How to choose your first AI project

First projects tend to fail by being too ambitious, or by having no way to tell whether they worked. A good first project is a little dull.

Score your candidates

• Frequency: does it happen at least weekly?
• Time: does it take real hours in total?
• Checkable: can someone tell quickly if the output is right?
• Low stakes: is a mistake cheap to fix?
• Owner: is someone keen to use it and willing to help test?

Write down five tasks and give each a yes or no on these questions. Choose the one with the most yeses. Weekly reports, ticket sorting and pulling fields from documents often score well.

Avoid these first

Skip anything customer-facing with no review, anything where the data is scattered or a mess, and anything that needs three departments to agree. They may be worth doing later, once you've had a win and learned how your team works with these tools.

Worked scoring (invented)

• Weekly status report: yes to all five. A strong first project.
• Sorting support tickets: frequent and checkable, with a small risk if misrouted. Good second choice.
• Customer-facing chatbot: frequent, but mistakes are public and hard to check. Not first.
• Forecasting next year's sales: rare and hard to check. Not first.

Before you commit, name the person who will use it and the number you'll measure.

A scoring sheet to copy

Task | Weekly? | Real hours? | Quick to check? | Cheap if wrong? | Keen owner? | Yeses
[task 1] | Y/N | Y/N | Y/N | Y/N | Y/N | __
[task 2] | Y/N | Y/N | Y/N | Y/N | Y/N | __
[task 3] | Y/N | Y/N | Y/N | Y/N | Y/N | __

Winner: ________
Owner: ________
Number we'll measure: ________