The Promise and Pitfalls of Web-Agents in Insurance

In the fast-evolving landscape of AI technology, Web-Agents - smart bots that can reason through webpages to perform a task - have emerged as a promising solution to automate complex processes and rethink RPA. At Gaya, we’ve been closely monitoring and experimenting with this technology. Today, we’d like to share our insights on the current state of Web-Agents, their potential, and the challenges they face. 

The Current State of RPAs in Insurance

Robotic Process Automation (RPA) has been a buzzword in the insurance industry for years. These systems promise to automate repetitive tasks, reducing human error and freeing up valuable time. However, traditional RPAs often fall short of expectations. They’re prone to breaking when websites update, require constant maintenance, and struggle with complex, multi-step processes.

Language Action Models: A New Hope?

The advent of advanced language models has breathed new life into RPA. These AI-powered systems promise to understand and navigate web interfaces much like a human would, adapting to changes and handling complex tasks with ease. It’s an exciting prospect. Many startups like Browser-use, Skyvern, MultiOn (Please) are working on this technology. Even the giants released their versions: Anthropic's Claude Computer Use and most recently OpenAI's Operator.

We have tried most of them. Beyond the hypes of Linkedin and Twitter demos, where these Web-Agents magically manage to book calendar meetings and flights with natural language, they’re far from being ready for the complex, high-stakes world of insurance.

The Reality Check

While the potential of AI-powered Web-Agents in insurance is immense, the current reality is more sobering. Let’s break it down:

• Success rates: A typical insurance transaction (like changing a mortgage lender) might involve interacting with 30 different web pages. If a Web-Agent has a 95% success rate on each page (which is very optimistic), the overall success rate for the entire transaction drops to just 21% (0.95^30).
• Cost analysis: Processing a single page with 50,000 tokens (a reasonable estimate for a complex web form) costs about 10 cents. For a 30-page transaction, that’s $3. While this might seem reasonable, remember that a human could complete this task in about 5 minutes. At a wage of $36/hour, the human cost is $3 as well—and with a much higher success rate. Outsourcing to a VA or a BPO would drop the cost even further.
• Edge case handling: Many of these demos consist of Web-Agents being over-prompted to deal with edge cases. “Prompting your way through edge cases” makes Web-Agents more akin to a pre-recorded macro again, negating any major scalability gains.

Our Viewpoint on Web-Agents

Given these realities, we at Gaya are taking a measured approach:

• We’re not rushing into full automation. The technology isn’t quite there yet for complex insurance processes. All the Web-Agents we mentioned above are evaluating themselves against a simplistic benchmark (WebVoyager) that is far from reflecting the complexities human insurance agents and CSRs deal with.
• We are focusing on Human-assisted tools. Our current system allows humans to execute tasks on carrier portals with 95% accuracy, with easy human intervention for the remaining 5%.
• We are addressing the 80% that matters most. The biggest time sink for insurance agents isn’t necessarily navigating web pages but filling out forms. Form-filling accounts for about 80% of the time spent on quoting tasks. By streamlining this aspect, we can make a significant impact on agent productivity.
• Verticalized Web-Agents will dominate. Domain expertise is going to be the be-all, end-all for Web-Agents. By embedding specialized knowledge into their design, web-agents can operate with "System 1" thinking—quickly and intuitively managing tasks across portals without over-analyzing every step. Most insurance agents/ brokers rely on habit-based processes rather than deep reflection for routine tasks, and this habitual efficiency will find its perfect counterpart in verticalized Web-Agents tailored to specific industries like insurance.
• Robotic Process Automation is not dead. Even if it is cool to say otherwise, the ground reality in most businesses and enterprises is very different. "System 1" is kind of robotic and it will always have a place within the realm of automation in insurance.

The Future of Web-Agents at Gaya

While we’re cautious about current Web-Agent technology, we’re excited about its future potential. We believe the key lies in combining traditional RPAs with vision models and large language models (LLMs).

We’re working on a system that can execute pre-recorded macros (like traditional RPAs) but use AI vision and language understanding to adapt when it encounters errors or changes. This approach forms the foundation of our upcoming Gaya Service Tools—think of it as a marketplace of smart bots that can handle various service transactions across different carrier portals and agency management systems.

Conclusion

While Web-Agents hold immense promise for the insurance industry, we’re not quite there yet in terms of reliability and cost-effectiveness for complex processes. At Gaya, we’re committed to developing practical, efficient solutions that combine the best of AI technology with human expertise. Our focus on "Supercopy" and "Superpaste" functionality is already making a significant impact on agent productivity.

We invite you to join us on this journey of innovation - stay tuned for more updates as we continue to push the boundaries of what’s possible in insurance.