The journey from AI experimentation to production deployment is accelerating rapidly, with companies now building end-to-end agentic workflows that deliver measurable business value rather than isolated proof-of-concept projects. Companies are now creating BBQ chatbots in just 8 weeks.
In a conversation at Google Cloud AI Live with Matthijs van den Berg, Director of Customer Engineering at Google Cloud, he shared insights into how the gap between where organizations want to be with AI and where they currently are is closing faster than many expected. The evolution from simple chatbots to sophisticated autonomous agents represents a fundamental shift in how businesses approach AI implementation.
The evolution from chatbots to agentic workflows
Three years into the current AI journey, organizations have progressed well beyond the initial fascination with chatbot capabilities. While the past two years focused heavily on implementing single-use case agents, the industry is now witnessing companies building comprehensive end-to-end agentic workflows integrated directly into their business processes.
“What we’re now seeing is that companies are actually starting to build end-to-end agentic workflows in their businesses,” van den Berg explains. “And that really makes a difference because now you can really see the benefits of AI in your organization.” The combination of Google’s models, platform capabilities, and cloud infrastructure with organizations’ domain knowledge, processes, and data creates what he describes as a “flywheel of innovation” that generates tangible returns on investment.
Also read: Google Gemini Enterprise to become the AI platform for everyone
Autonomous AI development in action
One of the most striking examples van den Berg shares involves a staffing company where an initially skeptical engineer transformed his approach to AI-assisted development. This engineer, who was reluctant to adopt AI due to quality concerns, took several weeks to immerse himself in the technology and built a comprehensive framework that met his standards.
The results were remarkable. “Instead of having larger development teams, they have smaller teams, people that understand the business and understand technology, and they’re actually creating code at speed, at scale with AI,” van den Berg notes. The most impressive aspect is the autonomous nature of the system: “They let the agent run, it generates code, and they only check the output and stop the agent only if the output is not good enough for them. And it runs for days and days, sometimes even weeks in a row without stopping, without correcting it.”
This engineer’s confidence in the system has grown to the point where he claims he could rebuild the entire application from scratch using AI in just six months with sufficient token allocation. This level of autonomous operation represents a significant milestone in AI-assisted development, though van den Berg acknowledges that early implementations required building custom frameworks to manage context and task decomposition, capabilities now increasingly built into newer models with larger context windows.
Democratizing AI development with AI Studio
For organizations seeking to begin their AI journey or scale beyond experimental projects, Google’s AI Studio provides an accessible entry point. Van den Berg uses AI Studio extensively for both hobby projects and customer demonstrations, highlighting its rapid deployment capabilities.
“You just give it what you want. It creates a demo for you. And with one click, you deploy to Cloud Run. You deploy it behind a secure IAM proxy. It has a Firebase database below, and it just works. It’s secure,” he describes. This streamlined approach removes much of the infrastructure complexity that previously slowed AI project development.
The platform strategy addresses different maturity levels within organizations. AI Studio targets low-code agent development suitable for business users, while Agent Builder, Agent Designer, and tools like Antigravity cater to high-code agent requirements involving complex processes and nuanced datasets. The underlying principle remains consistent: democratize AI development by enabling more people across the organization to create agents, capturing diverse inspiration and expertise in AI processes.
The Gemini Enterprise Agent Platform
As AI implementations matured, the industry recognized the need for comprehensive frameworks similar to those developed for cloud computing. “Two years ago, there was no agent registry. There was no observability framework. There was no security framework for agents or how we handle those. And that is all in place now,” van den Berg explains.
The Gemini Enterprise Agent Platform addresses this need by providing the infrastructure plumbing necessary to run AI agents at scale with appropriate confidence levels. It handles security, audit trails, compliance with rules and regulations, and observability. These are the critical requirements for enterprise deployment that individual teams would otherwise need to build themselves.
Van den Berg envisions a model in which IT departments retain responsibility for the secure, audited platform infrastructure while empowering users with coding tools such as Antigravity. A translation framework ensures generated code meets organizational specifications, adheres to design principles, and deploys to approved platforms. Security integration tools like CodeMender further ensure code compliance and vulnerability management throughout the development lifecycle, embodying the “shift left” principle of building security in from the start.
