Senior Agent Product Manager

How Decagon uses AI for design system saturation. Jenny Xie Editor, Figma The fast-growing customer experience platform explains how Figma MCP and Figma Make helped them scale a new design system and keep pace with customer requests. Share How Decagon uses AI for design system saturation. Customer service is the next industry AI is poised to reshape, and Decagon is building the platform to do it. Just three years in, the company's AI agents span voice, chat, and email, replacing the ticket queues and hold times that have defined customer service for decades. That's how Decagon landed on the CNBC Disruptor 50 list - but what got them there isn't just what they build. It's how they build. At an AI-native company like Decagon, the design-to-code loop moves fast, but it can break down just as quickly. Coding agents need precise inputs to produce reliable output. But without a design system, every component comes down to a judgment call - and that doesn't scale. When Product Designer Jennifer Xu joined Decagon, there was no design system; just a growing product and a team moving fast enough to feel the absence of one. Inconsistency across the platform undermined the polish they needed to bring to enterprise customers, and back-and-forth revs between design and development cost time and effort. To build their design system, called Deco, designers and engineers worked hand-in-hand from the very beginning. "We were able to think about a lot of the edge cases and what already exists in code from the start," says Jennifer. That meant working through questions that would've been easy to defer and expensive to revisit later: What happens in focus mode? Which states does each component need - disabled, read-only, error, warning? Placeholder or no placeholder? The result is an org-wide library in Figma that now includes hundreds of components, styles, and variables encompassing the vast majority of use cases across the whole platform for multiple teams. According to Figma's library analytics, it logged tens of thousands of inserts in 30 days, a sign that people are using it. "With a built-out library, we aren't debating styles or implementation," says Jennifer. "Engineers have a clear view of which button or table should be used. It also means we have a shared vocabulary. As a designer, I can anchor on very similar primitives that an engineer does, so when we're building a product or thinking about a flow, it's a lot easier to communicate." Deco is published to the whole organization, helping designers assemble new screens from existing components, rather than building them from scratch. "It's also easier for developers to create work that is aligned to our overall design goals," says Jennifer. "As we're trying to ship quickly, we want to make sure everything is still aligned, and a design system is a really great way to have a single source of truth." Coding agents don't interpret specs the way a developer might; they work only with what they're given. If what they're given is inconsistent, the output reflects it. At Decagon, that handoff used to mean a familiar back-and-forth: designers exported specs, developers interpreted them, mismatches got caught in review, and the cycle repeated. The Figma MCP server changed that. Decagon's engineering team moved the design system's components into Storybook, then created a skill for their coding agents to use exact components when implementing designs. Another skill lets designers add new components, keeping Figma and code in continuous parity. "Our agents have the Figma MCP enabled, so all the specs, code, and canvas stay in one loop instead of having to bounce back and forth," says Jennifer. Connecting to Deco through the Figma MCP sped up the pace of iteration. Now, coding agents can read directly from Figma to create high-fidelity starting points. "With MCP, I can just copy and paste a link of my Figma into the coding agent, and it'll not only use the skill of getting design context, but also map it to our design system components," says Jennifer. "It can create things that are really high-fidelity and close to our design without having to do a lot of nit iterations." Roughly 70 percent of Decagon's product roadmap comes directly from customers. It's not just their philosophy; it's core to how the team builds. "We'll do hands-on roadmap sessions with a customer's team to make sure that the features we're shipping are addressing what they need," says Bihan Jiang, director of product at Decagon. Being able to present mockups and working prototypes in Figma is central to that process. "Tools like Figma Make and connectors to all the systems we use enable us to raise the ceiling on the products we're building," she continues. Instead of gathering requirements and presenting an end product, the team can now build 10 different prototypes and show them to 10 different customers. Says Bihan, "The result is a better product that is built much more quickly." Interactive dashboards, metrics, and AI insights are core to what Decagon offers their customers who need to understand both quantitative and qualitative data - like Customer Satisfaction Scores and sentiment in conversations across channels. One Decagon PM uses Make to prototype new graphs and redesign pages to surface different data sets based on what customers are asking for. "The PM could prompt, 'I want you to do it in the style of this page' and paste in a Figma screenshot, and that would allow us to think about the tool and the graph in the context of the overall platform," says Jennifer. "We could then talk to our engineering team and our customer-facing team to see whether or not the change was actually useful before we start investing a lot of engineering and design resources into it." The same principle applied when the team wanted to revamp an interactive chart to be cleaner and more user-friendly. The PM fed the original chart into Make, described the changes they wanted, and sent that to the developer. Jennifer says, "It was a lot simpler than the process of translating that to a designer, a designer creating a mock, and then the engineer developing it. It cut down on development time." High-fidelity prototypes allow the team to work in a visual medium from the start, rather than translate ideas from docs to wireframes. "As a designer, I'm thinking visually, and when you're working on analytics, the problem is visual," says Jennifer. "I'm looking at it as a picture, not as a doc. So starting to think in the modality the end product is in, from the very beginning, is useful." That instinct to get concrete - and get it in front of customers early - has become part of Decagon's DNA: "It's a big part of our culture to use Figma as part of our brainstorming. This allows collaboration to be faster and at a higher fidelity than it used to be." Decagon centers AI not just in the platform they offer, but in how they build it. The design system keeps craft consistent as the team scales. MCP keeps design and code aligned, so nothing gets lost in translation. And Figma Make puts visual thinking in the hands of anyone who has an idea worth testing. "This new world of design is really exciting," says Jennifer, "and I think we're really at the forefront of it. It allows us to ship high-quality things very quickly." For Decagon, gaining speed without sacrificing craft is the real edge. Jenny Xie is a writer and editor at Figma and the author of the novel Holding Pattern. Her work has appeared in places like The Atlantic, Esquire, and Dwell, where she was previously the Executive Editor.

