Why Discovery is going slow on AI coding agents - Derek Wilcocks — Discovery’s group CIO, Derek Wilcocks

Discovery group CIO Derek Wilcocks has pushed back on the idea that AI is about to make software developers obsolete, saying the productivity gains his teams are measuring from AI coding tools are real but a fraction of what the industry’s loudest voices are promising – so far.

Speaking on an upcoming episode of TechCentral’s Meet the CIO podcast, Wilcocks said Discovery is seeing efficiency gains of roughly 20-25% in some areas from AI coding assistants. He immediately qualified the figure: developers in large corporates were typically spending only 25-30% of their time writing code in the first place, with the rest going to meetings, code reviews, testing and sessions with business users.

Discovery builds much of its own software from scratch in its biggest businesses and buys little by way of traditional enterprise software at all in health and life, Wilcocks said.

Three problems stand in the way of fully agentic AI coding, according to Wilcocks: explainability, efficiency and maintainability of AI-generated code. None is solved today. Anyone who has run a significant project on AI-generated code knows that a year later, after model drift, they are left with something even an AI is not quite sure the previous AI wrote, he said.

That is more dangerous than it sounds, because code is not merely a set of instructions that execute something. It is also the systemisation of a business’s processes. Wilcocks expects meaningful human oversight to remain necessary for the foreseeable future.

Discovery’s senior developers use all the major agentic coding frameworks, but the group has spent heavily on constraining the tools. The effort has gone into limiting what the agents are allowed to do, so that output meets Discovery’s coding standards, its security and privacy requirements, and works with what is already in place.

The scaffolding

The calculus is different for a start-up, Wilcocks conceded: a company founded today would have to have its “head in the sand” not to generate its code almost entirely with AI. But in large regulated enterprises, writing code that touches multiple legacy systems, across multiple teams, he said Discovery remains deliberately conservative – a pattern he is hearing from other CIOs, too.

His broader bet is that the winners will not be the companies with access to the best frontier model. They will be the ones that build what is increasingly called the harness: the scaffolding that lets AI be deployed safely and at scale. The companies investing in that now will be the AI winners of the future, he said.

Read: Discovery teams up with Google to launch AI-powered Vitality platform

That view feeds into how Discovery hires. Wilcocks said the industry had come to the end of a decade in which the ability to code was enough on its own to walk into a job.

Coding still matters, he said, but so does the grounding a computer science degree or its equivalent provides: how architectures work, how to think about algorithms, how design-led approaches work and how to handle risk and governance. Add the ability to work in a team and communicate, and the most important remaining skill is simply to keep learning. Building a career on a short programming course will be harder over the coming decade than it was over the last, he said.

Discovery Health — Discovery’s head office in Sandton. Image: Boogertman and Partners

Discovery’s annual grad hack, where teams from universities and coding academies compete over a weekend, has shown him how far the tooling has moved. Teams now ship a complete product in two days, something out of reach eight or 10 years ago.

Retraining the existing workforce has surprised him more. Discovery made well-known online academies freely available to its IT staff, and Wilcocks said he had expected 3-4% to take them up. Instead, 50-60% have completed at least one AWS or Google programme. Hiring an outright specialist in agentic security or advanced AI models is close to impossible today, he said. Bringing in graduates who can learn it quickly is not.