How to choose your first AI project

Look at where work slows down, needs correction, or leaves customers waiting. Ask the people doing it what they spend time finding, checking, or chasing.

Consider a few possible projects. Compare the value of improving each task with the effort required to connect the tools, prepare the information, and change how people work.

Record how the work happens today. How often does it occur? How long does it take? Where do mistakes appear? How much time goes into checking and correcting it?

Use a representative sample of actual work to establish your starting point.

For the account brief, useful measures would include preparation and review time, missing or incorrect information, and whether the salesperson has what they need for the conversation.

Choose a primary outcome and a few checks for quality and cost. Agree on the target before the pilot begins.

Match the measure to the job. An account brief should improve preparation. A lead-qualification project might also track whether suitable inquiries progress to meetings.

Identify the information the project needs and who is allowed to use it. Confirm that the sources are current, consistent, and accessible.

For the sales example, that could include the inquiry, customer records, approved product information, and company research. Decide which source to trust when two records disagree and how missing information should be handled.

Check the connections, too. Can the system read the necessary records? If it needs to make changes, can those changes be made through a supported connection?

Give it access to the information and actions its job requires. Resolve gaps before committing to a build.

Different tasks call for different approaches.

A project can combine all three.

For the account brief, a workflow with an AI step may be enough. An agent may become useful when the work requires deciding which information to gather next or which action to take.

Choose the simplest approach that can meet the requirements.

Decide which actions can run automatically, which need review, and who receives work the system cannot complete.

In the first version of the sales project, the system might prepare a brief while a salesperson reviews it and decides how to respond. Additional actions can be introduced as the project proves useful.

Test examples from everyday work. Include missing details, conflicting records, and a system that is temporarily unavailable. Check that the system asks for help when it reaches the limits of its information or authority.

Verify the result where the work happens. If an agent updates a record, confirm that the correct record changed. If it hands work to a person, confirm that the person receives the information needed to continue.

Keep these examples and repeat the tests when instructions, sources, or tools change.

Choose a defined set of work, involve the people who will use the result, and set a review date. Agree on what would justify expanding the project, improving it, or stopping.

Include the full effort in your assessment: setup, integrations, software, usage charges, review, corrections, and maintenance.

Describe how any time freed up will be used. For a sales team, that might mean researching more prospects or preparing more thoroughly for important conversations.

Review whether people use the output and whether it improves the job. Gather their feedback alongside the performance measures.

Assign an owner to keep the information current, monitor quality and costs, and address problems. Expand the scope when the results support it.

Before choosing a tool, put the project on one page.

• The job: What specific task should improve, and who will use the result?
• The outcome: What does a completed, correct result look like?
• The information: Which sources and systems does the project need?
• The limits: What can run automatically, and what needs a person?
• The measure: How will you compare quality, time, and cost with today’s work?
• The owner: Who will use, review, and maintain the system?
• The pilot decision: What is the budget, when will you review the results, and what would justify the next step?

Use any unanswered questions to shape discovery before you commit to implementation.