Databricks AI + data: ai-native development is reshaping software creation. Founders from several AI startups explain how AI agents are accelerating engineering, boosting worker productivity and changing enterprise operations, but not replacing humans. Bob is the managing editor of BizTech magazine. As generative AI continues to mature, one question looms large: If foundation models are becoming increasingly powerful, where will the next wave of value creation occur? According to leaders from three rapidly growing AI startups, the answer lies not in the models themselves but in the application layer that sits above them. During a panel discussion at Databricks Data + AI Summit, executives from Cognition, Glean and Decagon described a future in which natural language becomes the primary interface for building software, automating workflows and interacting with enterprise systems. But despite the rapid advances in large language models, they argued that significant engineering challenges remain. "A foundation model by itself doesn't provide many of the capabilities enterprises require," said T.R. Vishwanath, co-founder and CTO of Glean. Organizations need systems that understand enterprise data, enforce governance policies and deliver information tailored to specific users and tasks, he said. Those requirements have helped fuel the growth of application-layer AI companies. Rather than simply passing prompts to foundation models, these platforms combine multiple models, enterprise data sources, security controls and orchestration layers to solve specific business problems. For Glean, an AI-powered work assistant that connects to organizations' business applications, that means connecting to enterprise knowledge systems while preserving permissions and governance. For Cognition, maker of the AI coding agent Devin, it means creating infrastructure that allows agents to understand code bases, validate work and proactively assist development teams. For Decagon, which builds AI-powered customer service agents, the challenge is delivering accurate, low-latency interactions at enterprise scale. "Even if you have 99% performance, that remaining 1% at enterprise scale is 10,000 times a day" that the AI may be hallucinating, said Ashwin Sreenivas, Decagon's co-founder and president. As a result, companies are investing heavily in safeguards, testing frameworks and specialized models designed for specific tasks, rather than relying solely on a single frontier model. Why AI success depends on more than a single model. One recurring theme throughout the discussion was the growing importance of model orchestration. Rather than standardizing on a single foundation model provider, panelists described environments where dozens of models may be used for different tasks, balancing performance, latency and cost. Jeff Wang, president of new enterprise at Cognition, said his company evaluates models continuously and routes workloads based on performance and economics. Token costs, he noted, have become one of the most common topics in executive conversations. Similarly, Glean supports multiple frontier and open-source models while automatically selecting the most appropriate option for a given task. Decagon takes an even more specialized approach, Sreenivas said, using teams of smaller models that each perform a specific function, such as gathering information, generating responses or detecting errors. The result is an AI stack that increasingly resembles a coordinated system of agents rather than a single monolithic model. That complexity extends beyond customer-facing products. Panelists said their own organizations are aggressively using AI internally and measuring usage patterns to identify high-value applications. Sreenivas described how Decagon analyzes internal AI consumption and studies how top users are achieving productivity gains. In many cases, employees are using AI to create highly personalized customer briefings, automate administrative work and streamline customer engagement processes. AI is changing workflows, not eliminating the need for people. The conversation also addressed one of the most debated questions in AI: whether automation will replace workers or simply make them more productive. The consensus among panelists leaned strongly toward augmentation. Wang said many of the most successful use cases today involve work that is repetitive, tedious or frequently backlogged. Examples include software vulnerability remediation, bug replication and application modernization projects. Rather than reducing the need for engineers, he argued, AI often enables organizations to pursue larger ambitions. "We are hiring more engineers because they're more productive, and they're getting our roadmap ahead," he said. Vishwanath echoed that view, describing how AI allows employees to complete tasks in minutes that previously required hours. Product teams can analyze hundreds of customer calls, generate presentations and synthesize large volumes of information on a routine basis. The technology is also reshaping how AI-native companies organize themselves. Several panelists described moving toward smaller, more autonomous teams supported by coding agents and automation tools. These teams can build and ship software more quickly, but they also require stronger testing, governance and coordination mechanisms. Despite the rapid pace of change, the panelists rejected the notion that AI will eliminate the need for specialized software companies or human expertise. Instead, they argued that as models become more capable, opportunities for innovation will expand alongside them. "The ceiling is so high," Vishwanath said. "If the models do more, then we do more."