Dhruva Kapur

As a part of LLM/NLP models, our businesses have deployed few AI capabilities in terms of level 1 data processing and organising. The ops team used to manually enrich data extracted from the source systems which has now been taken over by AI assistance in terms of organising and enrichment. To build this model, few operation analysts were actively observed, and logics were built accordingly for organising data. For example, the debit entries are summated with a negative sign and in red font whereas the credit entries are summated in green colour. Further the data entries are enriched from golden source viz. Master Data Maintenance/operational data store for items like address, date of birth, real time balances (ledger vs clear/statement). Since the Level 1 is now done by AI in terms of organising and enrichment, the continuous focus has shifted from hiring starters/analysts to hiring techno functional resources, model owners and second line oversight. With testing automation in place using frameworks like cucumber, the role of business analysts have shifted towards that of data analysts/leads who are continuously involved in breaking down data, building controls for key data elements, applying dimensional data controls viz. accuracy, consistency, comprehensiveness, conformity. The data engineers are continuously engaged with quality testers and product owners play more of a business analyst role and requirement gathering. The hiring criteria is including aspects of not only analysis, but also higher numeracy. The personas for hiring mandatorily include technical skills like Java, Selenium, Python along with statistical analysis. The following skills have become necessary 1. Numeracy and Statistical Analysis 2. Capability to breakdown work/functional decomposition 3. Writing short requirements using INVEST criteria 4. Understanding Machines 5. Data Mastery 6. Legal and Regulatory Compliance (to protect from machine doing wrong things) The roles that are becoming redundant include 1. Manual testers 2. Business Analysts (not Data Analysts) 3. Level 1 Ops/analysts 4. Project managers/scrum masters In the given example where AI does data organisation and enrichment, the hiring focus is more on data architects. Data team are continuously involved in two way consistency checks to protect organisation from regulatory penalties, thresholding data dimensions b/w 95%-100%. Reliance is more on technical knowledge such as credit cards follow luhn algorithm so that analysts know where machine can go wrong. Similarly focus is on framing error free algorithms like IBAN. The hiring also focuses on the judgement and oversight meaning a stronger defence system which continuously monitors the models and work as model owners. Assuming the role of a Data Analyst, the hiring conditions will now be:- Critical Analysis and Judgment: More focus is on pattern recognition. Problem Framing: Doing right first time, well it is believed that intelligence is required even if one must cheat. Therefore, human intelligence should be the core. Explainability: One of the critical criteria is model explainability like shaps. Understanding models in a more mature way will be required. Statistical Analysis: Data analysts are required to do result oriented statistical and mathematical analysis to study models. Old Criteria Raw Data analytical speed: Models can handle this effectively without human intervention Manual Reporting: The Data Analysts do not have to do any manual reporting anymore as models are good enough to handle those. Data Organisation and Modelling – Analysts are no longer required to do data cleaning, organising and modelling which AI can handle well.

There can definitely be instances where AI predictions can be ignored or on the flip side be flawed. it all depends on the law of the confusion matrix and for example sake I would like to discuss the two dimensions of the confusion matrix. False Positive - Model or agent predicts positive, however it is not true. False Negative - Model or agent predicts negative, however is true or positive. while the item No 1 is classed as type 1 error in hypothesis testing, type 2 error could be seen in item no 2 above, a beta risk. it means False Positive do less harm but in principle, False Negative can be really harmful. in our organisation, there is a team who monitors models or is otherwise called ML ops, where the model owner continuously studies these processes by means of SHAP values. The model owner also studies something as recall rate and the moment there is a decline to the recall rate, the human intervention is triggered or human in the loop. Usually recall rate is set high as a number so that model predictions could be studied properly. recall rate also mean the probability of model predicting the correct value or desired result in percentage or otherwise Recall rate is True Positive/(True Positives+False Negatives). This whole process is called Model Monitoring. there are other principles like r squared, thresholds which are also important in a model. monitoring process.

The models or AI are built via human intervention but have intelligent brain and decision making powers. The AI’s recommendation is followed, and things go wrong? The AI’s recommendation is ignored, and things go wrong? In my opinion for both he above scenarios, there has to be a logic behindz AI means artifical human and should have the same functions as of RACI Matrix- Responsibility., Accountability, Consulted and Informed. Removing Accontability, i think Ai can be held responsible, can be consulted or informed for further course of action. AI must not be help accountable as it is not considered a resource with high influence or power. This would further mean that if AI is made accountable, then the day is not far that AI will control humans too.

AI that we use in my organisation mainly provides a recall rate which the possibility of something being true had a human carried the same exercise. Model Monitoring is a critical process and it helps understand whether AI is working as expected. The models are trained over a period of time and baseline data. if the data sets are higher, the model prediction in terms of a recall rate will be more than 99%. Also the other metric we use is the confidence level of 95%, meaning model is allowed to make 5% errors. now these errors are man tally investigated and the model is retrained basis the results of these failures. this is how model is groomed on a day to day basis and therefore to summarise:- capture metrics in tableau/dashboard read recall rate and confidence levels tune recall rate at 99% and confidence at 95% if confidence falls below expected , then manually investigate the errors and retrain the model on the manual investigation outcomes this is how model will be able to get a new baseline

In my opinion there are processes where there is drift in decision making. The larger the data set is, baseline changes or drifts from point A to Point B and sometimes skewed. I had a process AI model to detect false positives and we monitored it through tableau. It worked on 95pc confidence and 90pc recall rate i.e. possibility pf outcome being true. One day the file or the batch job failed which had around circa 1m records and model operated as usual and that it dis not met the percent of recall rate and also the confidence level. This was the biggest file which got missed and was specific to UK and all other rest of the world records were low in number. In order to manage this, a self check algorithm was developed and ensured that model runs only after the files are received and was governed by model Monitorig process

Are theories like Emperical Risk Minimization, Data agnosticism and CI/CD helpful in this context? I work for Fin Crime Domain and I observe most automation fails due to data agnostic capabilities whether it is robotics, AI/ML. Data Retraining and continuously feeding back in the machines are helpful techniques though.