.png)
Agents have featured in the AI conversation for years—but actual usage was relatively rare, and mostly limited to software development. Recently, that has started to change. The rise of personal agents, and the increasing ease with which organizations can build agentic workflows, have created the sense that the age of the agent has finally begun.
Defining an agent as a system that can “set goals, reason through multi-step tasks, use tools and application programming interfaces (APIs), and coordinate work with people or other agents”, a recent survey shows that a quarter of organizations are currently using agentic AI at least moderately. The report predicts this will rise to three quarters over the next two years.
This shift has happened fast. At this year’s Google Cloud Summit in London, which showcased some of the latest developments in AI, cloud infrastructure, and data analytics, successful agentic enterprises featured in nearly every session. “I went back and I looked at all of the keynotes from last year, and we only mentioned agentic once,” said Maureen Costello, Google Cloud’s vice-president of UK, Ireland, and sub-Saharan Africa, as the event kicked off. “This year, you’ll hear it many more times—because the agentic enterprise is happening right here, right now.”
Later that day, energy company OVO revealed how its customer service agents take actions to reduce call volumes. Home-improvement company Kingfisher shared how its price-optimizing agents have increased profit margins. And the insurer Hiscox showed how equipping teams with agents—and having these agents communicate with each other—has reduced information-gathering time during underwriting from hours to minutes.
At the same time, speakers acknowledged that deploying this technology comes with challenges, particularly when it comes to building trust.
Across sessions, organizational leaders and AI specialists discussed how to successfully navigate becoming an agentic enterprise. Here are four of their key pieces of advice…
“The first thing, I’d say, is just having a laser focus on outcomes,” said Felix Reilly, a program lead at the UK government’s Incubator For AI. Successful organizations, he says, target a key area where agents can improve a workflow or service, such as onboarding clients or handling customer queries, and then work deeply on this. In addition to channeling energies in the right direction, it helps staff get behind adopting agents. “You buy a lot of trust, because people can see things that are going to solve a problem for them.”
It’s important to remember, however, that a conservative scope shouldn’t mean a conservative execution. Don’t fall into the trap of limiting yourself by building a dummy agent or sandboxing with dummy data. “Companies should develop real tools that actually work from the very start,” advised Alex Rutter, EMEA managing director of AI at Google Cloud. “If the business loves it, and it does what it says on the tin, then you can scale it rapidly.”
“Tokenomics” was the buzzword of the summit. Because AI services are charged by the token, unpredictable and rapidly escalating costs have become a primary challenge for chief technology officers and financial leaders alike.
At the event, representatives from organizations highlighted that managing these costs can be a critical hurdle to scaling AI initiatives. To address this, the conversation is shifting toward selecting models with a highly efficient performance-to-cost ratio, which are engineered to deliver high performance at a lower operational cost.
"Unpredictable costs can be a deterrent to enterprise AI adoption," said Rutter. "To move past experimentation, organizations need a platform like Gemini Enterprise Agent Platform that offers both high-performing, cost-efficient models like Gemini and the granular FinOps controls necessary to predict and manage token spend at scale."
As part of that drive for cost-effectiveness, it’s important to pick the right model for the right task. When Kingfisher added shopping-assistant agents to its websites, for instance, it built some of their functionality with traditional machine learning algorithms. “The compute cost for that is near to zero,” said Mohsen Ghasempour, Chief AI Officer at Kingfisher, but it achieved the same effect.
Thanks to products such as Google’s Agent Designer, AI agents can now be created using no-code and low-code tools. This means coding skills aren’t essential for building agents, Rutter explained, but AI-complementary ones are.
“Can you think critically? Can you explain business processes logically? Can you prompt in a way that gets the output you’re looking for?” These are the sorts of adjacent skills that can help people make effective agents that do what was intended.
Where you will need specialists, however, is for connecting AI agents properly to complex legacy data sets. “In order to get things done, you have to know the nitty-gritty of the organization, its legacy systems, everything,” said Ghasempour.
At Kingfisher, many of those responsible for connecting its AI systems to its data were employees. But for companies that don’t already have this expertise on staff, it’s possible to partner with a model provider and have them provide the technical know-how. Google, Anthropic, and other AI companies all offer teams, known as “forward-deployed engineers”, to help organizations with this kind of work.
The constant acceleration of AI can be exciting. As models progress, it creates new possibilities for agents you’re developing. Monika Delekta-Ebbage, head of data science at Hiscox London Market, was part of the team that developed the insurer’s underwriting agents. “New models were coming out every week,” she told the summit, reflecting on the project’s build phase. “You could experiment, change the prompts.”
But once agents are built and tested, the situation flips: in deployment you want stable performance—and changes to an underlying model don’t necessarily mean an agent will get better. With Hiscox’s agents, “it was about how we maintained that stability and control when we productionized it,” Delekta-Ebbage said.
To avoid drift, agents need to be regularly evaluated. “Evals are one of the most critical things,” said Caroline Matthews, AI and machine learning architect at Anthropic. When building agents, she suggests creating a set of at least 20 to 30 benchmark tests that let you assess accuracy and quality. “This isn’t an afterthought,” Matthews said. “We actually want to do it right at the beginning.” If you have lots of agents, you can use LLMs to help judge agentic performance. But make sure to have a human in the loop to spot check the quality of the LLM’s work.
In addition to checking outputs, also scrutinize transcripts—logs of what agents are thinking and doing as they work—to make sure their processes remain optimized. Circular reasoning or unnecessary steps waste tokens.
AI agents may have been a big new talking point of this year’s summit, but don’t expect to see a similar shift in the agenda next year—agentic AI will likely continue to be a key business trend and the concept will keep evolving. It’s probable that in coming years, multi-modal agents able to process visual and audio information will start becoming more commonplace, allowing them to experience the world more like humans do, while the integration of AI and robotics will see agents, over a longer time horizon, move into the physical realm.
Rutter believes that the shift to agent-enabled work will be quick, so enterprises need to push forward to stay competitive. “The pace is accelerating rapidly, and enterprises can no longer afford to sit back—this is the moment to seize the lead and actively build what comes next.”
Whether you're a senior leader or a hands-on builder, click here to access Google Cloud's on-demand portal and watch the defining keynotes and "best of" sessions from the London Summit whenever it suits you.
