Post 1 Introduction: Removing Barriers to Getting Models into Production

How much time, effort, and resources are invested into the development of models that never make it into production?

Our hypothesis is that models do not make it into production because we do not proactively identify the business, organizational and technical barriers that stand between the Proof of Concept (POC) and getting the model into production. More directly, we often settle for just enough support to get to a POC, without sufficient upfront conversations regarding why the model, if the POC is a success, would not then go into production.

Why is this important?

POCs have served to statistically prove out the validity of a model but for a model to have the opportunity to demonstrate business value it needs time learning and operating in the real-world i.e. in production. Particularly with Machine Learning, business value accumulates with time and as the model improves the business value accelerates.

In this blog series we are going to flip the script and instead of defining all the reasons why a model should make it into production we are going to focus on removing barriers clearing the way to production.

Below is a list of examples that in hindsight often explain why a model didn’t make it into production.

· Skepticism and/or misunderstanding of AI/ML e.g. creating a capability that preempts bad things from happening is too mind bending for people to change their reactive behavior.

· The model is statistically accurate, but executives rightfully question if the model reflects the real-world, in other words the training data is not comprehensive.

· Lack of co-pilot ‘one executive can approve a POC but we need at least two to get something into prod’ either IT needs a funding partner or a business leaders needs IT to do the technical work.

· Insufficient business justification to cover the costs to get and keep the model in production — people don’t see a direct line behind the requisite funding and the payback or cost take out.

· Insufficient MLOPs skillsets i.e. IS does not how to operationalize a model

· Conflicting priorities across departments i.e. getting the right people engaged for sufficient period of time

· Inflexible or prohibitive legacy systems causes delays and the POC’s momentum attenuates i.e. steps A, B, and C need to be completed and before long people move on from the POC

In the next post we explore why it is important to give the model time operating and learning in the real-world.

What obstacles have you encountered or overcome getting your models into prod?

— — — — — — — — — — — — — — — — — — — — — — — — — — — — — -

From streamlining supply chains to curing cancer, data is the key to a better future. Slalom partners deeply with our clients to understand their business imperatives and translate data into real-world results. We can help you deliver modern data platforms, beautiful dashboards, and game-changing AI solutions grounded in strategic thinking — unlocking new opportunities to cut costs, accelerate innovation, and win your customers’ hearts.

David leads the Philadelphia Data and Analytics practice, Slalom Data Science Community and Philadelphia Snowflake Meetup and a member of Slalom AI Center of Purpose SteerCo.