Artificial intelligence (AI) has slipped into the global lexicon without a common understanding of what it means. Often, AI is shorthand for any kind of automation. Even though AI has shown promise in the areas of facial recognition and advertising, it has been disappointing in areas that matter for large companies-- e.g., high-impact decision support, time series prediction, chatbots, contracting, and anticipating client requests-- these are a few areas where attempts to implement AI have mostly failed. Even so, in the last few years, full automation of particular processes—called robotic process automation (RPA)—has been gaining ground in many organizations. Simply put, RPA replaces an existing process (or may build on an existing process) to replace human labor. Many fear automation will replace most workers & employees. RPA is one strand of many current AI discussions.
Other popular AI discussions explore how to simulate human cognitive processes. These discussions lead to concerns that anthropromorphized AIs will gain consciousness (however defined) and become an artificial general intelligence (AGI). Hollywood has anchored its AI-related imagination on these AGIs such as the ruthless Skynet (Terminator), evolving Samantha (Her), or a hive mind like the Borg (Star Trek). These movies influence how many people visualize AI. Unfortunately, these highly unlikely AGI outcomes are not directly relevant to organizations wrestling with automation today. Both RPA & AGI deter discussions away from four AIs we don't talk about enough: 1) Augmented intelligence; 2) Intelligent automation (flip the "I" and the "A" to fit in with this list); 3) Assessed intelligence; and 4) Adaptive intelligence. Let's discuss why these four AIs should be dominating our AI-related discussions.
Augmented intelligence: AI to boost human productivity
While computers are phenomenally good at executing billions of calculations and can reproduce data without degradation (assuming systems are maintained), they still do a terrible job of navigating complex, ambiguous circumstances requiring creativity & the mixing of tacit and explicit data & information. For example, a physician could become much more productive using machine-learning-enabled systems to assist diagnosis rather than making the diagnosis. Emerging research indicates that while artificial intelligence by itself typically matches or outperforms a physician, a physician plus augmented intelligence does better than the machine alone given that the leveraged human outperforms on the subset of diagnoses where the circumstances are unusual or unprecedented. Swiss Re has augmented its underwriters by implementing algorithms that machine read contracts to identify language that may be a concern. Thus, an underwriter can focus on high-concern areas. While this technology still needs development, early implementation demonstrates how underwriters' productivity can be improved. In another application, insurance claims adjusters could use automated systems to determine which claims to scrutinize more thoroughly rather than replacing the entire process with machines. The point is that better & cheaper processes can be engineered by supporting humans with augmented intelligence rather than replacing them entirely with artificial intelligence. Not only can performance be improved, the downside risk of unanticipated scenarios is better addressed by an AI-supported human decision maker.
Intelligent automation (technically IA; however, I will call this another "AI"): Keeping humans in the loop
Even though RPA will continue to insinuate itself into our companies & governments, we should collectively resist full-blown automation and insist on keeping humans in the loop. Returning to claims processing, full RPA will likely lead to higher frequencies of fraud as the machines are "tricked" by clever hackers. Instead, relevant processes should be distilled into a collection of sub-workflows where process engineers can include humans in ways that keeps processes robust, reduces fraud, and assigns tasks that match the strengths of AIs and humans, respectively. We already have some notions of how to divide process responsibilities among AIs and humans.
AIs excel in the following areas: Computation, Consistency, Ubiquity, Details, Networkability (at scale)
Humans excel in the following areas: Adaptivity, Autonomy, Tacit/subtle understanding, Imagination (particularly for troubleshooting), Emotional/social understanding, Robustness to system failure (can figure out workarounds)
A key takeaway from this comparison is humans should be integrated into vulnerable and/or fragile points in a system or process. That is, without proper integration of humans into a process loop it is likely that more money will be lost in times of failure than saved in times of normal processing. Many processes within a typical insurance company will benefit from a deeper dive into how to intelligently automate:
1) Marketing and sales
2) Policy administration
3) Claims processing
4) Capital allocation
5) Asset/liability management
These processes are not the only ones that could benefit; however, they are the ones that seem ripe for re-engineering from the perspective of materially transforming the insurance industry.
Assessed intelligence: Continuous evaluation is the primary path to improvement
Surprisingly, many systems, processes, models, and software components are implemented with assessment done only at time of implementation & infrequent inspection. Often assessment happens after a system failure—wildly increasing costs. Augmented intelligence in an intelligently automated organization can be used to lighten the burden of frequent and regular assessment of the increasing mix of software/systems (defined broadly) that accrete within our society. These assessments should focus not only on performance (suitably defined), but also computational cost, latency, robustness, and ease of diagnosability. This last point of explaining/interpreting the underlying processes used for decision support is often dismissed when some "black box" algorithms (like deep learning) perform so well on specific tasks. Unfortunately, performance often trumps prudence. This said, mode/process/system explainability & interpretability are criteria we should not forget. Mixing assessment systems and humans will likely lead to discontinuous improvement in any software/system used in complex organizations and provide the side benefit of increased system resilience & robustness.
Adaptive intelligence: Learning how to learn
This fourth AI is still mostly in the future. Many human advantages relative to current AI algorithms result from adaptivity. Some of the newer machine learning approaches (e.g., meta-learning, ensemble modeling, data augmentation, certain types of reinforcement learning, etc.) are leading the AI industry toward adaptive intelligence. This said, these adaptive AIs are still more hope than reality. Even so, this development path may lead to much better AI tools. An interesting approach in this context combines a subject-matter expert's characterization of a problem, say, the impact of an earthquake on a city's infrastructure with actual data from real earthquakes. The Swiss Re Institute is working with one of its technology partners and one of its clients to determine how this adaptive machine learning could change the way non-damage business interruption insurance is priced, sold, and managed. In this case, an adaptive intelligence toolset could augment existing underwriting processes in a way that shows promise as an effective method for reducing the insurance protection gap in an area where most corporations & governments are underinsured. The difficulty with using typical AI (e.g., deep learning) in this context (i.e., how catastrophic perils create widespread business interruption) arises from the small number of data points available to train a model. This new adaptive approach that merges subject-matter-expert modeling with machine learning addresses this data problem with simulation—eliminating the need for large training datasets. This could lead to augmented-intelligence systems that adapt based on human input and renewed simulation. Again, the engineered interaction of humans & machines improves system outcomes. The result is likely to be better insurance underwriting and further reduction of the insurance protection gap. Adaptive intelligence is the new AI frontier.
These four AIs require more engineering, conversations, empirical research, and experiments. Companies & governments wrestling with these new technologies should focus on these four AIs instead of hosting frequent discussions that focus on the extremes of AI-enabled utopias or AI-triggered apocalypses. I predict neither extreme will come to pass. Rather, we will see flashes of brilliance in organizations who figure out how to unlock the synergies of humans supported—not replaced-- by machines. Everyone else will continue to fail to transform their organizations with these promising tools.
-- Dr. Jeffrey R. Bohn, Chief Research & Innovation Officer, Swiss Re Institute
Opmerkingen