Commentary: How AI and machine learning improve supply chain visibility, shipping insurance

Historically, the insurance underwriting process has relied on static data that is updated only periodically. Such updates are based on deterministic assumptions about the associated uncertainties and risks for which insurance is being sought.

In a May article, How to Adapt to a VUCA+V World, Frederik Bisbjerg conveniently creates and summarizes an easy-to-digest framework that insurers and the companies that rely on them for shipping insurance can use to think about how to adapt to a world that is becoming increasingly Volatile, Uncertain, Complex and Ambiguous. He calls it VUCA+V. It is pictured in the diagram below, which is sourced from the article.

Bisbjerg suggests that beneficial cargo owners and their insurers need to strive for three things in a new normal that is characterized by VUCA. First, they must acquire strategic savviness to create a Vision for the future. Second, they must adopt and implement artificial intelligence to help them gain Understanding and Clarity within their business operations. Third, they must strive for the organizational Agility that is required to function in a Virtual environment that requires Valor.

Insights from the field — PAXAFE combines insurance, intelligence and IoT sensors

Ilya Preston is CEO and co-founder of Milwaukee-based PAXAFE, a seed-stage startup that develops an AI-enabled IoT platform that classifies and contextualizes data to de-risk B2B shipments in the supply chain, better predicts adverse events, and enables dynamic cargo insurance.

I asked him to tell me about the problem PAXAFE solves for its customers.

“The global supply chain is not very safe for products — there’s lots of product theft, counterfeiting, damage and loss that occurs throughout the shipment lifecycle,” he said. “A lack of real-time, product-level visibility compounds these issues and adds additional operational inefficiencies into the mix. Not knowing the location or precise condition of a shipment makes it difficult to drive accountability and enforce SLAs with supply chain partners, and also makes it difficult to optimize for various warehousing KPIs like inventory, buffer stock and labor.”

But, according to Preston, the problem does not end there: “Going one step deeper, it’s not just the shippers that suffer from this lack of visibility and data from the supply chain. It’s also the cargo insurers. You don’t have to look far past the industry’s loss ratios over the past five years to understand that it’s incredibly difficult to accurately price shipment risk. And that shouldn’t be a surprise.”

Explaining further, he says, “After an adverse event, the insurer is going to approach the shipper and the receiver [of the products], the logistics carriers, 3PLs, distributors, and any other stakeholders from the value chain to collect each stakeholder’s own, qualitative version of events [not backed by data]. From there, the insurer’s claims adjusters are going to try to make sense of all of this collected information, and make a determination as to what they believe happened, and who will be held financially responsible.”

Preston points out that “the outputs of these qualitative ‘root cause’ investigations form the basis for future underwriting policies. Policies that are static in nature, updated [at best] quarterly and [at worst] annually, rather than dynamically based on shifting risk parameters in real time. This is an even messier process when an insurance broker is involved.”

But that is not all.

According to Preston, “The asset tracking and visibility platform market is massive [$18 billion], but it’s extremely fragmented. There’s no set data standard in this market, and as such, insurers find it difficult to trust the accuracy, consistency and reliability of IoT data.”

He says, “Existing providers are doing a great job of aggregating more and more raw data sources into their platforms, but that data is not actionable for the people that are actually doing the shipping, or the underwriters who make their living off of accurately pricing the risk associated with each shipment. The data is not contextual. And if it’s not contextual, it’s very difficult to form accurate and consistent prediction algorithms to understand when, where and under which conditions future excursions are likely to occur. And if you can’t predict accurately, you can’t prescribe to customers how to best avoid specific risks.”

Understanding the customer landscape

I asked Preston, “Who is the typical PAXAFE customer?”

“Our primary target customers are the intermediaries between shippers and receivers: logistics carriers, 3PLs, freight forwarders, brokers, TMS platforms and outsourced logistics providers,” he said. 

“This group is typically seeking operational efficiency improvements — improving on-time deliveries and service levels while decreasing fees and penalties. We’ll also work directly with original equipment manufacturers (OEMs) across various target industries like health care, food, manufacturing, insurance and technology. These groups benefit from a supply-chain-of-custody that enables the enforcement of service level agreements, enhances inbound estimated-time-of-arrival accuracy, and enables optimization of inventory and warehouse labor.”

Lastly, he says, “We also work with insurance carriers, where our solution focuses on reducing fraudulent and real claims payouts, digitizing and automating portions of the claims workflow and enabling underwriting commensurate to actual shipment risk parameters via our Adaptive Insurance risk engine.”

What’s under the hood of PAXAFE’s Adaptive Insurance risk engine?

I asked Preston to tell me a bit about the secret sauce that makes PAXAFE successful at solving its customers’ problems.

“Our CONTXT platform uses AI and machine learning to classify and contextualize supply chain data to de-risk B2B shipments and enable dynamic cargo insurance. This platform provides a digital footprint for the physical supply chain capable of answering what, when, where, how and why an adverse event took place.” 

Preston emphasizes that “the how and why here is critical, and it’s also our core differentiator. All the other answers are commodities, but understanding how and why something happened is integral to getting cross-stakeholder buy-in (amongst shippers, carriers, receivers and insurers) and is the first step in establishing a data standard for the visibility provider industry.”

Further, he said, “PAXAFE believes that data context is vital to establishing an accurate, consistent and sustainable prediction engine.” 

To help me understand why this matters, he added that “Understanding how and why an excursion happens allows all parties to accurately identify a root cause, and put controls in place to minimize the likelihood of repetition.

