AI Summit London: Managing legacy IT and the pace of AI development
While there appears to be no stopping the progress of AI, businesses need to get a handle on technical debt and decide when to build or buy AI
Panel members on the AI as a Competitive Advantage session held at the AI Summit in London this week discussed the reality businesses face when trying to move artificial intelligence projects into production.
Data presented in the Summit’s AI at Scale stream, suggests that 80% of proof-of-concept AI projects fail to move into production. While the panel discussion did not focus heavily on moving beyond AI proof-of-concept projects, as that topic was discussed in a previous session, two members of the panel did raise the issue of how AI aligns with enterprise IT.
This can be a challenge for moving beyond proof of concepts, especially when the new technology being piloted needs to integrate with existing IT infrastructure and enterprise datasets.
Ravi Rabheru, head of AI centre of excellence at Intel for EMEA, noted that the big challenge businesses face is around technical debt.
Dara Sosulski, head of AI and model management for HSBC, added: “The bigger the company, the more the technical debt and the more the complexity.”
It is an area of concern both in large enterprises and in government, which has an agenda to push AI-enablement across the public sector. The Public Account Committee’s (PAC’s) Use of AI in government report from March 2025 noted that AI relies on high-quality data to learn. However, the committee was told by the Department for Science, Innovation and Technology (DSIT) that government data is often of poor quality and locked away in out-of-date legacy IT systems.
Sosulski noted that IT leaders need to assess whether their data infrastructure is right for AI applications to prioritise and understand what is achievable: “Infrastructure is the thing that unlocks the keys of the Kingdom, in a way. You then have something that is a backbone and it’s modular and interoperable. [With such IT infrastructure], you can access applications from other places and you can connect to other things.”
However, she acknowledged that it may not be possible to provide a date as to when all the components needed for AI will be in place in some organisations.
Build or buy?
But the industry is keen to promote the value of ready-made AI capabilities. Sosulski believes that the build-versus-buy question essentially comes down to the use cases the business wants to tackle, saying: “I think all enterprises have now adopted a very similar set of tools to solve problems like software development, drafting emails, translating a document and document Q&A.”
Given that there are products that cater for such use cases, she added: “There are some use cases that are so generalist, we would consider them core capabilities. Those are ones that we consider buying as an enterprise wide solution that is tested and integrates well with our other IT infrastructure. Everybody realises you don’t solve them internally at great expense.”
While business and leaders tackle technical debt and balance when to build and when to buy AI functionality, they also need to keep abreast of the latest developments.
While the whole tech industry appears to be steamrolling agentic AI and artificial general intelligence (AGI), Sosulski recommended that technology decision-makers look at what developments are relevant to the business.
Sosulski felt that there’s less of a need to keep up with the latest AI foundation model. “Despite everything changing constantly, a lot of these models wind up being a much of a muchness for what you want to do,” she said. “We don’t need new foundation models every six months. HSBC and most companies are that way and so, at some point, you just get familiar with the ins and outs of what the models can and can’t do.”
With a selection of some open source and proprietary models, Sosulski urged delegates to focus on assessing which models work best for their use cases. These can then be piloted in proof-of-concept projects to prove they work. She also recommended putting in a place a control framework and IT infrastructure that enables retraining and iterating quickly.
With such a setup, she said: “You can keep moving things out into production without having to overhaul everything every six or 12 months.”
