The AI Factory
Operationalize the End-to-End AI Lifecycle to Achieve AI at Scale and at Speed
With co-authors John Thomas and Kristen Summers
Just like a physical factory creates physical products, reliably at scale and at speed, an AI Factory delivers AI solutions for business — reliably at scale and at speed. An AI Factory combines data, people, process, product and platform to move beyond science experiments and deliver AI that drives business value. The AI Factory builds on the importance of creating a solid information architecture for sustained AI success. It combines DataOps, ModelOps, and MLOps to accelerate AI innovations to market.
In this post, you will learn about the top inhibitors to success in AI at Scale, and how to overcome those inhibitors by combining data, people, processes, product, and platform in an AI Factory.
Gartner predicted that by 2020, 85% of CIOs will be piloting AI programs (Gartner report G00384355). This represents a massive investment in data science and AI across all industries. However, they also report that over 50% of data science and AI projects never get deployed.
To achieve AI at Scale, it is important to understand the inhibitors to AI success
The top 3 inhibitors in 2018, were (1) trust and security of data and model outcomes, (2) time to implement from POV to production, and (3) integration of siloed data and analytics tools.
As a result, the major AI vendors started a race against time to deliver platforms that are today based on built-in governance, lineage, security, bias detection, and explainable AI — all essential tools for trusted AI. Container-based architectures, multi-cloud build and deploy, no-code dashboards, and no-code deployment reduce the time to production. Persona-based collaborative platforms include a variety of popular tools that speed up co-creation. Data virtualization significantly reduces data integration needs. DataOps, ModelOps, and ML Ops now complement devOps with the joint vision of automating AI at scale. As the new generation of AI platforms started to address the mostly technology-focused inhibitors identified in 2018, the top inhibitors to AI success started to change.
A year later, the top 3 inhibitors reported by business leaders were focused on people and business impact, rather than technology: (1) having the right skills, (2) understanding the benefits and uses of AI, and (3) being able to define the data scope and quality required to achieve business goals.
New skill programs emerged to upskill a new AI-focused workforce, ranging from new university degrees to apprenticeship programs. In parallel, new tools emerged to lower the barrier of entry to AI, most notably AutoAI — capable of automatically building and evaluating AI models. New courses were created to help educate people on the benefits and uses of AI. Design Thinking became a core discipline considered essential in defining data scope and data strategy to achieve business goals, while Data Journalists have gained ground in making the stories behind Data and AI consumable to the business and the public alike.
With today’s plethora of available tools, techniques, methodologies, and personas, it can be difficult to make sense of where to start, how to set up an AI Center-of-Excellence (CoE), and how to operationalize AI at scale, all while improving business KPIs.
The AI Factory approach brings clarity to chaos by focusing on the 5 pillars of data, process, product, platform, and people.
By focusing on AI as a Factory, as opposed to a scientific experiment or POC shop, one can create a prioritized AI roadmap based on current state and future direction, while defining risks and success criteria, just like with a physical factory or business.
Data — the raw materials of the AI Factory.
Just like with a physical factory, you need to understand the supply chain of your data: where the data is sourced from, whether the suppliers are trusted, how the data will be used, which products will be produced from those data, who the consumers of the data are, and whether the consumers themselves can be trusted to use the data in an ethical way. You also need to have efficient storage of your data, being able to deal with varied types of data, whether structured or unstructured, and being able to discover the right data required for your AI processes. Within your data storage, data should be trusted, discoverable, shareable and traceable. DataOps enables agile data collaboration throughout the data lifecycle. Today, it is an essential methodology to manage the data supply chain for your AI Factory.
Process — just like raw materials feed a manufacturing process to produce a product, data feeds an AI process to produce a model or AI service.
In thinking of the AI process or pipelines, a common mistake is to focus on only the technology and neglect the business process. An effective AI process must start by considering the business and market need for the AI innovation. Design Thinking is a great methodology to align, plan, and prioritize with business and technical stakeholders, and to map the business KPIs to data science metrics. Once the business need and metrics are established, it is important to agree on policies and workstreams, to create an agile delivery process and CI/CD pipelines. MLOps is the methodology of choice to manage the processes in your AI Factory, creating the link between the data science and AI experts, and the operations teams. Other Data Science Best Practices include engaging the business stakeholders early on and keeping them engaged, investing in top talent for long-term success, and asking operational questions early in the POC phase of your project to architect your factory.
Product — a physical factory produces products at speed and scale, and an AI Factory produces models, pipelines, and other AI assets at speed and scale.
Just like with a physical business, it is important to understand the market demand for your AI product, whether models, pipelines, or other AI assets. Once again, Design Thinking is a great tool to define the scope and user experience of AI “products”. As with any product, it is important to engage sponsor users and continuously gather end-user feedback. Unlike mainstream products like cars or computers, consumers are not yet at the point where they fully trust AI. It is therefore important to build explainability and bias detection into your AI products, so that people who are not AI experts can understand why AI is recommending certain actions or finding certain outcomes. As with the raw materials of the AI Factory, namely data, the products of the AI factory, such as pipelines, models, and services, need to be discoverable, shareable, reusable, and marketable to achieve long-term success. ModelOps is the methodology that ensures a smooth transition from your POCs to production, as well as monitor your AI “products” for performance and bias.
Platform — the AI Factory floor is the Data and AI Platform.
First and foremost, the AI Factory platform must be a trusted environment, where governance, explainability, and bias detection are built into the platform. Discoverability, sharing, and reuse of AI Factory assets like data, models, and pipelines happen via the AI Factory Catalog, which is analogous to a physical factory warehouse and cataloging system.
Secondly, the AI Factory platform supports all personas who play a role in bringing the AI “products” to market. It therefore includes a variety of popular tools for data engineers, data scientists, visualization experts, analysts, and other factory personas, inside an open ecosystem that facilitates collaboration. It also includes tools that lower the barrier of entry to building AI products, such as AutoAI, no-code dashboards, and click-to-deploy.
Thirdly, the AI Factory platform supports the AI process from POC to production. It provides different environments for sandbox, test, and production, as well as self-serve analytics (such as Industry Accelerators) to speed up the path to production. Leading AI Factory platforms are based on containers and Kubernetes to automatically scale, manage, and deploy AI applications to any cloud.
People — the core to the success of any AI Factory.
An AI Factory requires a team of people with a variety of skills, roles and responsibilities to be successful, just like a physical factory. AI development typically involves cross-functional or full-stack technical teams. That said, it is important to consider the AI Factory not just as a technical shop, but as a market-driven business. In designing the AI Factory, consider all the “jobs-to-be-done”, including the business, AI, and IT stakeholders, data scientists, data engineers, data journalists, IT support, business analysts, marketing, operations, and sales. Assign people with clear ownership, roles and responsibilities to all these jobs. As with any business, the teams who drive the AI Factory need to continuously evolve their collective skills via enablement, certifications, recruiting, and outsourcing, to be the leaders in their market.
The AI Factory comes to life when combining data, people, process, product, and platform in symphony. It is not built overnight, and well worth careful planning and design to ensure a long-term competitive advantage with AI.