Model selection and cost optimization
As AI capabilities have improved, the question of model selection has become increasingly nuanced. Van den Berg emphasizes that for most use cases, models are now “good enough,” making it crucial to balance capabilities against cost considerations.
Google Cloud has the model garden, which offers various models, including smaller, specialized, and industry-specific options, enabling organizations to match model capabilities to specific use cases rather than applying a one-size-fits-all approach. These models can help balance capabilities and costs. Also, Google’s cheaper models are now more capable, according to Van den Berg.
“I think Gemini 3.5 Flash is a very strong model because it strikes a strong balance between capabilities and cost,” he notes, referencing reports of companies spending half a billion dollars on AI tokens in a single month, clearly not a sustainable situation for most organizations. The principle of “AI fluency” becomes important: understanding which models suit which tasks rather than defaulting to the most powerful option for every query.
Van den Berg provides a practical example: “We have a chatbot with a big bank. They actually use 2.5 Flash, and that works perfectly because for their use case […] this is the right model versus the right cost.” Conversely, deep research tasks that require extensive web research and large context windows justify using top-tier models like Gemini Pro or Claude to enable genuine AI-driven inspiration and insight.
The platform includes infrastructure tools for continuous validation, ensuring that model changes or upgrades maintain answer quality and correctness. This systematic approach to model management helps organizations optimize their AI investments while maintaining performance standards.
From experimentation to production deployment
Many organizations remain in what van den Berg characterizes as an “experimental phase,” running small AI projects without clear paths to production. However, the technology and frameworks have matured significantly. “I think the play days are over and we’re now actually building more intentful towards business solutions,” he observes.
The key insight is that low-hanging fruit use cases now have established examples in the market. Organizations no longer need to figure everything out from scratch. For customer service chatbots, SaaS solutions like Gemini Enterprise for Customer Experience enable drag-and-drop agent creation. The technology itself is no longer the primary barrier to production deployment.
Van den Berg shares the example of Jumbo, a company that appeared on stage at the Google Cloud AI Live event. “Jumbo actually told us that they went from the start of the project to a working chatbot on their website in just 8 weeks,” he reveals. This wasn’t a simple FAQ bot answering store opening times, it’s a sophisticated system where customers can describe needs like planning a barbecue, receive product recommendations for meat, meatless options, and other items, add them to a cart, and proceed through checkout, all integrated with customer data and purchase history.
This rapid deployment required courage and willingness to be uncomfortable on Jumbo’s part, but it demonstrates what’s achievable with current AI platforms and tools when organizations commit to moving beyond pilot projects.
Strategic considerations for AI implementation
Organizations embarking on AI implementation should consider several strategic factors. First, clarify the primary business objective: increasing revenue (topline), reducing costs, or mitigating risk. These three drivers shape implementation priorities and success metrics.
Second, balance budget constraints with innovation needs. While setting token usage limits prevents unexpected bills, van den Berg notes that overly restrictive budgets can hamstring productive engineers. The better approach involves establishing reasonable guardrails while demonstrating that AI investments deliver measurable benefits and efficiency gains, creating a virtuous cycle that justifies increased investment.
Third, recognize the emergence of “unicorn engineers”, individuals who understand business processes, technology, customers, and organizational dynamics. While rare, AI tools can help bridge this gap by enabling more people to combine business understanding with technical implementation capabilities.
The path forward
The convergence of improved model capabilities, comprehensive enterprise platforms, established best practices, and real-world success stories has created an environment where AI implementation can progress rapidly from concept to production. Organizations can now adopt AI and innovate faster than ever. Google and several other companies are providing the tools to do so. We have left the pilot phase, and many organizations are now running AI in production, combining AI technology with their own domain expertise. It’s now time to make a move; organizations that wait any longer risk being disrupted by their direct competitors. As van den Berg’s examples demonstrate, the gap between AI aspirations and reality is now very narrow, making this an opportune moment for organizations to accelerate their AI journeys.