Skip to main content Skip to footer


DevOps for AI: Are you ready to scale?


April 22, 2021

What does it take to bring powerful change? Speed + quality + scale + flexibility. Just as DevOps is bringing this to software development, you can use DevOps to power up your AI model delivery.

The next wave of competitive advantage is expected to come from technology such as artificial intelligence (AI), emerging technologies like blockchain and the imperative for every company to move to the cloud. Overnight, we’ve the seen speed of the shift to these digital technologies go from years to months or even days. And to support this change, the number of intelligent automation and AI implementations are increasing fast. To integrate AI into your company’s DNA, DevOps principles for AI are essential.

How DevOps for AI can help

When applied to AI, DevOps enables AI at scale through the operationalization of machine learning models from design to production. DevOps for AI ensures that the right AI delivery processes are in place and can bring the flexibility and “fail fast” approach needed during these times of constant change and technological transformation. Essentially, DevOps will facilitate continuous delivery, deployment and the monitoring of models through:

  • Speed: Improve time-to-market by reducing non value-added activities in AI delivery
  • Quality: Accelerate cleaning datasets and promote continuous learning to improve AI model quality
  • Scalability: Preempt scalability considerations and ensure AI models can scale on-demand
  • Stability: Monitor deployed AI models to keep them reliable, stable and accurate

How to make it repeatable

To meet the demand for AI implementations, best practices need to be applied to AI model operationalization. However, the best practices around AI delivery are in a state of constant flux. DevOps principles address this challenge and present a repeatable, yet adaptable, approach to increase the maturity of your AI delivery through ongoing change.

We’ve broken down AI operationalization using DevOps practices into four stages:

  • Data Preparation: Preparing the right datasets for developing AI models is a crucial starting step as model accuracy depends on the quality and the size of the training dataset. Traditionally, data preparation—data extraction, data cleansing, data labeling and data validation—is a manual and cumbersome task where data scientists typically spend around 70% of their time. DevOps for AI automates such steps and enables data pipelines to handle big data. This improves the quality and the size of datasets and frees up data scientists to focus on feature engineering and AI model development.
  • AI Model Development: Even with the right data, developing an AI model takes a lot of time. AI model development comprises three main activities—feature engineering, algorithm selection and dataset training. Model development is an iterative process that requires multiple rounds of model training to arrive at an optimal solution. Usually this happens in the local machines of data scientists and without much collaboration between various AI teams. DevOps practices speed up AI model development by providing the elastic infrastructure and the processes for parallel development, parallel testing and model versioning. This reduces the time and effort required to arrive at the optimal model.
  • AI Model Deployment: Deploying an AI model in production is an extremely challenging area for many organizations. Problems especially arise when individual data scientists deploy a model developed in silos in their local workstations. For AI models to perform well in production, they should be able to handle incoming streams of data in real-time on highly scalable and distributed platforms. DevOps methods make AI models portable and modular. Such architecture allows AI at scale in operationalizing AI.
  • AI Model Monitoring & Continuous Learning: Once deployed, models face the threat of ‘model drift’. This is where the models deployed in production were developed based on historical datasets initially. Then as time passes, the data and the model become outdated and the model accuracy decreases resulting in ‘the drift’. DevOps for AI brings in the concept of continuous learning, where data and model performance indicators such as drift and accuracy, are monitored to ensure they stay relevant for a longer time. This results in better and more responsible AI solutions in the market.
Four parts to DevOps for AI: Data Preparation, AI Model Development, AI Model Monitoring & Continuous Learning and AI Model Deployment

DevOps for AI presents a promising solution for organizations looking to accelerate and improve AI solutions, AI-driven innovation and intelligent automation. It speeds up data preparation and model development tasks and brings in standardized processes to make AI at scale a reality. But despite the clear benefits, AI operationalization is often unaddressed. The time is now to make AI operationalization a core business objective.

Reach out to learn more about how Accenture can help you assess and execute your end-to-end AI life cycle management, including our intelligent automation platform Accenture myWizard® with myWizard for DevXOps to automate DevXOps across technologies and platforms.


Rajendra Prasad (RP)

Chief Information and Asset Engineering Officer