Start with the smallest useful AI project

Choose one job, not a department

Follow one piece of work from beginning to end: answering a common enquiry, preparing a proposal, or finding a policy before a call. Look for copying, repeated questions, and decisions waiting in an inbox.

A useful first project names one owner, one starting point, and one outcome. “Help the front desk prepare a reply to booking questions” is easier to test than “improve customer service with AI.”

• One repeatable job with a clear beginning and end
• A person who knows when the result is useful
• An outcome the team can inspect without a new dashboard

Start from sources your team already trusts

Give the system a small, maintained set of material: approved service notes, a current price list, or an internal FAQ. This makes the first answer easier to check and shows where information is missing or out of date.

For example, a small hotel could let staff search its own room, transfer, and seasonal policy notes before drafting a guest reply. The system should point back to those notes instead of inventing a policy.

Design the fallback and handoff first

A first project needs an obvious way to say “I am not sure.” Let the system return a draft, a source link, or a short exception list when it cannot answer from the approved material.

Decide who receives the exception and what context travels with it. A handoff that includes the original question and the source gap is far more useful than a generic warning.

Pilot with a small, real slice of work

Use a limited set of real requests with the people who already do the work. Keep the old method available while the team compares the result. The point is not to prove that AI is always right; it is to learn where it helps and where the workflow needs a person.

A pilot can be as simple as reviewing drafts for one common type of enquiry during a normal workweek. Avoid connecting it to every channel or record system before the team has seen the output.

Measure a useful change

Choose a few observations the team can collect without ceremony: how often a draft needed a major rewrite, where the source material was incomplete, and whether a recurring task became easier to finish.

These are learning signals, not a promise of a universal result. They help decide whether to improve the source material, adjust the workflow, or stop the experiment.

Earn the next step

When the pilot is useful, expand one dimension at a time: another question type, another trusted source, or a clearer review step. Keep the same owner and the same visible feedback loop.

The goal is not to remove people from the process. It is to return attention to the parts of work that need judgement, empathy, and taste.