and where a person remains responsible for the final decision.
A bounded workflow is easier to understand than a broad promise to “use AI across the firm”. It gives colleagues something concrete to test, makes the information flow visible, and helps you notice when a model’s answer is not good enough.
Map the task as it happens today
Choose a real process and write down its steps from beginning to end. Who starts it? What information do they receive? Which systems or documents do they consult? What do they produce, who checks it and what happens next? Include the exceptions and interruptions, not just the ideal path.
This simple map often reveals that the time-consuming part is not the apparent task. A “draft a reply” request may actually involve locating the latest policy, checking a client history, applying a tone and routing the message for approval. AI might help with one part and be inappropriate for another.
Use a shortlist of practical tests
- Is the task repeated? A recurring task gives you chances to learn and compare, without assuming that every case is the same.
- Can you bound the inputs? Identify the exact files, fields or approved references the system should use. Avoid opening a whole shared drive by default.
- Can a person review the output? Decide who checks it, what they should verify and whether the draft is clearly marked as unapproved.
- What happens when it is wrong? Consider the consequences of an omission, invented detail or misrouted answer. Begin where errors can be caught before they affect a client or a decision.
- Can you describe “better”? Agree how you will judge usefulness: fewer repeated steps, a clearer first draft, or easier retrieval of source material—not a vague sense that the tool is impressive.
Make the human check part of the workflow
A reviewer needs more than a final answer to approve. Show which sources informed it, keep the originals available and give people a way to flag missing or conflicting evidence. Define what must never be generated or sent automatically.
Use appropriate test material and limit access to the people involved. If real client information is necessary, involve the firm’s privacy, security and professional-risk owners first. A local system does not remove that review.
Test against ordinary work, not a perfect example
Collect representative cases, including awkward documents and examples that should trigger a cautious answer. Have experienced staff compare AI-assisted work with the current process, recording failure modes as well as useful results. Do not judge only polished responses.
Ask whether the proposed step fits the working day. A system that adds queues, duplicate entry or difficult corrections may not be useful, even if its output looks convincing.
Decide what comes next
After a trial, choose deliberately: refine the task, widen it with new safeguards, keep it optional, or stop. Document what information it can use, what it may do and who owns review. A thoughtful “not yet” is a productive outcome.
Starting with one well-understood task helps a firm learn about its data, infrastructure and working practices before making a larger commitment. It also keeps the central question in view: does this help people do useful work while keeping appropriate control?