September 20, 2019 - 786 views
Businesses are scaling AI beyond the pilot stage. here's how.
Scaling AI beyond the pilot phase is no simple endeavor. Businesses that are overcoming the barriers are getting these three things right.
2019 has been a good year to be in the business of applying AI to solve business problems. Many organizations now realize AI’s power and are gearing up to infuse machine intelligence into their business core. This has personally been very cathartic, as it’s something I’ve been advocating for many years.
But why now? I see a number of common trends culminating to form an AI inflection point:
- Multiple business units across organizations have tested AI against a selection of use cases and started to prove the benefits.
- Organizations continue to explore cost take-out in order to become more competitive. AI has emerged as a key to unlocking efficiencies across the front, middle and back offices.
- Digital efforts need to deliver value from all the data they generate, and AI is becoming the logical way to do this.
Plenty of AI Hurdles
Despite this, very few organizations find AI to be a simple endeavor. Major barriers often prevent them from scaling AI beyond pilots. For example:
- Their infrastructure isn’t mature enough to cope with the demands of both developing and deploying machine learning at scale.
- Questions abound on where AI should sit in the organization. Is it IT? The business? A federated or centralized model?
- Time to value – from AI project inception to live production – is still too high to contemplate moving forward with tens if not hundreds of use cases.
- ROI is still up for debate, a common occurrence with not only AI but also most advanced technology initiatives.
- Resource constraints are pervasive, from data engineering and data scientists, to SMEs who understand existing business processes and the role AI can play in advancing strategic objectives.
Three Foundational Pieces for Scaling AI
Luckily, there’s a way through these barriers. From my experience, you need to get three things right to successfully scale AI pilots:
- Rethink the AI operating model to deliver proper support, strategic focus and governance. This requires senior buy-in, agreement on where AI should sit in the organization, governance and new processes. For a large insurer, this meant adopting a hub-and-spoke model, with an AI center of excellence sitting in the business and reporting into the COO, as well as small teams sitting in ”spokes,” such as marketing and customer service.
- Focus on data management, analytics and algorithms. Data and analytics architectures at many organizations are built for business intelligence purposes. This means that some data may not be captured at a sufficient granularity or history to enable machine-learning development. AI places new demands on how data is managed; feature stores, or data stores, specifically built for machine-learning data, are fast becoming the de facto way to do this.
AI also places new demands on how compute capacity is managed; Evolutionary AI, or machine-learning algorithms that learn as they go, can improve existing models given the right compute resources. For a large bank, the combination of a feature store built for machine-learning development and a model deployment infrastructure reduced the time to deliver a model into production to two weeks.
- Embrace new ways of working, using both accelerator frameworks and AI prototype factories to prioritize the vast number of potential use cases. Moreover, agile digital engineering pods are required to bring together data, science and business into one team to push through to delivery. At a global media company, an AI prototype factory has helped the business assess tens of AI use cases in rapid succession. The assessment includes ROI estimation, deployment considerations and recommended process changes.
Tapping the Value of Fully Scaled AI
It’s clear that AI will transform how we run our organizations, and there is massive value in successfully scaling AI. It’s time now to take a step back and work out whether your organization is built for success. If not, it’s back to basic blocking and tackling because no amount of automated machine learning is going to solve the structural issues that must be addressed before undertaking an initiative as complex – and potentially rewarding to the business – as AI.