THALES IN THE UK
problems in production. Defence and security applications demand higher standards before deployment.
Ajay advocates integrating verification from the start rather than treating assurance as a final stage.“ Even when you’ re innovating, think about what kind of assurance steps you’ re going to put in,” he says.“ If you’ re using data to train a model, then start thinking about where the data is coming from. What is the provenance of it?”
This approach extends through the entire lifecycle. Training procedures require documentation and verification. Deployment demands monitoring systems that catch edge cases. Maintenance processes need to account for model drift and changing operational environments.“ As long as you’ re able to create an agile process which enables pragmatic assurance of these AI systems, you can come up with novel ideas and develop them with assurance at the back of your mind,” he says.
The pace of AI development has accelerated markedly since transformer models emerged and became widely accessible. ChatGPT democratised large language models in a way that changed adoption patterns across industries.“ Now it’ s almost at a logarithmic scale where, even if you are a week late, you’ re already late,” Ajay says.“ It’ s very difficult to keep up with what’ s happening in the world of AI.”
He pushes for what he calls an AI-first mindset, where considering AI tools becomes the default starting point for any task.“ If you haven’ t asked if you can
use AI to do the task for you and automate it, or if AI can help you in reducing your cognitive load to do that task, I think that needs to be the default,” he says.
But rapid adoption creates risks when organisations follow hype rather than solving defined problems.“ If you look on social media, everything is agentic AI, it’ s going to solve all the problems of your world. But what exactly is it going to solve? What is the problem we are trying to solve? Where is the data for it, and how effective is it?”
60 February 2026