“It also enables visibility providers to prescribe how to minimize shipment risk to their partners. For example, if a product comes in damaged, the shippers and insurers don’t gain anything by reading a shock sensor that tells them when and where this particular shipment experienced a shock — they can’t draw any actionable conclusions. Instead, they need to understand the force of each shock, and delineate whether that damaged product was the result of a package drop, a package that tilted over in response to a truck swerve or pothole, an abrupt forklift pickup, a courier mishandling the package, etc. Same thing for a perishable product that comes in spoiled: Was this the result of human error in not properly following a packaging standard operating procedure (SOP), the reefer being off during an unplanned stop, the package being positioned too far from the vent or poor air circulation, the package sitting on the tarmac or the receiving dock for too long, etc.”

Explaining more about the AI and machine learning that PAXAFE employs, PAXAFE co-founder and COO Ashok Seetharam said, “Deep learning lies at the core of many PAXAFE algorithms. Our entire data contextualization platform is built using deep learning architecture that evolves and improves its accuracy as we collect more data.”

He continued, “To give an example, we have an ML model to detect the change in mode of transportation (road, air, ocean, rail) for any intermodal shipments.”

He added, “The easiest approach would have been to just look at the travel time between two points and calculate the speed to make a prediction if it was air vs. ground travel. But that does not work for scenarios where the GPS tracker fails to communicate due to lack of cell service or when it’s on a flight or traveling over the ocean, or if that average speed comes out to be 100 mph — a bit too fast for road travel, but too slow for a flight. To solve for these obscurities, we had to leverage both spatial and temporal data to train our model to accurately predict the change in mode of transport even when there is limited data coming from the hardware.

“Our environmental sensors record a variety of measurements like temperature and humidity, pressure, location, tilt, shock, G-force, acceleration, light, etc. While these readings provide limited insight in and of themselves, a layer of contextualization gives us actionable insight at a much more granular level — all thanks to the proprietary DL algorithms we use,” he said.

I asked how PAXAFE handles the issues around a lack of high quality data for AI and machine learning systems. 

“A machine learning model is as good as the data it is trained on,” he said. “The number of data sources that a typical supply chain journey sees, coupled with the inconsistency in data measurement and lack of ‘labeled data’ for different events and scenarios, can be a data scientist’s worst nightmare. PAXAFE tackles this predicament in two ways.

“First, by using creative filters to improve imperfect data using a supervised learning approach instead of guessing whether the incoming data is meaningful vs. noise, we trained a machine learning classifier to detect outliers and, as such, are able to predict the difference between useful data generated as a result of real events vs. artifacts from noise.” 

Seetharam added, “Second, there’s a lack of labeled data for accurate classification and prediction. It’s impossible to have customers or users label every event manually when the shipment is in transit. We had to come up with creative approaches and experiments to initially classify events and generate labeled data using unsupervised ML models in conjunction with third-party data and enterprise data.”

“The generated label data is then used to train our prediction engine,” he said.

Giving an example of what this means, he said, “One such example would be our predictive route and dynamic ETA feature using just the scan information from the large carriers, without our hardware. Scan information provided by carriers have several problems — missed scans are often not recorded at regular intervals and they lack context. The first problem is tackled by having a data cleansing module built into the very core of our platform which handles any missed scans, consolidates duplicate scans and also cleans the rest of the data to a high standard. The latter problem is tackled by our intelligence platform to which clean, high-quality data is supplied after cleansing.”

While I do not have enough details to know if my assumption is true, it appears to me that PAXAFE is developing automated approaches to scrubbing the data it needs, similar to what we learned Transmetrics does in Commentary: Bulgaria’s Transmetrics uses augmented intelligence to help customers. 

PAXAFE’S early customers

Finally, I asked Preston about PAXAFE’s progress in winning customers — aware that PAXAFE is a seed-stage startup.

He said, “We’ve partnered with pilot customers from a number of industries, including health care, logistics, food, technology, manufacturing and insurance. We decided to go broad rather than deep to test out the interoperability and true universal application of our platform and determine where the value is the strongest. While we can’t yet publicly share details regarding our partnerships, we expect to start documenting partner case studies in 2021.”

Conclusion

It will be interesting to observe how PAXAFE and other startups developing AI products and systems for the insurance industry progress. 

In an article titled Artificial Intelligence, which was last updated on March 27, the Center for Insurance Policy and Research of the National Association of Insurance Commissioners says: “The insurance industry has only begun its venture into AI, with many traditional insurers experimenting with new ways to incorporate it into their day-to-day operations in anticipation of further technological development. InsurTech startups are also utilizing AI to develop solutions to streamline operations, create better underwriting models and enhance customer service. However, while AI provides opportunities for traditional insurers to modernize themselves, implementing AI is not straightforward. Traditional insurers could face challenges integrating AI into their existing technology due to issues such as data quality, privacy and infrastructure compatibility.”

If you are a team working on innovations that you believe have the potential to significantly refashion global supply chains, we’d love to tell your story in FreightWaves. I am easy to reach on LinkedIn and Twitter. Alternatively, you can reach out to any member of the editorial team at FreightWaves at media@freightwaves.com.

Dig deeper into the #AIinSupplyChain Series with FreightWaves.

• Commentary: Optimal Dynamics — the decision layer of logistics?
• Commentary: Combine optimization, machine learning and simulation to move freight
• Commentary: SmartHop brings AI to owner-operators and brokers
• Commentary: Optimizing a truck fleet using artificial intelligence
• Commentary: FleetOps tries to solve data fragmentation issues in trucking
• Commentary: Bulgaria’s Transmetrics uses augmented intelligence to help customers
• Commentary: Applying AI to decision-making in shipping and commodities markets
• Commentary: The enabling technologies for the factories of the future
• Commentary: The enabling technologies for the networks of the future
• Commentary: Understanding the data issues that slow adoption of industrial AI

Author’s disclosure: I am not an investor in any early stage startups mentioned in this article, either personally or through REFASHIOND Ventures. I have no other financial relationship with any entities mentioned in this article.