Seizing the AI opportunity with CSPs

We built the foundation first, with AWS, across all our markets on public cloud alongside our own on-premise cloud, and integrated the data sources into it. Then we trained people, with 10% of all employees gaining a license, for cloud and analytics fluency. And we democratised access, so anyone who needs data gets it there and then without going to someone else in the organisation. On that foundation we built the core use cases. Customer value management, understanding what a customer needs so we can be relevant to a segment of one, with the right product at the right time through the right channel. An internal app store for our analytics products, so the whole organisation can discover and use them. And external monetisation through Crowd Insights, where anonymised, GDPR-compliant location data goes to large enterprises.

Generative AI then goes across all four pillars of the strategy rather than into a corner, and we call the ambition a tenfold productivity improvement. On the customer pillar, we generate marketing material that stays consistent with the brand and relevant to the person, work that people used to do by hand. Our chatbots become genuinely conversational, with the information the customer needs supplied through embedding and a vector database, so updating an offer needs no retraining. On the assisted side, an agent who used to search thousands of knowledge articles now gets the answer directly, and ideally the customer reaches that answer without the agent at all. On the network, anomaly detection with correction and prevention, plus data cleansing and AI ops. On sales, better B2B offers. And across the company, productivity through the Office Copilot and GitHub Copilot.

The architecture sits on the data foundation: sources and cleansing underneath, identity and access management on the user side, the chatbots at the surface, and an embedding engine feeding a vector database and the large language model. We stay open to any foundational model through API access rather than locking into one, which also lets us play with performance.

The enabler that makes Copilot worth anything is easy to miss. Over three years of the transformation we harmonised the entire software repository of the company, every market and every function, into a single GitHub instance. Only because that consolidation happened can a coding assistant see the whole codebase, and every new piece of software goes back into the same place.

Adoption is uneven and worth saying so. The Copilot pilot showed real variance, with younger developers more willing than older ones, which makes this a cultural change and a change management job as much as a technical one. We have about 450 people working on AI and analytics now, and several thousand internal business users. The most striking effect is a kind of crowdsourcing: give people access to the data and the tools, and the creativity that comes back is mind-blowing.