The shift from experimental artificial intelligence (AI) projects to enterprise-grade autonomous systems was the defining theme of the Google Cloud Summit in London last week.
As UK enterprises move to action-oriented agentic workflows, technology leaders encounter a new set of challenges. To build a prototype agent is simple, but to deploy hundreds of them at scale demands rigorous data infrastructure, strict security governance and fine-grained cost management.
This transition is also redefining the economics of IT, with “tokenomics” – the tracking and optimisation of large language model (LLM) token consumption – emerging as a major focus for chief financial officers (CFOs) and technology directors.
We sat in on a Google-hosted roundtable that featured Jo Drake, chief technology officer of platforms at THG Ingenuity; Mohsin Ghazpour, chief AI officer at Kingfisher; Steve Pimblett, chief data officer at Rightmove; and Hayley McKelvey, chief AI officer at Deloitte.
The panel discussed business outcomes in their initial agent roll-outs, the practical methodologies they use to track and manage token budgets, and the cultural shifts required to prepare engineering teams and wider workforces for an autonomous enterprise.
From chatbots to autonomous agent deployments
Q: We are seeing a market transition from basic AI experimentation to concrete, real-world agent deployments. How are proactive agentic models transforming user experiences and business outcomes in your respective sectors?
Jo Drake: The benefits we have seen are threefold. First is the customer experience. Shopping is conversational by definition – you walk into a store, someone greets you, and you explain what you are looking for, whether it is for a wedding or an athletic event. We have brought that conversational experience to e-commerce websites and apps using our shopping assistant. We can layer hyper-personalisation onto that, knowing the customer’s size, colour preferences and buying goals.
The second benefit is performance. For MyProtein, where we piloted this, we saw a massive increase in the conversion rate for first-time visitors, including a 22% increase in basket size and a 20% increase in average order value.
The third benefit is the data loop. The insights into what customers are actually conversing about with the agents are extremely valuable for brands, showing us trends like customer interest in GLP-1 supplements or specific dietary requirements.
Steve Pimblett: Rightmove holds a vast amount of proprietary property data, including billions of minutes of browsing activity. Our estate agent partners publish around 10,000 listings every single day. Previously, this data ecosystem was siloed. We migrated to a unified cloud stack to bring it together and drive value across the network.
Because property search is highly multimodal, containing images, floor plans and virtual tours, we used Google Cloud to extract features from a billion property images to build a new metadata language. This allows consumers to engage in “conversational search” – they can talk to the local area and describe the uniqueness of what they want in a home. The hubs develop the guardrails, while the spokes innovate. This model allowed us to launch our conversational search tool in just six weeks.
Mohsin Ghazpour: In home improvement retail, customers do not just buy a tool for the sake of it – they are trying to solve a project, like wallpapering or putting up a shelf. We started our AI journey a few years ago with Google Cloud, and launched our first conversational search agent in December 2023.
Moving from traditional keyword search to proactive shopping agents helps customers discover and shop for home improvement projects more naturally. To make this work at enterprise scale, we built our unified data platform, Nucleus, which lets us launch new products quickly across our 10 multi-brand international entities. Once a product launches for B&Q, we can redeploy it for Screwfix just two weeks later.
Hayley McKelvey: Moving from transactional chat interfaces to autonomous agents is a completely new paradigm. At our London AI Studio, we help clients see the art of the possible by getting hands on keyboards.
We are starting to see substantial enterprise impact. For example, there was a recent report pointing to a 30% uplift in software coding efficiency being driven by AI in 2026. However, you cannot just drop these agents into an organisation and expect immediate value. The implementation has to be wrapped in a broader conversation around integration, trust, risk and data governance. Activations require access to high-quality data.
The economics of AI and the rise of ‘tokenomics’
Q: As agentic deployments scale, token efficiency and real-time cost management have become top priorities for technology chiefs. How are your organisations monitoring and managing token usage, and how does this tie into your cloud resource management?
Hayley McKelvey: Tokenomics is all we hear about right now. Organisations have signed consumption-based contracts with frontier AI labs, but many were not commercially ready for the financial impact. Technology leaders are suddenly focused on predicting and monitoring costs in real time. Cost transparency is achievable, but simply spending what you budgeted does not mean you have realised value.
