What is the role of ML model in ROI generation?

Question

I am working on a binary classification problem with 75:25 class portion for 1000 records. Objective is to find out whethet the supplier met the target or not.

Currently my recall for minority class stands at 40%. Meaning, it identifies 22 out of 55 actual negatives correctly.

Now, my business wants go live with this performance and they also say that they will consider this project/model a success if in real time, at least 1 out of 22 (identified by model) as less likley to meet the target (converts to meeting the target upon follow up by our sales team). So, basically if we even get 1000$ from the suppliers (identified as less likley to meet the target) after follow up, we consider this model a success (because it results in revenue for the company. Otherwise, our sales team may not know who to follow up etc).

My concern is as below

a) Even if my recall had been 90%, do you think it is possible to hold model responsible for ROI? Meaning, whether there is increase in revenue or no revenue at all, should model be credited for it?

Because, I feel model can only provide some strategy as to who to follow up, identifying the right suppliers etc. But do you think model can be deemed as success in real life only if it generates revenue? I feel the generation/non generation of revenue could also be due to the skills of sales team, supplier suddem demand etc.

I feel model assessment is over once we assess the performance on test data? Am I wrong to think that way?

But why is model considered a success only based on the ROI? Is it the right way to assess the usefulness of model? Is it possible to seperate the role of model from other factors that influence the ROI?

So, based on your experience, how and when a model is considered a success in real world?

Answer 1

In the end, enterprise ML applications require more than just performance on the test set -- they have to perform in the real world. This performance often is qualitative and binary; the model passes some threshold where non-technical users say:

This model / signal was substantively useful.
My life is better with this model than without it.

That's a (the?) fundamental real test for applied ML. Not just statistical impact, but substantive impact.

In your case, there are a number of reasons the model's impact could be conflated with other factors causing a positive or negative ROI, or at least be swamped by unexplained variance. But, in the end, your job is to positively impact the business. You will achieve success by building a model good enough that, no matter the variance in other factors, the model is obviously having a positive business impact. Success in both a rigorous statistical sense, but also substantively.