MLOps: industrialised AI

The promises and challenges of ML

Banks have been relying on data-analytics for a long time, but with Artificial Intelligence (AI) and Machine Learning (ML) technologies they can radically step up their game. “AI and ML are widely seen as key drivers to unleashing a bank’s digital potential,” says Riona Arjoon, manager at Deloitte Consulting, with a focus on data and analytics in the financial services industry. “All the banks are experimenting with AI and ML.”

Yet despite the growth in ML adoption, few organisations manage to realise the broader, transformative benefits of AI and ML. In a survey among nearly 750 business decision-makers, only 8 percent considered their companies’ ML programs to be sophisticated. “A lot of ML projects do not get beyond the proof-of-concept phase,” says Leon Kortekaas, director Cloud Engineering at Deloitte. “They struggle with a lack of expertise and production-ready data, as well as immature development and deployment processes.” AI and ML can transform the way business is done, but only if organisations reshape their operations and structurally embed AI and ML throughout the company, he concludes.

MLOps: DevOps meets Machine Learning

MLOps, which means applying DevOps tools and methods to ML, is the answer to these challenges. “About fifteen years ago, DevOps transformed the way many IT teams delivered applications and services,” explains Arjoon. “By standardising and automating application development, deployment and management, organisations dramatically improved development efficiency, delivery schedules and software quality.” Now, organisations are getting ready to apply DevOps principles to ML. “This may have similar revolutionary effects and will help to realise the transformative benefits of AI and ML,” says Arjoon.

Like DevOps, MLOps features automated development pipelines, processes and tools that streamline ML model development and operations. “It’s an automated sequence to structure your modelling, deployment and management that allows you to get fast feedback,” Kortekaas explains. “It makes the process more transparent and efficient.” Structuring the application of ML models will allow companies to reduce operational costs and scale more quickly, says Kortekaas. He adds that cloud services help to make ML and MLOps easier to use for companies, as they reduce the complexity of having to manage the analytical services and infrastructure yourself. “MLOps creates a highway for ML that everyone can use safely, rather than lots of small dirt tracks.”

Another principle of MLOps is the collaboration between experts in multidisciplinary teams. “Together, data scientists, ML engineers, business analysts and IT-operations professionals design, develop, operate and maintain ML applications in production,” says Arjoon. “Multitalented teams help to create efficiency, scale and business value.”

Detecting fraud, improving efficiency, and new services

TMNL, a collaboration of five Dutch banks, already uses MLOps as standard practice. It deploys AI and ML to monitor payment transactions for signals that could indicate money laundering or the financing of terrorism. “At TMNL, the entire data infrastructure is cloud based, which makes it easy to leverage ML,” says Kortekaas, who currently leads the IT track at TMNL from Deloitte. “Tech, business and IT people work together and own the model end to end.”

Other use cases of deploying ML in the banking industry include models to automatically decide whether a consumer of a company is eligible for a loan or a mortgage. “At one bank, we reduced the response time to loan requests from small and medium business owners from several weeks to a couple of minutes with the use of ML,” tells Kortekaas. Arjoon sees opportunities in banks through creating personalised customer experiences. “Customer service agent augmentation is an example where natural language processing algorithms and sentiment analysis can be applied to better understand customers behavior and offer products and services which better suit their profiles.”

MLOps to promote trust

Whereas the technical and business opportunities of AI and ML are developing rapidly, Kortekaas observes that companies are lagging when it comes to ensuring governance challenges such as accountability, security and ethics. “ML models are becoming easier to use for people without a technical background,” he says. “This opens up amazing opportunities, but also results in a higher risk of data leakage.”

MLOps helps to mitigate these risks and address data management challenges such as accountability and transparency, regulation and compliance, and ethics. “By standardising and automating ML models you can embed ethical, regulatory and cybersecurity requirements in the MLOps pipeline,” says Arjoon. For instance, banks can provide consumers insights into automated decisions. If you deny someone a mortgage based on an AI model’s outcome, you can explain it to a customer. “With MLOps, you can set up your AI and ML models in such a way that every modification is recorded,” says Arjoon. “This helps to make your model auditable.”

The way forward

MLOps enables multitalented teams to work together more efficiently and to get more done in a standardised manner. By creating automated development pipelines, processes, and tools that streamline ML model development and operations, banks can scale ML models and reduce costs. Moreover, MLOps allows AI and ML teams to promote trust by embedding accountability and transparency, regulation and compliance, and ethics.

Last but not least, automating the more mundane tasks with MLOps will give AI and ML professionals more time for exploration and innovation. “MLOps makes the work more fun interesting,” says Kortekaas. “This is an important factor to attract and retain scarce tech and data talent.” In the forthcoming years, MLOps will become a standard practice and an important driver for change in the banking industry, says Arjoon. “Several banks are thinking about MLOps on a strategic level,” she says. “Now it is time to make it a reality.”