Successful AI projects follow a lifecycle that looks more like product development than a one-off analysis.
1. Frame the Problem
Define the decision or task, who benefits, and how success will be measured. Establish a baseline: how is it done today, and how well?
2. Get and Understand the Data
Find the data, check its quality, licences and privacy constraints, and explore it. Many projects stall here because the data doesn't exist, isn't labelled or can't legally be used.
3. Prepare the Data
Clean it, handle missing values, engineer features and create honest train, validation and test splits.
4. Build and Evaluate Models
Start simple, compare against the baseline, and evaluate with metrics that match the business goal. Check performance across relevant groups, not just on average.
5. Deploy
Integrate the model into a product or process. Decide how predictions reach users, how uncertain cases are handled, and who can override the model.
6. Monitor and Maintain
Track prediction quality, input data drift, latency and costs. Retrain when performance degrades. Keep documentation — a model card — up to date.
Where Projects Fail
Most failures come from unclear goals, missing or poor data, no plan for deployment, or no ownership after launch — not from choosing the wrong algorithm.