The five-step approach to successful AI deployment

Deploying Artificial Intelligence into their operational workflow still feels like a leap into the unknown for many insurers. But in fact, the project methodologies and technical solutions are in place to deploy AI safely, step by step, with business outcomes very much in focus.

Claims processing keeps coming up in our conversations with Life and Pensions firms as an area in need of radical optimisation. Current processes, even amongst well-performing organisations are too manual, too slow and too expensive. We can imagine that an insurer may have an objective of being able to process half-a-million claims per administrator, and to resolve 95% of claims within ten days.

Technically, this involves selecting and training combinations of insurance-specific AI algorithms to prepare documents for OCR (Optical Character Recognition) tools to read with the required speed and accuracy, then automating the instantaneous processing of validated documents in the Policy Administration System (PAS). Using Lumera’s experience from working with one such insurer, we can show how a five-step project process can de-risk early forays into the world of AI.

Fact-find workshop

We began our claims processing optimisation project with a fact find workshop. This is an engagement of just a few days, with the aim of establishing the client’s desired business outcomes. A key output of the workshop is an agreed problem statement for the project to resolve.

The workshop will also confirm that we have access to the data and data sources necessary and that our library of techniques and algorithms give us a strong starting point. Using exploratory data analysis, unsupervised AI can search the data for hidden segments or quality issues that can be addressed early in the project.

Discovery Deep Dive

During the second phase, the Discovery Deep Dive, we set up a research environment for experimenting to find the most effective combination of AI techniques. Given our experience, we had a shortlist of algorithms from our source code library to be tested.

In this case, the technical problem was to standardise and cleanse incoming documents so that an OCR tool could read the key data from the documents with sufficient confidence in their accuracy that the process could be fully automated.

Challenges that a human can overcome without thought, such as correcting for tiny misalignments, or recognising where a field of text or numbers begin and ends can take many days to perfect. Removing watermarks or other official stamps, without eliminating the data we needed to validate, is a good example of a task that took many iterations to solve. By the end of the Discovery Deep Dive, the techniques and algorithms we believe will deliver the desired outcomes will be identified.

Proof of Concept

During the Proof of Concept (POC) phase, we scale up the training of the AI algorithms to verify that the business outcomes can be delivered at scale. From a business point of view, the most powerful feature of AI is that confidence and accuracy improve with scale, as the model learns from more examples. Traditional, rules-based algorithms, suffer reduced accuracy as they encounter variations on the sample upon which their rules were based. In our claims process optimisation project we were able to validate death claims within 3 seconds with over 95% confidence. Even documents which the AI model read with only 30% confidence were still read accurately. As the volume of documents grew, the accuracy and confidence levels increased, as we would expect for the AI models.

In the background, AI-Operations were working on optimising the live set-up; how best to expose the AI model to the PAS as a Service via the API layer, the deployment processes, and the optimisation of memory, latency and hardware footprint.

Proof of Value

The aim of the Proof of Value (POV) phase is to validate that AI models will execute as expected in the client’s operational environment. During this phase, AI-Operations come to the fore, working with the client’s IT team to deploy the AI models into the live environment. Some technical tuning is likely required in the face of live operations, but the client will only have authorised the Proof of Value phase if the AI models have shown that they should deliver the desired outcomes during the Proof of Concept. The five-phase project approach gives the client complete control to only proceed once the probability of success has been established and risk of failure minimised at each prior phase.

Transition to Live

The final phase is the transition to live operation and business adoption, setting up the processes to support and manage the ongoing performance of the models. AI-Operations put in place monitoring of end-to-end logs to catch any foreseeable errors within the models and implement error handlers for predictable issues like invalid dates.

AI techniques such as Anomaly Detection (AD) algorithms are deployed to continuously analyse and learn the behaviour of each AI model’s inputs and outputs, and to alert when the AI pipelines deviate from the norm. In addition to the automated monitoring, robust actions are coupled with alerts to accelerate resolution time, which reduces risk and maximises business value.

Lumera’s mission is to enable the Prudent Revolution for the Life and Pensions industry, providing safe pathways to embrace next generation technology. Our five-phase approach for AI projects gives clients step-by-step control to realise their desired business outcomes through embedded AI use cases within their ecosystem. Allowing for periods of testing, client oversight and authorisations, deployment can be completed in five to six months, delivering operational ROI by the end of the final phase.

Take me back