The Power of the AI Learning Effect with Colin Toal, CTO, Chisel AI

Artificial Intelligence, Natural Language Processing - January 29 2020

I recently had the opportunity to sit down with our Chief Technology Officer, Colin Toal, to ask him ten frequently asked questions about machine learning, natural language processing, and the “AI Learning Effect” in commercial insurance. These are questions we get asked quite often by insurance professionals who are looking to implement AI-driven solutions. Colin shared his insights into how the AI learning effect works, human involvement in AI-enabled processes, system requirements, and the value delivered to commercial insurance companies.

Colin, please share with us your definition of the AI Learning Effect.

The AI Learning Effect is a term I learned from one of our board members. It describes a system that becomes more valuable with usage. In the simplest form of machine learning, you start with two "loops". The first loop occurs with customers using the machine learning (ML) based system by asking it to complete some useful task, and in return getting the benefit of that task being done more automatically. The second loop (the most common use case of supervised learning), is a 'supervisor' – a part of the system (usually another person) that identifies the patterns or examples that the system does not know how to handle yet and “teaches” the system the correct way to respond to it. Over some period of usage, the system learns new patterns and capabilities from these loops, and the result is that it becomes more useful and appealing to more customers – and this is when the Learning Effect kicks in.  

Tesla Motors' self-driving features are a great example. Each Tesla car is a data collector – and every time someone drives a Tesla, it sends valuable data about how the car was driven back to Tesla's computing cloud. That data is analyzed and labeled by complex processes that help the algorithms in Tesla's cars drive more accurately. More people want a Tesla because it drives itself, and so more people buy them for that benefit and each new driver adds more training data. Each new Tesla owner gets the benefit of a self-driving feature that has been trained by the millions of miles driven by every previous Tesla owner. That's the AI Learning Effect.

The more breadth and depth of data provided to the machine learning-based system, the more accurate and useful it becomes. Most ML systems are 'supervised' learning systems – which means human knowledge workers identify patterns and correct responses. When the machine misses, the human supervisor can give it a little more information, so it stands a better chance of matching the pattern the next time. Siri, Alexa, Google Assistant, and Tesla's autopilot are all systems that use great features to drive usage and supervised learning to turn that usage into new features, or better performance for existing features.

Now, if we think about AI solutions designed for insurance, which is a very document-intensive industry, we can design a system that gets more useful with the number and variety of documents that are fed into the system. The more policy documents or submissions or binders or quotes that are fed into the system, the more the system becomes familiar with the document format and data. The system is then able to extract, understand and summarize the data patterns, providing more value to the user.

At Chisel, we’re focused on building a Learning Effect system that processes insurance documents, making them machine operable – making it possible for novel applications to be built on top of them that gives the customer value and drives usage. We use supervised learning “teach” the machine how to read and extract data from insurance documents, and automatically execute key processes on these extracted values. We continuously learn from our usage, and our customers. Both insurance brokers and carriers can realize tremendous business value from using an AI platform that is domain specific, understands the industry jargon, and has read millions of pages of documents, much like the Tesla fleet of cars that have driven millions of miles.

Great examples! AI is not for the faint of heart, though. There is a technical nature to it, so do insurance professionals such as underwriters need to be technical geniuses to use a Natural Language Processing or Machine Learning solution?

Insurance is already very technically demanding. The benefits of machine learning can be applied across the insurance value chain, especially during the submission process. Our machine learning based approach allows Chisel AI to meet the industry on their terms and reduce the software technology expertise they need to deploy and use our AI platform. Instead of insurance companies changing their systems or building multiple integrations to APIs to interconnect systems in an attempt to gain business value, we can take existing insurance artifacts such as binders, applications, submissions, policies, etc., and teach a machine to contextually understand the data in these documents, eliminating the need for insurance professionals to manually transcribe, cross-reference, manipulate, and rekey data into multiple core systems. By automating manual, error-prone data touchpoints, insurance professionals are free to focus on more valuable customer-facing initiatives such as providing one-to-one client advice.

