What applied AI is and how to pick a problem

What it isCopy Anchor Link

Research AI builds new models and methods. Applied AI takes a model, connects it to data and tools, tests it and fits it into a job someone already does. Most of the AI work in the world is applied.

Most applied problems fall into a few kinds of task. Predict a number (next month's sales). Sort things into groups (fraud or not fraud). Rank things (which video to show next). Read or write language. Understand images or sound. Choose actions in the physical world (a robot arm). Each kind has its own tools, covered in the lessons that follow.

Picking a problemCopy Anchor Link

Ask four questions. How often does the task happen? A task done once a year is not worth automating. What does a mistake cost? A wrong movie suggestion costs nothing, a wrong drug dose costs a lot. Do you have data, or can you get it? Can you tell if the answer was right? If you cannot measure success, you cannot improve it.

A worked example: start with a baselineCopy Anchor Link

A shop wants to predict which customers will cancel. Before any model, write the simplest rule: "customers who have not logged in for 30 days will cancel". Say it catches 40 of the 100 who cancel. That rule is your baseline. A model is only worth its cost if it beats the baseline by enough to matter. Many projects find that the rule gets most of the value.

Scoring a task before you buildCopy Anchor Link

Count the hours first. Frequency times minutes per case gives the yearly or monthly load. Tagging 300 support tickets a month at 2 minutes each is 10 hours. Reviewing 40 contracts a month at 90 minutes each is 60 hours. The second is the bigger prize, but a mistake there costs more and needs a lawyer to check. Neither wins on size alone.

1. Write the task as input and output.
2. Multiply frequency by time per case.
3. Name the cost of a wrong output.
4. Find 50 past cases with known right answers.

If step 4 is impossible, you cannot tell whether a model works, so stop there.

A second example: when a rule is enoughCopy Anchor Link

A store wants to flag refunds over $500 for a manager. That is a threshold, not a prediction. One line of code does it, never tires and can be explained to an auditor. A model adds cost and uncertainty with nothing to gain. Use learned models when the pattern is too fuzzy to write down, such as spotting a likely cancellation from dozens of weak signals.

What goes wrong: a good score nobody can useCopy Anchor Link

A shop's cancellation model is better than the 30 day rule. It catches 55 of the 100 who cancel, against 40. But the retention team can phone only 50 customers a month. What matters is how many real cancellations sit in the 50 names the model ranks highest, not the overall score. To confirm, rank by the model's score, take the top 50, and count the real cancellations among them. Test against the capacity you actually have.

Common mistakeCopy Anchor Link

Starting with "we should use AI" instead of "this task costs us 200 hours a month". A tool searching for a problem usually finds a demo, not a result.