Operationalize AI: You built an AI model, now what?

Five imperatives for adapting DevOps for AI

By | 4 minute read | June 16, 2021

coworkers in conference room

Global AI Adoption Index 2021 reports the top drivers of AI adoption in organizations are: 1. Advances in AI that make it more accessible (46%); 2. Business needs (46%); and 3. Changing business needs due to COVID-19 (44%). To bring AI models into production, businesses are also mitigating the following AI modeling and management issues:

66%      Lack of clarity on provenance of training data

64%      Lack of collaboration across roles involved in AI model development and deployment

63%      Lack of AI policies

63%      Monitoring AI across cloud and AI environments.

Given the acceleration of AI adoption and the need to solve AI implementation challenges, AI engineering is rising to the top of agenda for technology leaders. Software engineering and DevOps leaders can empower developers to become AI experts and play a pivotal role in ModelOps. This blog will discuss the five imperatives in operationalizing AI that can help teams boost their chances for success while addressing common challenges pre-and post-deployment.

Automate and simplify AI lifecycles

Having built DevOps, many software and technology leaders are adept at optimizing the Software Development Lifecycle (SDLC).  More development organizations are expanding the responsibilities of deploying data and AI services as part of the development lifecycle. Advances in automated AI lifecycles can bridge the skills gap, streamline processes across teams and help synchronize cadences between DevOps and ModelOps. By uniting tools, talent, and processes, you can build your DevOps practices to be AI-ready and realize returns as you move through Day-2 operations and beyond.

Implement trustworthy AI

The disruption caused by COVID-19 and other world events this past year may have pushed consumers past a tipping point: an organizational stance for sustainability and social responsibility is no longer one of the considerations but can be a deciding factor for engaging with a brand or not, let alone buy. Misbehaving models and concerns about AI bias and risk are part of the checklist for go or no-go decisions to implement AI. Trustworthy AI drives business transformation with responsible AI solutions that address human needs, safety, and privacy. Explainability, fairness, robustness, transparency, and privacy are the five pillars of trustworthy AI.

Further, the evolving nature of AI-related regulations and varying policy responses make explainable AI implementation one of the top concerns for businesses. IBM Research donates Trusted AI toolkits to the Linux Foundation AI so that developers and data scientists can access toolkits in adversarial robustnessfairness, and explainability. IBM Cloud Pak for Data contains mitigation algorithms for foundations of trustworthy AI.

Support model scalability, resiliency, and governance

As discussed earlier, training data is the number one concern in AI development and deployment as it can have a substantial impact on model performance. Collecting, organizing, and analyzing a sufficient volume of relevant, high quality data to train models under enterprise constraints can be challenging, especially for distributed, heterogeneous environments. Federated learning enables organizations to achieve better model accuracy by securing model training without having to transfer data to a centralized location, minimizing privacy and compliance risks. A data and AI platform with model transparency and auditability as well as model governance with access control and security can seamlessly integrate with DevOps toolchains and framework.

Run any AI model – language, computer vision and other custom AI models

Successful software development teams are also not only integrating off-the-shelf AI services like chatbots but also building custom AI models to drive real business value. For example, development teams can combine deep learning models for speech-to-text, custom machine learning models predicting the next best offers and decision optimization models for workforce scheduling to be deployed with an app for a better customer experience. Beyond packaged machine learning, businesses can now more easily architect a solution that consists of a diverse set of AI models using language, computer vision and other AI techniques aided by industry accelerators.

Get more from your application, AI, and cloud investments

As a development team, you are familiar with the power of innovation in an open, modern environment. By using a modern data and AI platform you can enjoy the flexibility to run your AI-powered applications across various environments—from edge to hybrid clouds—and rapidly move ideas from development to production. Watson Studio on IBM Cloud Pak for Data with Red Hat OpenShift helps you build and deploy AI-powered apps anywhere while taking advantage of one of the richest open source ecosystems with secure, enterprise-grade Kubernetes orchestration. You can start with one use case and build on your success using the same tools and processes. As you take the next steps in the journey to AI, Watson Studio can be a natural fit for building AI in your development and DevOps practices.

Next Steps