At the AI Summit in London this week, panel members tackled the real hurdles businesses face when trying to push AI projects into production. The numbers are striking—80% of proof-of-concept AI projects never make it past the trial stage. While the panel didn’t delve deeply into this issue, two members touched on how AI needs to fit within existing enterprise IT systems.
Ravi Rabheru from Intel highlighted that technical debt poses a major challenge. Dara Sosulski from HSBC emphasized that larger companies often grapple with even more technical debt and complexity. This challenge resonates not just in big businesses but also in government sectors pushing for AI integration. A recent report from the Public Accounts Committee pointed out that AI relies heavily on high-quality data. Unfortunately, the Department for Science, Innovation and Technology revealed that much of the government’s data is outdated and poorly organized.
Sosulski urged IT leaders to evaluate whether their data infrastructure is equipped for AI. “Infrastructure is key,” she stated. It should serve as a modular backbone that allows for interoperability, enabling access to various applications. Still, she acknowledged that some organizations may struggle to put all necessary components in place within a specific timeline.
The conversation also touched on the decision to build or buy AI solutions. Sosulski mentioned that this choice boils down to specific use cases. Many enterprises have started using similar tools for common tasks like software development and document translation. When use cases are generic, it makes more sense to purchase tested solutions rather than developing them in-house at a high cost.
As companies manage technical debt and navigate the build-versus-buy dilemma, they can’t afford to lose sight of evolving tech. While the buzz around agentic AI and artificial general intelligence grows, Sosulski pointed out that it’s crucial for decision-makers to focus on what truly matters for their business. “There’s less need to chase every new foundation model,” she argued. “Most of them end up being quite similar for practical purposes.”
She recommended that organizations explore both open-source and proprietary models to find the best fit for their needs. After identifying viable options, they should pilot these in proof-of-concept projects. Sosulski also stressed the importance of establishing a control framework and IT infrastructure to support quick retraining and iteration. With the right setup, organizations can continue to roll out new solutions efficiently without necessitating a complete overhaul every few months.
