Why do we need it?
While the world of Generative AI has become quite interesting, for good & bad reasons (read the OpenAI saga that unfolded last week for more context), companies, both big & small, are still struggling to adopt GenAI fruitfully. While there are a lot of initiatives being run, including pockets of pilots, MVPs and independent POCs, GenAI bringing a value addition to businesses is still a far cry. Well, the reason cannot be that the CXOs do not understand this technology yet. In fact, the GPTs & Co-pilots of the world have ensured that the barriers to entry in the tech space are lowered than ever before. Nor is the problem in understanding the business needs, these are smart guys who run the business. While we will try to speak about our experience of implementing successful initiatives in this space, a question to ponder for everyone reading this, why do we need technology? Is it because it is available, easy to implement, or I have a lot of people who know tech and I want to give them jobs? Or it solves an actual problem, makes life easier, brings innovation to what we do and eventually help build a moat for the business? Put a pin to that thought.
A few weeks back we were in a discussion with a CIO who has been successfully managing the Motor Insurance portfolio for the last 5 years, has digitized a lot of workflows, increased the STP rate for claims by over 25% with intelligent automation, so all in all, has good credentials. 15 mins into the call & he is impressed with our capabilities, jumps into what he wants. His next left me speechless. “Rishabh, I have a KRA set for my team to bring in GenAI into every customer facing process by the 2nd quarter of 2024, so we are very bullish about adopting GenAI”. While it took me a min to gather my senses, I did ask a follow up question, “That’s great (well, actually it is not, but….), you have the intent, but what is your goal of implementing GenAI across these processes, let’s pick 1 usecase & try to understand your strategy”. The response was from the VP of Customer Experience “We want to get started with our post sales customer support for the Motor Insurance segment, as we have the maximum queries there and want GPT to answer all these questions”. My response “Well that’s great (Seriously it is not, but again….), but the reason for integrating GPT to your post sales support is? Is the CSAT going down with manual support or the costs of operations are high, or you want to better the customer experience with faster resolution?”. The VP’s response “No, the costs are pretty optimized, our agents are doing a great job with high customer ratings & low QRTs, but we feel GPT will give us a competitive egde, plus it is a KRA for the team next quarter”. Well, I can go on about this, but you got a gist of where this was heading. While eventually we did do a pilot with the customer to automate their Claims Settlement & Policy Enquiry usecase, but only after due diligence with the customer support team to understand the pain points & how GenAI (& not GPT) can help solve some of these (also because we are a bootstrapped start-up & need to pay salaries).
The above conversation is not a 1 of scenarios, a lot of requirement discussions with 1st time adopters in this area are like this. Taking a technology 1st approach & not a usecase base approach. While this strategy did work for a lot of businesses during the cloud boom, where they exited data centers to avoid huge capex with lift & shift migrations, to realize later that the bills are higher. But eventually good FinOps practices helped them reduce their bills in the long run. But a technology like Generative AI, which is so intrusive to your process’ DNA, that if it does not work, the cost of terminating or migrating from such systems can be catastrophic to the business.
While every technological shift, be it structural, like movement from monolithic to micro-services architecture, or foundational, like the cloud adoption, rides on a wave initially, where voyeurism is a major growth driver, but with Generative AI, the case is quite different. 2 reasons
- 80% of the usecases that can be solved with GenAI are not new, they were being solved earlier by ML powered systems. The way we approach these has changed drastically, not to mention the reduction in effort & much higher accuracy.
- The pace at which the technology is evolving, you do not have a chance to spend months on creating a 10 – year roadmap.
Thus, welcome to the realm of Generative AI, where innovation meets transformation, and both improve parallely. I really don’t want to inundate you with a step-by-step guide, I’m pretty sure you’d have read multiple of those; instead, we’ll embark on a directional journey, illuminating a path to build a Generative AI strategy, usecase 1st. While we have heard of this phase “Build smaller, fail fast, learn faster, scale”, now is the time to actually implement this. Picture it as a recipe, rather than a mere set of instructions.
Identifying the North Star: Use-Case First
Every transformative journey needs a starting point, and for enterprises delving into Generative AI, it begins with identifying the right use case. Let’s delve into a real-world example – a US-based Motor Insurance company navigating the complexities of Claims Settlement.
The Claims Revolution
Imagine a world where claims settlement happens within 24 hours, and 80% of claims undergo a seamless, straight-through processing. This wasn’t a distant dream for our motor insurance partner; it was a strategic vision brought to life through Generative AI, specifically leveraging the prowess of GPT-4.
But how did the journey start? Very clearly, not with a statement “We want to implement GPT for our Claims Settlement process”, it started with a statement “Our claims settlement is lagging indicator, only 22% claims are STP and the time to settle claims is 3X more than the industry average, how do we solve this?”. A process discussion helped understand that the current Claims process has at least 6 human touchpoints, which are workflow based, repetitive assessments. 3 out of these can be removed with entity recognition from automobile images to assess the damage, while 2 were knowledge extraction & comparison steps with the policy document to identify the right claim amount & any edge cases, the final step was sending instructions for claims settlement to the bank or escalating up the ladder for exceptional usecases. All 6 interventions were candidates for automation and that’s when GPT-4 came into picture, as it brought in capabilities to solve 4 out of the 6 steps, with high efficiency, with a combination of Finetuning on RAG, with the user’s policy context.
What helped going live in 6 weeks, meeting Claims 80% STP and settling STP claims within 24 hrs?
Strategic Execution: A Step Beyond Implementation
- Understanding the process
Before diving into the technology, understand the nuances of the process, identify the limitations & dependency. Most importantly, are any of the steps workflow based, good candidates for GenAI powered automation. What challenges are currently faced by the team in automating? What pain points can Generative AI alleviate?
- Pilot Prowess
Initiate with a pilot – a manageable yet impactful use case. In our scenario, Claims Settlement was the ideal candidate. Aiming for a 24-hour turnaround was audacious but feasible, because it has been achieved in the industry and once the problem was identified, technology was there to achieve the target.
- Learning from Early Flights
The pilot phase isn’t just about results; it’s about insights. What worked? What could be optimized? Harness these early learnings to fine-tune your strategy. Like in our case, the chat logs were not documented well, early in the pilot we had identified that using these logs will result in low accuracy & increased escalations. Quite fast, we looked for an alternative, which was the SOP documents in the company database.
Building the Framework: From Pilot to Powerhouse
- Scalability Matters
As the claims settlement pilot gained traction, scalability emerged as a critical factor. With more automation, we were now building additional data at breakneck pace to power GPT-4 to improve accuracy & expand usecases.
- Framework Expansion
Extend the success of your pilot into a robust framework. Scale your Generative AI capabilities, incorporating the lessons learned and ensuring adaptability for future use cases. The Vector DB built for the Claims Settlement usecase was now leveraged to automate the Claims enquiry usecase in less than a week.
The Road Ahead: Future-Ready Generative AI
While the 1st usecase was tough, it did help the organization realize the potential of the technology. But keep in mind, while the outcome was achieved due to the tech capabilities, you will be enticed to add usecases immediately onto the GPT bandwagon, the approach to replicate is, process & usecase 1st. The framework is ready (& will continue to evolve), use it well!
I will now go back to conundrum I had narrated at the beginning of this blog, why do we need technology?
