The Coming AI Revolution

By H.P Bunaes, CEO and founder of AI Powered Banking.

There is a revolution in AI coming and it’s going to render legacy data and model governance practices obsolete.

The revolution will manifest in three ways:

• Automated machine learning platforms like DataRobot, H2O.ai, Dataiku, and rapidminer are making data scientists more productive. A lot more productive. One company told me that they were seeing 7x as many models from their data science group shortly after the implementation of a leading autoML platform. The increase in model output will quickly reveal bottlenecks in model validation, production implementation, and model operation and management.
• The increasing popularity of tools aimed at “citizen data scientists”, local data literate subject matter experts in the business without formal data science training who nevertheless know a good model and a good use case when they see it, will turn a large percentage of technically savvy business people into model developers. Models developed by citizen data scientists will quickly dwarf the volume of models created by formal data science organizations adding further strain on existing procedures and revealing gaps in governance.
• Availability of nearly unlimited capacity on demand for both data storage and computing power from cloud providers will lead to the proliferation of sophisticated predictive models that can learn from broad swaths of data; structured (your existing databases, for example), semi-structured (your documents), and even unstructured (such as images), sniffing out the data that is relevant to any one particular prediction or population. Demand for more, and different kinds of data for modeling, and the need to integrate model results into downstream dataflows and IT applications, will make data platforms and data flows significantly more complex, harder to manage, and increase points of failure.

What this all adds up to is an explosion in the volume of predictive models and of the data in motion in your organization. Where there were no models, there will suddenly be many. Where there was one model, you may find there are now hundreds. And the pipes providing data into and delivering results out of these models are going to proliferate. Operational and reputational risk from model failure will rise significantly as companies outgrow their existing data and model governance frameworks and legacy procedures.

Making this worse, many banks are starting from a weak position. The demand for more and better models (descriptive and predictive) has already led to a thicket of overlapping, partially inconsistent data flows to a multitude of models. Model outputs themselves have become part of the data flow to downstream data marts, BI, apps and even to other models as inputs. It is the rare organization that knows where all that data is coming from, where it is going, how it is being used, and can identify the potential impacts of changes to data and to the models that consume it.

Certainly there has been much improvement in recent years in data governance at most large organizations. Data quality, data standards, data integration, and data accessibility on robust platforms (increasingly cloud based) have all gotten better. And most organizations now have robust model risk management practices in place, to test and validate models before they go into production use.

But these worlds are about to collide. Data and analytics, once distinct and manageable separately are going to become inextricably intertwined. As brilliantly explained in a paper by several smart people at Google (“The Hidden Technical Debt in Machine Learning System​s”),​ we will rapidly reach the point where “changing anything changes everything.”​

Take a simple example, what differentiates data on a client from a CRM system from data on a client created by a predictive model? The answer: nothing. Yet they are managed today by different groups. The former is typically managed by Data Governance, which is usually led by the Chief Data Officer. The latter is usually the province of Model Risk Management often found in the Corporate Risk Management organization.

But when model outputs become inputs to reports, to business processes, to critical operational or client facing systems, or to other models, they need to be governed just like any other data.

The perfect illustration of this challenge is in change management. Often you will find data change management in the chief data officer’s organization and model change management in the model risk organization. But changes in the data can, and often do, effect models in sometimes unpredictable fashion. And changes to models can change outputs and have major impacts to downstream consumers of those results if they are not prepared for the coming changes.

Managing them separately and distinctly will therefore no longer be sufficient. How to tackle this?

• First and foremost, you must have a complete catalog of all models including metadata describing model inputs and their source and model outputs along with their destination and uses. There are a number of solutions now coming on the market for this purpose including Verta.ai, ModelOp, and Algorithmia.
• Second, data management needs to expand to include not only source data but also all the results (predictions, descriptions) produced by models.
• Third, model management too needs to expand its remit, not just focusing on model testing and validation prior to model implementation but also monitoring model performance and managing model changes after the fact​ ​.
• Fourth there must be formal procedures for keeping model management and data management mutually informed and closely coordinated. Data cannot change without assessing model impact, and models cannot change without assessing data impact.

Organizationally, it may be infeasible to combine legacy organizations across traditional lines of responsibility. And it may be better to leverage existing expertise across model management, data engineering, data management, and IT. But a new partnership model, new tools, and new procedures will be needed.

The explosion in AI is upon us. To use AI safely and effectively you need to get your data and analytics house in order and make sure the right mechanisms are in place to keep it so. Regulators have taken note of the risks of poorly managed AI, and it is only a matter of time before they dictate minimum standards. Combining, or at least tightly coupling, data and model governance is where to start.

 

This is a Sponsored Feature.