We must avoid letting tokenomics become solely a CFO cost-monitoring exercise at the expense of value realisation. We address this in two ways. First, we manage the total cost of ownership by matching the task to the right model – using cheaper open source models for baseline tasks and reserving frontier models for complex logic. Second, we are building autonomous control planes to automatically route prompts to the most cost-efficient model without human intervention.
Jo Drake: Token management depends entirely on the use case. Some of our engineers are massive consumers of tokens during development. To manage this and speed up our delivery roadmap, we run “limitless engineering” sprints where we attempt to deliver a two-week sprint in two hours.
When pricing up products like the shopping assistant or virtual trials, we look closely at which models we use for specific tasks within the product. Some tasks do not require a frontier model – a baseline model is good enough.
We also focus on the tooling. Our shopping assistant is configured as a self-serve product on the Google Cloud Marketplace. Retailers can configure and test it without using a single minute of engineering resource, which prevents engineering capacity from becoming a scaling bottleneck.
Mohsin Ghazpour: At Kingfisher, we divide AI consumption into two distinct buckets. The first is commercial products, like our AI shopping assistant. For these public-facing products, everything – including conversion returns and token costs – is meticulously measured.
The second bucket is user-based engineering. We provide our developers with a platform that allows access to 14 different large language models, letting them test which model works best for their code. We apply user caps and quotas for experimentation to ensure costs do not run away.
Steve Pimblett: We call this AI Ops and FinOps. It is about understanding the exact use case and matching it with the right tool for the job.
We focus on balancing the cost of token consumption against the specific outcome and business value we are trying to achieve. As we scale autonomous agents, our KPIs [key performance indicators] and expectations around return on investment must evolve. We cannot measure these new agentic workflows using legacy IT metrics.
Evolving the workforce, culture and tech estate
Q: Transitioning to an agentic enterprise introduces significant cultural and operational changes. What are the key hurdles you are navigating in terms of workforce fear, team structuring, and the future of the enterprise tech estate?
Hayley McKelvey: The human experience is a critical aspect of this shift. Staff read about job displacement and labour market disruption, which can trigger a visceral “fight or flight” reaction. Leaders must address these emotional responses by creating an environment of psychological safety and trust.
Operationally, we are seeing the value of intergenerational leadership. We pair younger, AI-native staff with our senior leadership team to drive innovation, as the younger generation brings entirely new perspectives. We also run executive sessions with hands on keyboards to ensure leaders actually use the technology rather than just talking about it.
Jo Drake: We have seen our AI adoption drive from the bottom up. For the past two years, we have run an internal weekly AI podcast showcasing what employees across different departments have automated or streamlined, and the direct business return.
Culturally, we had to educate our teams as cost management became a prominent topic. Structuring the team has also evolved. Our engineers went through a massive cultural shift during our migration from on-premise datacentres to Google Cloud, as they became directly accountable for the costs of the cloud resources they consume. They have transitioned to a product mindset – working in multidisciplinary squads with product managers and UX designers focused on end-to-end customer features.
Mohsin Ghazpour: We run masterclasses for senior leaders and work with Google to educate our staff. The most fascinating cultural trend we observed is that as soon as you start educating operational employees, their fear of the technology turns into curiosity. They go from avoiding the conversation to suggesting 10 different ideas to make their roles more efficient.
From a team makeup perspective, AI is ultimately a tool of the mission – it is just another way of writing software. But we are sourcing more people for the “translation layer” – roles that can bridge the gap between technical AI capabilities and real-world business domains. As agent-to-agent protocols mature, systems will talk to each other more systematically.
Steve Pimblett: Our operating model is based on a hub-and-spoke structure. The hubs manage the centralised guardrails, AI Ops and security risks, while the spokes allow domain experts to innovate quickly.
We embed our data and AI experts directly into business units. We call this co-creation – bringing the technical teams together with the business leads who actually understand the domain problems. Looking ahead, the enterprise tech estate will become much more modular. There will likely be a period of vendor consolidation as enterprises evaluate which trusted platforms will remain relevant in an AI-interoperable landscape.
