Choose your first AI project by looking at the work.

Start with something your team can describe from memory. A request comes in. Someone looks up a record, copies information, asks a colleague, and updates another system. That sequence gives you something concrete to assess.

The useful question is where AI can produce a checked result with less total effort. Count the time people spend reviewing and correcting it.

Follow a piece of work from beginning to end.

Choose a recent, ordinary example. Write down what arrived, where it came from, which information was needed, and how someone knew the work was finished. Include the uncomfortable parts: the missing attachment, conflicting record, or approval that sat unanswered.

For a customer schedule change, the inputs might be an email, a booking record, and availability. The result might be an approved schedule change, a reply to the customer, and matching records. An AI-written reply covers only part of that work.

1. Input “Could we move the installation?”

   A customer asks to move Friday’s installation to Monday. The booking record identifies the job.

2. Preparation A proposed change

   The service coordinator checks Monday’s slot and the crew record. AI prepares matching updates and a customer reply.

3. Decision The coordinator approves

   The coordinator approves Monday’s slot and the reply. If the crew record conflicts, hold every change for review.

4. Checked result Everyone has the same answer

   After approval, verify that the schedule, customer record, and sent reply all show Monday. A failed check stays open for the owner.

Illustrative workflow. The owner compares the full handling time with the current process, including approvals, corrections, and held cases. A faster draft alone does not establish a benefit.

Separate preparation from permission.

AI can prepare a draft, extract fields, or flag a missing item. Updating a consequential record or sending a message still needs the authority you agree to give the service.

Write down who can approve each action. Decide what happens when the input is unclear or a check fails. A process with an explicit review queue is easier to evaluate than one that assumes every case will go smoothly.

Choose work with a repeatable shape.

Look for inputs and rules that recur. If the next customer needs an entirely different application or a new set of decisions, the delivery work changes too.

Supported templates make the boundary visible. They define which sources, records, actions, and exceptions a service can handle. Your assessment should also identify the cases that sit outside those templates.

Measure the full job.

Take a small set of representative examples and record today’s effort and result. When you test the new approach, include review, corrections, failed attempts, and the work needed to resolve exceptions.

Compare the same kind of work at a similar volume. Faster drafting means little if a colleague has to spend the saved time repairing records afterward.

Make the first implementation a specific decision.

A useful recommendation names the work, expected benefit, required systems, review responsibilities, and assumptions to test. It should also say what would make the idea a poor fit. If no implementation is justified, the assessment should explain why.

That gives you a basis for deciding how far to go. Everyday AI can help prepare email, notes, and documents. Connected back-office routines carry work across systems. A company OS brings knowledge, work, and approvals into a shared view while covered routines are operated.

The right starting level depends on the work you need done and the delivery you can verify.