The Story of Synthpop: Building AI Agents That Actually Execute

The Story of Synthpop: Building AI Agents That Actually Execute

In a recent episode of Category Visionaries, Elad Ferber, CEO of Synthpop, revealed the origin story behind a company that’s redefining what AI agents can do. This isn’t another chatbot company promising to deflect a few support tickets. Synthpop built something fundamentally different: AI agents that can actually take action across a company’s entire tech stack.

The Insight That Started Everything

Elad’s journey into AI agents began with a simple observation about the limitations of existing automation tools. He watched companies deploy chatbot after chatbot, only to see the same pattern repeat: initial excitement, followed by disappointment when the bots couldn’t handle anything beyond basic FAQs.

The problem was architectural. As Elad explains: “The vast majority of agents out there, they don’t have any access to your back end systems. They don’t have any access to your database. They’re just LLMs that are trained on text, and they’re trying to provide good answers based on that text.”

These text-only agents could read knowledge bases and provide information, but they couldn’t execute actions. When a customer needed a refund processed, a subscription upgraded, or account information updated, the chatbot had to hand off to a human. Companies were automating conversations but not resolutions.

Elad saw an opportunity to build something different: agents that could not only understand customer requests but actually execute the necessary actions to resolve them. The vision was clear, but the path to building it would prove complex.

Building the Foundation: Tool Access and Execution

The technical architecture Synthpop developed represents a significant departure from traditional chatbot systems. Rather than treating AI agents as conversational interfaces sitting on top of support workflows, they built agents that plug directly into a company’s operational infrastructure.

Elad describes the approach: “We give agents access to all the tools that your support agents have access to. So that means access to Stripe, access to your database, access to your own internal APIs, to Salesforce, to your CRM.”

This required solving multiple hard problems simultaneously. The agents needed to understand natural language requests and map them to specific actions across different systems. They needed to handle authentication and permissions correctly, ensuring they could only execute actions they were authorized to perform. They needed to manage complex multi-step workflows where one action depends on the outcome of previous actions.

Most critically, they needed to do all of this reliably enough that companies would trust them to operate autonomously. A chatbot that gives a wrong answer is frustrating. An agent that processes an incorrect refund or modifies the wrong account is a liability.

The Path to 93% Autonomous Resolution

The milestone that proved Synthpop’s approach works came when their agents began consistently resolving 93% of customer interactions without any human intervention. This wasn’t ticket deflection—where a chatbot convinces customers to solve their own problems—but true end-to-end resolution.

Elad is candid about the remaining 7%: “There’s always going to be the super long tail of one off edge cases that require human intervention.” These are genuinely complex situations that require human judgment, creativity, or the ability to make exceptions to standard policies.

But by handling the vast majority autonomously, Synthpop’s agents transform the economics of customer support. Companies can scale support operations without proportionally scaling headcount. Support teams can focus their time on complex, high-value interactions rather than repetitive routine tasks.

The journey to 93% resolution wasn’t just about building better AI models. It required deep integration work for every customer deployment, connecting to each company’s specific payment processors, databases, CRMs, and proprietary systems. It required training agents on company-specific policies, edge cases, and workflow requirements.

This complexity became a feature, not a bug. The investment required to integrate Synthpop deeply into a company’s operations creates significant switching costs, making the platform stickier than shallow integrations that are easy to replace.

Learning to Sell Infrastructure Software

Building the technology was only half the challenge. Synthpop had to figure out how to sell a product that required significant implementation work and asked customers to trust AI agents with actions that had real business consequences.

The breakthrough came from leading with data rather than promises. Elad describes their evolved sales approach: “We actually go through historical support tickets with customers and kind of go tag by tag and kind of map out, okay, how much of your support volume can we actually automate?”

This data-driven discovery process transforms the sales conversation. Instead of asking prospects to believe in the theoretical capabilities of AI agents, Synthpop shows them their own historical data analyzed to predict exact automation rates. Prospects see concrete projections—60%, 70%, 80%+ of their specific ticket volume can be automated—before making any commitment.

The approach also helps qualify customers efficiently. If a prospect’s ticket composition doesn’t align with Synthpop’s capabilities, both sides discover that early in the process.

The Pricing Evolution

Figuring out how to price AI agents presented unique challenges. Elad recognized that market immaturity affects willingness to pay: “I think a lot of companies are still hesitant to pay a lot for AI agents because AI is so new.”

Rather than trying to capture maximum value immediately, Synthpop priced for adoption, knowing they could expand revenue once customers experienced the value in production. They settled on per-resolution pricing, charging only when the AI agent successfully resolves an issue.

This model aligns incentives elegantly. Customers pay for outcomes, not potential. Synthpop’s engineering priorities naturally focus on improving resolution rates. Revenue scales with realized value rather than arbitrary usage metrics.

The Vision: Beyond Customer Support

For Elad, customer support is just the beginning. He sees it as the ideal wedge into a much larger opportunity: autonomous business operations.

“Customer support is a great wedge because it’s like one of the most standardized and routine business processes,” Elad explains. The same architectural approach—giving AI agents tool access and execution capabilities—can extend far beyond support.

Sales operations, back-office functions, data entry, scheduling, procurement—most business processes are more routine and automatable than people assume. They just require agents that can interact with systems rather than merely conversing with humans.

Synthpop’s long-term vision is to become the platform that enables AI agents to handle entire categories of business operations autonomously. Starting with customer support lets them prove the model, build robustness, and establish trust. From there, the platform can expand horizontally into any business function that involves routine decision-making and system interactions.

The key insight driving this vision is that the current wave of AI agents is dramatically underutilized. Most companies are using sophisticated AI models to power glorified chatbots when those same models could be orchestrating complex workflows across entire organizations.

As Synthpop continues scaling, they’re not just building a customer support automation tool. They’re building the infrastructure layer that will enable the next generation of autonomous business operations—where AI agents don’t just advise humans on what actions to take, but actually execute those actions themselves.