Plan your AI: data, users and risks
Use an AI plan to think through users, data, outputs, risks and human checks, with a crop doctor app as an example.
Before building an AI, good teams plan it carefully. Let's plan one together. Before anyone builds an A I, a good team plans it. Who will use it? What data will it learn from? What could go wrong? Let's plan an example together: a crop doctor app that helps cassava farmers spot disease.
Your AI plan
First, the users. Who will use it? For our app, cassava farmers in Oyo State, many of whom speak Yoruba. Second, the data. What will it learn from? Thousands of leaf photos, labelled healthy, mosaic disease, or brown streak disease. Third, the output. What will it say or do? The disease name, how sure it is, and advice, in English or Yoruba. Fourth, the risks. What could go wrong? And fifth, the people. Who checks it? A farming expert checks the cases where the app is unsure.
Questions to ask about data
Data needs special care. Is it varied? Photos from many farms, in sun and shade, with different phones. Do we have permission to use it? Are the labels right? Experts should check them. And is it private? No faces, names or exact locations.
Risks, and fixes
Now think about risks. What if it says a sick plant is healthy? What if all the photos came from one region? What if many farmers have weak internet? For each risk, plan a fix. Show the confidence and suggest an expert when unsure. Collect photos from many regions. And make the app work offline. The most important question in any A I plan is this. Who could be left out, or harmed? Maybe farmers who can't read English, or farmers with older phones. When you find them, change your plan to protect them.
Quick check
Quick check. All the crop photos came from one farm. What's the risk? The app may not work well on other farms, with different soil, light or phones. Varied data makes fairer A I. Options: The app may not work well on other farms; The app will be too fast; The app will use too many colours; There is no risk.
Recap
Plan the users, data, output, risks and people. Data should be varied, used with permission, correctly labelled, and private. Plan a fix for every risk. And ask who could be left out, or harmed.
Key points
- An AI plan covers the users, the data, the output, the risks and the people who check it.
- Example: a crop doctor app for cassava farmers, trained on thousands of labelled leaf photos, giving the disease name, confidence and advice in English or Yoruba.
- Ask about data: is it varied, do we have permission, are the labels right, and is it private?
- Plan a fix for every risk, and always ask who could be left out or harmed.
Try it: What's the fix?
Choose the best fix for each risk in the crop doctor plan.