You’ve mentioned “the loop” a few times now and that a human knowledge worker needs to be in the loop. Can you provide more details about their responsibility? What is the task they are performing?

Human knowledge workers play a vital role in the AI Learning Effect loop as they receive value and apply value to the model. At the frontend of the AI Learning Effect loop, insurance knowledge workers are the recipients of value from the machine automatically matching the patterns that lead to valuable actions like real-time data extraction, automated, etc. This results in their jobs becoming easier and faster. We also leverage the judgment and experience of human knowledge workers in our 'supervisor' loop – they are responsible for providing the system with the information it needs to learn and figure out the next pattern.

These AI-based systems enable human knowledge workers to focus on the highest value of their profession by spending less time on tedious tasks. Human knowledge workers can augment their role by letting the machine focus on these tasks at the speed of light, freeing them up to spend more time on nurturing customer relationships, and risk assessment work that requires human judgment.

Can you share with our readers a few examples of “judgment” as it relates to commercial insurance processes?

Sound judgment, the wise evaluation of information or the rigorous assessment of risk is key to the livelihood of insurance companies. The success of insurance companies is predicated on effective pricing strategies and human judgment.

Brokers use judgment to inform the needs of a client based on the plans that the client has for their business or helping an underwriter evaluate the risk of potential exposure on their business. Judgment tends to be based on a great deal of context – a variety of data, facts, history, and hindsight gained through experience. This is not something that we are focusing on automating anytime soon.

In contrast, AI solutions like Chisel AI focus on common, repetitive tasks like the manual transcribing of insurance documents, data re-keying, validating policy jargon and coverages, manipulating data from a variety of documents to identify the best quote, etc. By automating these tedious, mind-numbing tasks, brokers, underwriters and others get time back to focus on high-value initiatives that are highly dependent on their knowledge, expertise, and judgment.

What is an AI model?

At a high level, a machine learning model is a useful summary of data. It’s useful in that it can answer questions for us and make predictions with some measure of reasonable accuracy even when it has not seen the data before. The simplest model is a regression model. If I tell you one iPhone costs $1000 and a second iPhone also costs a $1000, then you will have a reasonably accurate guess at what a third iPhone costs, even if I don’t tell you in advance. Machine learning models work with these kinds of summaries of data.

A great example is Google Home or Alexa. Users speak voice commands to interact with services within their home. These smart speakers were built on AI models that can take a representation of your voice, compare it to a statistical summary of the sound representations of lots of words and predict what words you are saying even though the speaker has never heard your unique voice before.

The power of a model is predicated on the breadth and depth of the data it is exposed to. For commercial insurance, an AI model uses multiples data sources extrapolated from multiple documents such as policies, binders, quotes, submissions, applications, statements of value, loss run reports, vehicle fleet summaries, property histories, etc. The data sources can be a block of text or a set of words from a document. Using the model, we can assert with reasonable certainty that a data point is a limit of liability value or named insured value based on the corpus of data representing document types, carriers, lines of business that have been summarized in a model.

At the heart of all artificial intelligence (AI) solutions are these models. The more data that the model summarizes and the more fidelity in how it is summarized, the better the accuracy.

How much and what type of insurance data is required to train a model?

The short answer is a lot of data is needed to train AI models. I know that sounds daunting and is not qualitative, so let me explain in more detail what I mean by “a lot.”

At Chisel AI, our models are built on a large corpus of millions of pages of data pooled from a set of commercial insurance carriers and brokers. Like Tesla, we have run the miles with our fleet. We are continuing to develop models that can usefully answer questions about the specifics of insurance. This enables us to leverage the model to automatically read an insurance policy and identify key data points such as the limits of liability in the policy, for example, with reasonable accuracy.

Colin Toal Blog Quote #1
Again, I can’t stress enough that the key to effectively training an AI model is the breadth and depth of the data. Operationally, for insurance companies considering AI or kicking off an AI pilot, hundreds to thousands of examples of each document type (e.g. binders, submissions, policies, endorsements, etc.) are required to effectively teach the machine. The more data available, the better, as it will accelerate the machines’ capacity to contextually understand and summarize the data patterns.

