5 Pot-holes on your AI/ML Roadtrip:

1. Data roadmap - 500 Pieces - each drawn with 80% accuracy.

Data integration is a $3.1 Billion+ Market, and rightly so! Sound data comes at a great price and one that must be paid in advance. In organizations where IT has evolved organically as a second-class citizen, the chances are that you will encounter traditional data silos and, ironically, silo-ed data-warehouses/marts/lakes. Duplication of data attributes and sometimes the variances in time-sensitive and critical values (have you got two different systems assigning risk ratings to clients using varied parameters?) makes the job of data analytics and provisioning a challenging problem to solve. Trying to start an ML/AI project without having a proper data strategy and a reasonably curated data model in place is akin to jumping head-first from a plane and then figuring out how to open the parachute! 80-85% of your ML delivery cost comes from data integration and analytics efforts. Let's get the map in one piece before we start marching towards the destination. Buying a vendor product in ML space? Unless the use case revolves around standardized data templates, you will pay far more than the vendor licensing costs in the form of data curation, integration, and, more often than not, substandard results.

2. The Bane of Supervised Learning - too much to pack!

If you thought that the data integration was a hard nut to crack, wait till you get onto the feature labeling and engineering street. Most of the data around use cases require manual tagging/labeling, pulling your existing skilled resources away from their day jobs. Invest in good tools to simplify labeling and use automation where applicable! Asking a team of Analysts to tag 11000 emails manually is borderline insane.

3. Model Silos - when each person tells you a different direction.

Got a model for everything? Great! Now you have hundreds of fire-ants in your decision-making process and each making a little decision at a time with its little mind! Like data silos, model silos are becoming an imminent threat. Especially beware of those unexplained black-box models (Explainable AI is a notch better, at least you know what's going on).?The correct path here is to have an AI strategy that in conjunction with a suitable data strategy, strives towards achieving an explainable and streamlined decision-making process. Or, wait for General Purpose AI to become mainstream (Google DeepMind, I am looking at you!).

4. (Fr)agile Transformation and Project Governance - Are we there yet?

While big organizations adapt to Agile Methodologies and their repercussions on project governance, AI/ML projects throw a curveball in the equation. AI/ML projects are more research-driven, and R&D will consume a good chunk of the initial delivery timeline. ML Models don't lend themselves to a vertical-slice-of-cake agile delivery model, and asking "are we there yet?" every sprint will just add to the frustration. Also, it is perfectly normal for the ML model to underperform in production settings (Machine Learning 101).?Wait before you count your success and give data & tech teams the necessary breathing space.

5. Probabilistic-Determinism Paradox - When things go wrong!

Understand the cost of model inaccuracies and put mitigating controls in place. Most of the ML models are baked on the foundations of probability. Expecting a model to make a decision with 100% confidence for every input is unrealistic. A mistake in improving photo quality is less costly than marking a high-risk customer as low-risk. It is important to be cognizant of the respective trade-offs. Before embarking on a new ML journey, start by answering the following questions:

• What happens when (not if) the model goes wrong in production? How is it identified?
• What is the cost of incorrect decisions?
• Does the use case require explanation for decisions taken by the Model?
• What is the process of recovery?
• Is there a scope for preventing error recurrences i.e retraining models or refining processes?

Disclaimer:

All opinions expressed in this article are the author's personal views and do not reflect the employer's collective stance on the matter.