Bringing Complex Conversational AI to Production with RPA

March 17, 2022

At the iMerit ML Data Ops summit, iMerit CRO Jeff Mills spoke with Ankit Jain, Co-founder & CEO of Infinitus Systems, about conversational AI and the process of integrating it into robotic processing automation (RPA) of enterprise clients.

The discussion centered on using AI for automation, briefly touching on the need for such technology, the challenges in building it, and the impact on people.

Automating Business Calls in the Healthcare Sector

Infinitus has spent two and half years building the AI digital assistant Eva, with the goal of reducing the waiting time and cost in healthcare. Eva is capable of having a phone conversation with representatives of patient information services, a call humans regularly make to clarify benefits and insurance processes. Eva can provide in a timely manner the correct information to authenticate a patient’s identity. These structured and predictable parts of the conversation are low hanging fruit for AI-based automation technology.

The Automation Opportunity

“There’s something like three and a half billion phone calls that are there between businesses just to exchange data, just to initiate processes, check the status of processes.”

Ankit Jain

As most support phone calls only have limited use cases – such as exchanging data – and following a predictable structure, we can reframe the whole interactions from a human conversation to a series of API calls. This can help drive the sought-after interoperability required in the healthcare industry.

With a strong proof of concept and a definitive market need, Infinitus secured $30 million funding from Google Venture to bring their solution into production. The challenge is crossing the chasm between R&D and production, especially in a highly regulated industry such as healthcare.


To do so, the first step is building a prototype. In the case of Eva, the proof of concept phase was looking to validate if machines could hold long conversations. Chatbots and voice assistants are prevalent across many industries today. However, the average phone call duration with institutions like hospitals is too large (35 minutes) to simply plug in existing automation. Here is where we require a bespoke AI voice assistant capable of complex and lengthy dialogue.

“All of us get phone calls from machines on our cell phones, and what’s the thing we do as soon as we hear that machine that says ‘We’ve got a new solar panel for you’? Well, we just hang up, right?”

Ankit Jain

Besides the technical challenges, AI projects also face resistance from an adoption perspective. Most people already have preconceived notions about automated phone calls, immediately being labeled and treated as spam. 

Upon surpassing technological and social challenges, the next step is getting all the security and privacy processes in place. To venture into the healthcare industry and deal with protected health information, companies must comply with the Health Insurance Portability and Accountability Act (HIPAA). Infinitus navigated these and other such regulatory checklists, now providing their AI solution to Fortune 10 and Fortune 100 companies.

Before we even get to some saying, “Can you show us a demonstration?”, they say, “Do you guys have the following regulatory or security and compliance checklists in place?”

Ankit Jain

Specialized AI for Specialized Use Cases

Most representations of AI in the media and popular culture refer to generalized artificial intelligence – a solution that can do anything for everything.

At this stage in the industry, in a real world scenario, an AI solution is more efficient and capable as its scope is better defined. That was a critical part of Infinitus’s success to data – the ability to bound the problem. For Eva, the initial scope is automating phone calls in healthcare. Within healthcare, it only focused on those business calls with clear patient authentication procedures that a machine could replicate. The calls would also need to conform to specific purposes for which they could build APIs. It should also be possible to express the conversation in terms of the data inputs and outputs of the APIs. All this ability to define the product and solution helped Infiintus release a market-ready product. 


From an initial working product adopted by leading players in the industry, Infinitus can now expand their feature set to also include patient facing calls.

The Helping Humans

As with all AI solutions, behind the neat solution that automates phone calls sit staff with years of experience making and receiving phone calls. These trained practitioners had to wear multiple hats. First, to label the data which feeds the AI models. Second, as manual testers who performed quality assurance (QA) on the developed products and services. Third, act as the safety driver – the humans-in-the-loop –  in instances where Eva, the AI assistant, came across situations that were not seen or not automated yet. The human operators would step in and take over the call to ensure the customers were satisfied. These types of new setting or edge cases would then be documented and fed back into the system to include these new use cases.

Impact on the Job Market

Fears over job security follow any news of automation, and in the instance where human phone operators are replaced with AI operators, the initial reaction from the public is resistance. However, even as Eva was deployed and scaled up in the past year, no clients had made any job cuts. It turned out to increase employee satisfaction by reducing the time spent on routine back-office calls. 

“It unlocked employees of our clients to do things that they wished they were doing.”

Ankit Jain

With Eva staying on hold and making requests for patient information or prior authorization status checks, healthcare practitioners like nurses were freed up and spent time on patient care.

If you wish to learn more about creating datasets for Machine Learning, please contact us to talk to an expert.