How an AI project takes shape: from idea to finished result

A good AI project solves one small problem

An AI project starts with an everyday problem, rather than a model. A good first project has a clear scope, produces a result you can check and can be stopped without causing harm. For example: how can we turn one approved set of instructions into a clearer draft of frequently asked questions?

Try your first small project

1. 1

   Name one problem

   Write the problem in one sentence. Avoid goals such as “build intelligent customer service”; choose something like “draft answers to five common questions” instead.

2. 2

   Choose harmless test material

   Use your own material, public material or invented material. Do not start with personal data about customers, employees or anyone else.

3. 3

   Set the criteria for success

   Decide in advance what you will check: for example, whether the answers are clear, whether they can be found in the source material and whether uncertain questions are passed to a person.

4. 4

   Test it and decide for yourself

   Try several ordinary and difficult questions. Do not publish or automate anything until the person responsible has checked the results and limits.

Two small project exercises

We want to turn our current, approved customer service instructions into a clearer draft of frequently asked questions. Suggest one narrow starting point covering only the five most common questions, and say who will check the draft before it is shown to anyone else. Do not suggest automating all customer service.

Check your own suggestion. Remove questions about prices, contract terms or an individual customer’s information from the list, and tell me how many questions remain.

Create a test checklist for an imaginary FAQ assistant. Include five ordinary questions, two unclear questions and two cases where the assistant must direct the person to the appropriate contact.

Turn a big idea into an experiment you can test

Draft five answers to public, frequently asked questions. Check each answer against the source instructions and have the person responsible approve it before publication. Pass unclear, personal and contract-related questions to a person.

Once you know the basics

The next step is to write short instructions for use and an assessment table covering the use case, approved material, prohibited material, test questions, the person responsible and a feedback channel. This helps an experiment grow into a service that is managed properly, rather than remaining a one-off demo.