Roundtable: UK tech chiefs on agentic AI, workforce tradition and tokenomics
The shift from experimental synthetic intelligence (AI) initiatives to enterprise-grade autonomous programs was the defining theme of the Google Cloud Summit in London final week.
As UK enterprises transfer to action-oriented agentic workflows, expertise leaders encounter a brand new set of challenges. To construct a prototype agent is easy, however to deploy tons of of them at scale calls for rigorous information infrastructure, strict safety governance and fine-grained value administration.
This transition can be redefining the economics of IT, with “tokenomics” – the monitoring and optimisation of huge language mannequin (LLM) token consumption – rising as a significant focus for chief monetary officers (CFOs) and expertise administrators.
We sat in on a Google-hosted roundtable that featured Jo Drake, chief expertise officer of platforms at THG Ingenuity; Mohsin Ghazpour, chief AI officer at Kingfisher; Steve Pimblett, chief information officer at Rightmove; and Hayley McKelvey, chief AI officer at Deloitte.
The panel mentioned enterprise outcomes of their preliminary agent roll-outs, the sensible methodologies they use to trace and handle token budgets, and the cultural shifts required to organize engineering groups and wider workforces for an autonomous enterprise.
From chatbots to autonomous agent deployments
Q: We’re seeing a market transition from fundamental AI experimentation to concrete, real-world agent deployments. How are proactive agentic fashions reworking person experiences and enterprise outcomes in your respective sectors?
Jo Drake: The advantages we’ve got seen are threefold. First is the shopper expertise. Purchasing is conversational by definition – you stroll right into a retailer, somebody greets you, and also you clarify what you might be in search of, whether or not it’s for a marriage or an athletic occasion. We’ve got introduced that conversational expertise to e-commerce web sites and apps utilizing our procuring assistant. We will layer hyper-personalisation onto that, understanding the shopper’s measurement, color preferences and shopping for objectives.
The second profit is efficiency. For MyProtein, the place we piloted this, we noticed a large enhance within the conversion fee for first-time guests, together with a 22% enhance in basket measurement and a 20% enhance in common order worth.
The third profit is the info loop. The insights into what clients are literally conversing about with the brokers are extraordinarily helpful for manufacturers, displaying us developments like buyer curiosity in GLP-1 dietary supplements or particular dietary necessities.
Steve Pimblett: Rightmove holds an enormous quantity of proprietary property information, together with billions of minutes of shopping exercise. Our property agent companions publish round 10,000 listings each single day. Beforehand, this information ecosystem was siloed. We migrated to a unified cloud stack to convey it collectively and drive worth throughout the community.
As a result of property search is extremely multimodal, containing pictures, ground plans and digital excursions, we used Google Cloud to extract options from a billion property pictures to construct a brand new metadata language. This permits customers to have interaction in “conversational search” – they will speak to the native space and describe the distinctiveness of what they need in a house. The hubs develop the guardrails, whereas the spokes innovate. This mannequin allowed us to launch our conversational search instrument in simply six weeks.
Mohsin Ghazpour: In dwelling enchancment retail, clients don’t simply purchase a instrument for the sake of it – they’re making an attempt to resolve a mission, like wallpapering or placing up a shelf. We began our AI journey just a few years in the past with Google Cloud, and launched our first conversational search agent in December 2023.
Transferring from conventional key phrase search to proactive procuring brokers helps clients uncover and store for dwelling enchancment initiatives extra naturally. To make this work at enterprise scale, we constructed our unified information platform, Nucleus, which lets us launch new merchandise shortly throughout our 10 multi-brand worldwide entities. As soon as a product launches for B&Q, we are able to redeploy it for Screwfix simply two weeks later.
Hayley McKelvey: Transferring from transactional chat interfaces to autonomous brokers is a totally new paradigm. At our London AI Studio, we assist purchasers see the artwork of the doable by getting arms on keyboards.
We’re beginning to see substantial enterprise impression. For instance, there was a current report pointing to a 30% uplift in software program coding effectivity being pushed by AI in 2026. Nonetheless, you can not simply drop these brokers into an organisation and count on quick worth. The implementation must be wrapped in a broader dialog round integration, belief, danger and information governance. Activations require entry to high-quality information.
The economics of AI and the rise of ‘tokenomics’
Q: As agentic deployments scale, token effectivity and real-time value administration have turn out to be high priorities for expertise chiefs. How are your organisations monitoring and managing token utilization, and the way does this tie into your cloud useful resource administration?
