Accelerated by a widespread move to digitize operations, the company is enthusiastically embracing AI. According to According to IDC’s AI InfrastructureView 2022 survey, 31% of enterprises say they now have AI in production, while the majority are actively piloting AI technologies. Increasingly, the adoption of AI is leading to increased profitability, with 27% of companies responding to a December 2021 McKinsey investigation claiming that at least 5% of their earnings before interest and tax (EBIT) are now attributable to AI.

But there remain many obstacles to the successful implementation of AI. Of the companies participating in the AI ​​InfrastructureView survey, only a third say they have reached a “mature” state of adoption where their entire organization benefits from an enterprise-wide AI strategy. Additionally, while nearly two-thirds of companies surveyed in the McKinsey survey say they will continue to increase their investment in AI over the next three years, half admitted to incurring project costs of AI higher than expected.

Disconnect from data science

Why is it so hard to get AI projects into production? The reasons vary, according to Jeff Boudier, head of product and growth at Hugging Face, the AI ​​language startup. But typically, companies fail to establish systems that would allow their data science teams — the teams responsible for deploying AI technologies — to properly version and share models, code, and datasets. ‘IA, he said. This creates more work for AI project managers, who need to keep track of all the models and datasets created by teams so they don’t reinvent the wheel for every business request.

“Today, data science is largely done in ‘single player’ mode, where code lives in notebooks on local machines,” Boudier told VentureBeat via email. “That’s how enterprise software was designed 15 years ago, before modern version control systems and…collaborative workflows changed the game. »

The emerging discipline of MLOps, which stands for “Machine Learning Operations” (a term coined by Gartner in 2017), aims to address the disparate and siled nature of AI development by establishing collaborative practices among scientists across data. By simplifying AI management processes, the goal of MLOps is to automate the deployment of AI models into an organization’s core software systems.

For example, startups like ZenML allow data scientists to express their workflows as pipelines which, with configuration changes, can accommodate different frameworks and development tools. These can be integrated into a framework to address reproducibility and release management issues, reducing the need for coordination between DevOps teams and data scientists.

Increasing size and data requirements

But collaboration isn’t the only hurdle facing companies adopting AI. Others are the consequences of the exponential growth of machine learning models, according to Boudier. Large models are often not suitable for commodity hardware and can be slow and expensive to run. Or they are locked into proprietary APIs and services and dubiously presented as universal problem solvers.

“[Proprietary models hamper] Adopting AI because…teams can’t dive into code and properly evaluate or improve models, and continues to confuse how to approach AI issues pragmatically,” said boudier. “Deploying large models in production to apply to large amounts of data requires diving into the graph from model to hardware, which requires skills that most companies don’t have.”

Sean Hughes, director of ecosystem at ServiceNow, explains that companies often expect too much from AI models without doing the work necessary to adapt them to their business. But this can lead to other problems, including a lack of data available to refine the models in each context where they will be used. In a 2019 Dun & Bradstreet investigationcompanies rated the lack of data on par with the lack of internal expertise as the top barriers to further implementing AI in their organizations.

“The hype and sensationalism generated when open source work by AI researchers that achieves new cutting-edge benchmark results can be misinterpreted by the general public as the same as ‘problem solved’. But the reality is that the state of the art for a specific AI solution can only achieve 78% accuracy for a well-defined and controlled setup,” Hughes told VentureBeat via email.[A major challenge is] the business user’s expectation that [an off-the-shelf]model will understand the nuances of the business environment in order to be useful in decision making… [Without the required data,] even with the ability for the AI ​​to suggest a best directionally correct next action, it cannot because it does not understand the context of the user’s intent in this business. »

On the same page

Feiyu Xu, SVP and Global Head of AI at SAP, agrees, adding that AI projects have the best chance of success when there is alignment between business lines and technology teams. ‘IA. This alignment can foster “targeted” and “scalable” solutions for delivering AI services, she argues, and address ethical issues that might arise during ideation, development or deployment.

“The best use cases for AI-powered applications ensure that AI technologies are fully integrated and automated for end users. Additionally, AI systems work best when experts safely use real business data to train, test, and deploy AI services,” Xu said. “Companies need to clearly define guidelines and safeguards to ensure that ethical issues are carefully considered in the development of new AI services from the outset. In addition, it is important to include external and independent experts to regularly review the cases and topics in question.

On data challenges in deploying AI, Xu highlights the emergence of platform-as-a-service solutions designed to help developers and non-developers connect data sources across different backend systems. . Torch.AI, for example, connects applications, systems, services, and databases to enable reconciliation and processing of unstructured and structured data for AI applications.

“AI plays a key role in enabling businesses and industries to become intelligent enterprises,” Xu said. “Most AI users have little experience developing software to design, modify and improve their own workflows and business applications. This is where an intuitive, no-code development environment for functions such as intelligent process automation, workflow management and robotic process automation can really help.”