Messy data, messier problems, messiest AI

Do AI and Generative AI help in solving the problems of the Healthcare industry? Yes, they do. And those of the financial services, manufacturing, IoT, banking, insurance, retail, e-commerce, education, transportation, logistics, media, entertainment, gaming, government, agriculture, legal, oil, real estate, hospitality, automobiles, pharma, biotech, construction, fashion, apparel, fintech, cybersecurity, genomics, utilities, clinical psychology, airlines, chemicals, human diseases, and telecommunications? Yes. The foundations of AI, ML, and LLMs apply across industries and functions.

They help in solving problems across departments: marketing, sales, HR, supply chain, IT, finance, service assurance, ordering, fulfilment, billing, payments, operations, and research and development.

How do real-world problems look?

Real problems don’t arrive with a tag of regression or classification. Neither do they announce themselves as solved through supervised or unsupervised modelling techniques. Sometimes, they do not even say whether AI will help solve them. At times, they don’t appear to be problems either. Only a deep exploration of data might reveal their identity. Talk about a pained customer or a painful customer. Both are problems worth solving. Talk about allowing a fraudulent transaction or blocking a genuine transaction. Both are problems.

How does real data look?

Messy. Not very clean. Not comfortable working with. It’s the same difference we see between a neatly drawn diagram of the human digestive system in schoolbooks and what you see when you open a body. You realise that the different organs and parts inside the body are not so separably identifiable as you thought they would. You pray you never see another dataset like the one in your current project. Then, you end up getting assigned to a project with messier data. The more you understand the data, the more you have yet to understand.

How does a real AI solution look?

Obviously, not like the ones you solve in tutorials. The iris and Boston house prices datasets, with about 150 and 495 rows, respectively, are nowhere near real-world datasets. Their AI solutions often arrive without much experimentation or failure. In contrast, ask someone who has picked up datasets from about twenty tables, each containing two million rows, replete with junk, missing, and untraceable records.

The best solution is often finalised after seven failed attempts (Why seven? I don’t know). It is never in black and white. Options sway this side or the other. Add to that the fact that we are not perfectionists at our jobs. There are mistakes, misassumptions, and a certain lack of expertise as well. If my colleague were asked to solve the same problem, they could have solved it better than I did. Therefore, an AI solution is anything but standard.

In addition

What separates impactful AI from noise is not sophistication, but honesty about complexity. Too much of the discourse around AI still clings to benchmark scores, model sizes, and neatly packaged demos, while real-world systems remain brittle, messy, and deeply contextual. The hard truth is that most AI projects fail quietly. The integration into business reality is poorly thought through. Building useful AI demands discomfort: sitting with ambiguity, challenging assumptions, and accepting that iteration is the dominant pattern of progress.

In conclusion

AI and Generative AI are not magic solutions. They are multipliers for how well we think. They amplify clarity, but they also amplify confusion when the problem itself is poorly understood. You can reach an incorrect solution pretty fast and deliver it! Across industries, the rigour in defining problems, interrogating data, and aligning outcomes with reality plays a pivotal role. The organisations and practitioners who benefit are not those chasing trends, but those willing to engage deeply with complexity and accept that meaningful solutions are rarely clean, fast, or final.

END OF ARTICLE