How to choose your first AI project

1. Choose a task you can observe

Look for a recurring activity with a recognizable starting point and result: finding a procedure, classifying a request or preparing a draft. “Improve customer service” is too broad for a first test. “Prepare answers to common questions about returns” gives you a way to define sources, users and review.

Ask the person doing the work where the effort goes and which errors cause rework. A recurring task can be a useful candidate when someone can judge its output.

2. Describe how the work happens today

Gather a representative sample and record working time, delays, corrections and the final result. Include easy and difficult cases. Separate a person’s active working time from time spent waiting between systems.

This baseline lets you compare the pilot with current practice. If an assistant prepares something quickly but requires extensive checking, that review effort belongs in the evaluation too.

3. Check the starting conditions

Identify the information required, its owner and permission to use it. Check whether it is current and whether there is an approved way to access the relevant systems. Decide who will handle exceptions.

A limited sample and draft output may be enough to learn about the task. Automating an action in another system also requires a review of its permissions and consequences.

4. Define what the pilot must demonstrate

Agree the criteria before testing: usefulness, support from source material, total time including review and cost per task. Include cases where the application should acknowledge missing information. The thresholds depend on your operation and the impact of an error.

Keep some examples for the final evaluation. This helps you check behavior on situations beyond those used to adjust the solution.

5. Agree a decision at the end

Decide who will review the results and what evidence would support expanding, changing or stopping the test. A convincing demonstration needs to become a repeatable evaluation. Record the dependencies that would remain before operation, including integration, maintenance, training and support.

A checklist for the first conversation

• One task, its users and a named owner.
• Examples of current work, including difficult cases.
• Available sources and authorization to use them.
• Criteria for quality, total time and cost.
• Review points and an exception path.
• An agreed decision at the end of the pilot.

I can help you assess these conditions through AI strategy and readiness. If you already have a task in mind, discuss your project with me.

To compare approaches, read chatbots, RAG and agents. If your pilot depends on documents, review how to prepare your company knowledge.