If we take a step back to the “loop”, is it critical for a knowledge worker to always be in the loop or at some point is a human no longer required?

Most effective machine learning processes today are humanly supervised. Advances are coming in places like reinforcement learning – but they are not close to being commercialized, and they don't address the very hard problems of context, experience, judgment, and bias.

At its heart, insurance is a human business about providing safety to other humans in the face of risk. We will always need people exercising that judgment and working on that highest value proposition of the business. We are just trying to make it easier for them to do this.

Can the AI Learning Effect be deployed across the insurance value chain? Who stands to benefit the most from AI-driven technology?

Yes. Insurance carriers and brokers who invest in and deploy AI-driven technology will gain the operational efficiencies to boost net written premiums without adding headcount. By deploying digital workers or software robots, they can automate tasks throughout the process including data extraction and classification, data entry, policy checking, auto-routing, and auto-declining submissions, freeing up underwriting and administrative staff to focus on risk assessment, pricing, and client advice.

Think of it this way – in the forestry business, an average lumberjack with an average chainsaw will cut way more wood than the best lumberjack with the best ax. Giving human knowledge workers AI-powered tools to automate and streamline high-volume, repetitive tasks will enable them to do more – they will be able to consult with more customers, win more business and deliver a better customer experience.

Ultimately, we want all our broker and carrier customers to be able to offer their customers a faster, more consistent insurance buying process.

In your opinion, what is the risk of not deploying an AI strategy?

First, the risk of not pursuing and adopting AI is being outplayed by the competition, losing market share or, worse, going out of business. Don’t get me wrong, there will always be a place in the market for specialty and niche insurance players, but for mainstream commercial insurance providers, competitive advantage will vaporize if they don’t consider the business benefits of AI.

Second, employment enrolment in insurance is eroding as the baby boomers retire. This is having an enormous impact on the industry as insurance companies compete and vie for new talent. To use my forestry analogy above, there are fewer great new lumberjacks who want to work with axes, most want to use a chainsaw! Innovation in the form of technologies like AI is attractive to the next-generation insurance worker and can help insurance companies attract the best talent and compete more vigorously.

Thirdly, AI gives insurance companies the ability to optimize operational efficiencies, drive faster processing through automation. It delivers speed and consistency by eliminating error-prone tasks like manual data re-keying of information in multiple systems which equals reduced errors and omissions and writing policies more quickly and safely.

As a seasoned business professional who is passionate about technology, what words of AI wisdom would you like to share with our readers?

If you take all the press, hype, doom, and gloom, from every angle, the main thing to understand is that AI is all simply about summarizing data and matching patterns. Yes, this might seem like I am trivializing it, but given the deluge of data being produced today, AI is enabling insurance companies to put the data to work to better serve their customers.

We are experiencing an “important moment” when computing power, technique and opportunity are coming together to make it possible for commercial insurance companies and brokers to fully embrace digital transformation.

Colin, thanks for sharing your insights. Now, how can our readers contact you and can you share with them what you like to do in your downtime when you’re not thinking about AI?

Sure – you can reach me on LinkedIn or drop me a line at  

I spend my downtime thinking about AI, insurance, and software! Apart from that, I try to spend at least a couple of hours each day reading. I usually have two or three books in progress and a backlog of Medium articles. I just finished The Score Takes Care of Itself (for the 2nd time), and I am reading Trillion Dollar Coach and the Amazon Shareholder Letters.

I spend the rest of my time as a faithful sidekick to my superhero wife and daughter. We recently saw Frozen II in the theatres and "binge-watched" Star Vs. The Forces of Evil on Disney+. Both are highly recommended and have important lessons in them for adults as much as kids!

Listen to the Podcast: The Crucial Role of Data in Your AI Strategy with Colin Toal

Tune in to the Chisel AI podcast AI Wisdom – Talking Innovation in Insurance Ep. 6 to hear Colin's conversation with host Ron Glozman.

Podcast with Colin Toal LinkedIn


Browse different topics

Recent Posts