Hayley McKelvey: Tokenomics is all we hear about proper now. Organisations have signed consumption-based contracts with frontier AI labs, however many weren’t commercially prepared for the monetary impression. Expertise leaders are immediately targeted on predicting and monitoring prices in actual time. Value transparency is achievable, however merely spending what you budgeted doesn’t imply you’ve got realised worth.
We should keep away from letting tokenomics turn out to be solely a CFO cost-monitoring train on the expense of worth realisation. We handle this in two methods. First, we handle the full value of possession by matching the duty to the fitting mannequin – utilizing cheaper open supply fashions for baseline duties and reserving frontier fashions for advanced logic. Second, we’re constructing autonomous management planes to routinely route prompts to probably the most cost-efficient mannequin with out human intervention.
Jo Drake: Token administration relies upon solely on the use case. A few of our engineers are large customers of tokens throughout growth. To handle this and velocity up our supply roadmap, we run “limitless engineering” sprints the place we try and ship a two-week dash in two hours.
When pricing up merchandise just like the procuring assistant or digital trials, we glance carefully at which fashions we use for particular duties inside the product. Some duties don’t require a frontier mannequin – a baseline mannequin is nice sufficient.
We additionally give attention to the tooling. Our procuring assistant is configured as a self-serve product on the Google Cloud Market. Retailers can configure and check it with out utilizing a single minute of engineering useful resource, which prevents engineering capability from changing into a scaling bottleneck.
Mohsin Ghazpour: At Kingfisher, we divide AI consumption into two distinct buckets. The primary is industrial merchandise, like our AI procuring assistant. For these public-facing merchandise, all the pieces – together with conversion returns and token prices – is meticulously measured.
The second bucket is user-based engineering. We offer our builders with a platform that permits entry to 14 totally different giant language fashions, letting them check which mannequin works greatest for his or her code. We apply person caps and quotas for experimentation to make sure prices don’t run away.
Steve Pimblett: We name this AI Ops and FinOps. It’s about understanding the precise use case and matching it with the fitting instrument for the job.
We give attention to balancing the price of token consumption towards the precise end result and enterprise worth we are attempting to realize. As we scale autonomous brokers, our KPIs [key performance indicators] and expectations round return on funding should evolve. We can not measure these new agentic workflows utilizing legacy IT metrics.
Evolving the workforce, tradition and tech property
Q: Transitioning to an agentic enterprise introduces vital cultural and operational modifications. What are the important thing hurdles you might be navigating when it comes to workforce worry, group structuring, and the way forward for the enterprise tech property?
Hayley McKelvey: The human expertise is a crucial side of this shift. Employees examine job displacement and labour market disruption, which might set off a visceral “struggle or flight” response. Leaders should handle these emotional responses by creating an surroundings of psychological security and belief.
Operationally, we’re seeing the worth of intergenerational management. We pair youthful, AI-native workers with our senior management group to drive innovation, because the youthful era brings solely new views. We additionally run govt periods with arms on keyboards to make sure leaders truly use the expertise slightly than simply speaking about it.
Jo Drake: We’ve got seen our AI adoption drive from the underside up. For the previous two years, we’ve got run an inside weekly AI podcast showcasing what staff throughout totally different departments have automated or streamlined, and the direct enterprise return.
Culturally, we needed to educate our groups as value administration turned a outstanding matter. Structuring the group has additionally developed. Our engineers went by way of a large cultural shift throughout our migration from on-premise datacentres to Google Cloud, as they turned straight accountable for the prices of the cloud assets they devour. They’ve transitioned to a product mindset – working in multidisciplinary squads with product managers and UX designers targeted on end-to-end buyer options.
Mohsin Ghazpour: We run masterclasses for senior leaders and work with Google to coach our workers. Probably the most fascinating cultural pattern we noticed is that as quickly as you begin educating operational staff, their worry of the expertise turns into curiosity. They go from avoiding the dialog to suggesting 10 totally different concepts to make their roles extra environment friendly.
From a group make-up perspective, AI is in the end a instrument of the mission – it’s simply one other means of writing software program. However we’re sourcing extra folks for the “translation layer” – roles that may bridge the hole between technical AI capabilities and real-world enterprise domains. As agent-to-agent protocols mature, programs will speak to one another extra systematically.
Steve Pimblett: Our working mannequin is predicated on a hub-and-spoke construction. The hubs handle the centralised guardrails, AI Ops and safety dangers, whereas the spokes enable area consultants to innovate shortly.
We embed our information and AI consultants straight into enterprise models. We name this co-creation – bringing the technical groups along with the enterprise leads who truly perceive the area issues. Trying forward, the enterprise tech property will turn out to be rather more modular. There’ll possible be a interval of vendor consolidation as enterprises consider which trusted platforms will stay related in an AI-interoperable panorama.
