Elevating Enterprise Customer Interactions Using ASAPP’s AI-Native Platform and Amazon Bedrock

Instead of offering a stilted, menu-driven interactive system, ASAPP wanted users to feel like they were talking to a real person. Whether the GenerativeAgent Platform engaged by voice or by chat, it needed to be able to use natural language, understand context and tone, and respond quickly and accurately to customers.

ASAPP also prioritized helping its customers protect sensitive user data. But its homegrown PII redaction layer slowed interactions and hurt the quality of GenerativeAgent’s responses. “It’s a bit of a headache when you’re working on a customer deployment and you see that GenerativeAgent would provide even better responses if it had more data to draw on,” says Nirmal Mukhi, vice president of AI engineering at ASAPP.

For example, lack of access to a user’s PII, such as an address, significantly increased the difficulty of rebooking a ticket for airline customers. Such tasks also required access to backend systems, a synthesis of complex policies, and the ability to reason. “Solving this problem for the most part takes human beings or complex ways of using generative AI to craft a solution,” says Mukhi.

Always looking for ways to be at the forefront of new technology, ASAPP changed from another AI provider to Amazon Bedrock, accessing Anthropic’s Claude models to complement homegrown LLMs. “We recognized that LLMs are a transformative technology,” says Mukhi. “We didn’t want to take incremental steps but rather wanted to rethink how to solve the customer problem.”

ASAPP’s GenerativeAgent Platform operates using a dynamic choreography of five to seven LLMs that perform different functions depending on the use case, latency requirements, and whether the AI agent is responding by voice or chat. “The power of the Claude Sonnet models is that they are a lot more fluid and natural than the prior models we were using,” says Mukhi. “They’re able to conduct a more human-like, delightful conversation, and that has made a difference. We get incredulous responses like, ‘Are you really AI?’ and ‘You’ve been so helpful,’ and that sort of thing.”

Because AWS provides built-in rigorous controls that keep information protected, ASAPP has eliminated its PII redaction layer using Amazon Bedrock. The GenerativeAgent Platform is producing higher-quality outputs and solving up to 40 percent more problems without escalating to a human agent. ASAPP maintains transparency about its use of AI even as it develops the increasingly human-like qualities of GenerativeAgent, all while emphasizing the importance of protecting private information. “We bring AI to the contact center in a way that encourages consumers to trust the product and doesn’t mislead them,” says Mackenzie Smith, head of partnerships and business operations at ASAPP. “For consumers to adjust, we have to provide extreme speed and extreme accuracy.”

The GenerativeAgent Platform automates more than 90 percent of contact center interactions over voice or chat. ASAPP clients have seen a 49 percent increase in their customers’ self-service engagements and a 77 percent reduction in cost per chat.

Through Amazon Bedrock, ASAPP has access to the latest models and tests them using its own evaluation tool alongside built-in features such as Amazon Bedrock Prompt Management, which tests prompts instantly without any deployment. The company also maximizes GenerativeAgent’s available resources and model availability by increasing throughput using cross-Region inference, which automatically selects the optimal AWS Region within a customer’s geography to process inference requests. Users of GenerativeAgent see faster response times, higher uptime, and smarter, more consistent responses.

As it continues to make the voice experience more fluid and natural, ASAPP is working alongside AWS to test Amazon Nova Sonic, a state-of-the-art speech-to-speech model that delivers near real-time, human-like voice conversations with industry-leading price performance and low latency. “The balance of brand recognition and consumer trust in AWS and Amazon Bedrock increases the likelihood that people are willing to take the leap on experimental technology,” says Smith. “That’s been very important for us from a commercial perspective.”

Meanwhile, ASAPP’s customers are able to reimagine their customer service experience and improve interactions with users, resolving 91 percent of issues on the first call. Human agents can handle up to triple the number of complex interactions simultaneously because GenerativeAgent handles repetitive tasks.

“Every bit of the product benefit is directly related to our ability to use powerful models,” says Mukhi. “And Amazon Bedrock is core to how we do that.”