Model development was always a challenge in any subject but is more so in an area where abstraction is the foundational attribute: AI. I have written earlier also that abstraction is the first step to building great software and especially in AI, because AI works on abstraction, which depends on mathematics of the highest level. Most people will say data but what they forget is that the model needs data and vast amounts of it, precisely because it has to abstract from it so as to work effectively whatever be the data. The model abstracts from the data so that it can work on any data! This is the crux of any model.
The number of alphabets in English is 26. On this ‘small’ foundation, we are able to build complex pictures and portrayals of almost any subject, we ‘discover’ new subjects such as for example Chaos theory which was born of observations of weather patterns. An AI model cannot function the way language does; it needs a vast universe of data to be trained on so that it can work on ‘new’ data too. There is an important qualification to be understood.
Why vast training data
The fundamental dimension about data as used in AI model development is that the model is being trained not just on data as discrete sets but principally in the interactions amongst the datasets. This is evident in healthcare, especially aetiology, which is the scientific study of causes of diseases. Invariably, the causal factors work together but there are instances where they don’t, making aetiology a contested terrain. There often are many overlapping factors along with a complex interplay of genetic, environmental and life style factors, rendering the diagnosis of an individual patient challenging.
It is no surprise that, in an Agentic AI model, the pressure to seize data increases with every increase in the number of parameters, because the model has to be trained on each parameter and the interactions amongst them. If the amount of data and the level of granularity is inadequate, the model will keep navigating the same limited training data and be completely useless outside – it will fail to predict. Now you can understand the frenzy to collect data at a granular level; whenever you visit any website for the first time, you are asked for your consent.
Data is fundamentally different as it can be anything depending on the subject area; the interactions too will therefore differ with different consequences or effects. A further possibility is that minor variations in data will have much different effects. Or the model will just not work. At all.
Model failures
The most signal failure is one of the most well-documented: IBM Watson Health Platform, a $4 billion failure, built in collaboration with MD Anderson Medical Center. An article titled ‘The $4 Billion AI Failure of IBM Watson for Oncology’, details the trajectory of the failure and points out that the core problem was that it treated medicine as a pure data analytics challenge rather than a contextual human discipline, as it failed to interpret the data in context. There were issues of failed data integration, clinical misalignment because it relied on pre-fed literature rather than dynamic real-world context. It is a lesson in what not to do and how not to go about model-building in healthcare.
Bernard Marr, who writes on management and technology, records five major AI model failures, one of which was in Zillow, a property services specialist. When “Zillow used machine learning to build a tool for automatically buying homes and flipping them for a profit, the results weren’t quite as expected. Its algorithmic model, designed to find optimum buying and selling prices to maximize trading profits, proved incapable of accurately predicting the chaotic behavior of the real estate market. This led to overpayments resulting in $500 million of losses”. (https://www.linkedin.com/pulse/5-big-ai-failures-show-what-can-go-wrong-bernard-marr-5si1e/)
Lack of data?
Surprising though this may seem, lack of data is a major cause of AI failures because the data was not AI-ready. An article by the tech magazine Informatica says that “The global CDO Insights 2025 survey offers insights on specific factors leading to these failures, citing the top obstacles as data quality and readiness (43%), the lack of technical maturity (43%) and the shortage of skills and data literacy (35%)”.
There are many ways to classify data but for our purpose let us say data that is regularly captured by whatever system is used and data specifically collected, outside the system, in any discrete project. It is obvious that each such project must develop its own process of data verification and validation but must also involve the system professionals managing the data captured by the system. If there is a mismatch, the system will encounter problems.
It is not just a question of missing values. Ordinarily, the practice of interpolation used in Statistics ought to work well in instances of missing values, but it is unlikely to work in modelling interactions among elements of data; on the contrary, it will seriously compromise the model. Adequacy or otherwise of data has to be evaluated in relation to the desired output, paying special attention to the interactions among the data elements.
Bias
The reality of algorithmic bias is now no longer questioned by anyone but that does not mean that the problem is resolved. Bias is perennial. Period.
Bias manifests itself in different ways in different problem areas or may fall a prey to incorrect correlations. In criminology and healthcare, this will have direct, serious consequences. One of the obvious biases in healthcare would be characterising some health issues as life-style related without a careful study of environmental factors. Or estimating the probability of recidivism in criminology with a bias against non-whites (as it actually happened in US prisons). Or it could be assumed negligence of workers in determining the cause of defective output, without considering machine-related factors. Such biases enter the system disguised as data. When you combine this observation with the fact that the system is designed to infer hidden correlations you can imagine how deeply flawed the outcome could be.
Developing or even working with an Agentic AI model is not easy; working with it suggests the ability to make adjustments to the model to suit the context, but ensuring that the model is not vitiated by the adjustments. This might be so simple as change management in software.
We can only hope that companies don’t sweep the mistakes and failures under the carpet but use them as grounds to learn and improve.
