About this Guide

Welcome to The Business of AI & Bots Playbook. This site serves as a go-to resource for business leaders to brush up on their understanding of where AI, machine learning, and bots are used across industries. With brief introductions to the different types of intelligent technologies available on the market along with a number of their associated use cases, industry professionals can refer to this to quickly grasp the breadth and depth of the AI landscape today.

Use this as a pocket guide for quick references to the AI landscape today, complete with industry-wide use cases and high level introductions into the world of machine learning and the artificial intelligence ecosystem.

With each description of the various types of artificial intelligence below, you can jump to different industries and applications for the given use case.

Computer Vision

Computer vision is concerned with the automatic extraction, analysis, and understanding of useful information from a single image or a sequence of images. Image recognition is the science of identifying objects from a given image as a strict subset of computer vision.

Image recognition identifies objects within an image and classifies those objects, whereas computer vision also encompasses functionality such as style transfer, facial recognition, and anomaly detection.

Computer vision and image recognition live at the intersection of Machine Learning, Linear Algebra, and Image Processing. These forms of machine learning have been used for applications in a large number of domains. Depending on the use case, different paradigms, such as machine learning and neural networks, have been used to implement Computer Vision.

This is an area that in itself and has been in existence since the 60s when it was meant to be an alternative to human vision as a system integrated into robotics. The idea here is to treat an image as a data source and annotate various aspects of the image that can be used for learning. The annotation can be unsupervised (eg. clustering images) or supervised (e.g. classifying images of cats vs. dogs, Facebook image tagging suggestion, etc.)

By using image processing and object detection, image recognition and computer vision have been successfully implemented in a wide array of use cases ranging from fraud detection to oncological testing.

Where is this used?

Jump to an Industry: Agriculture, Construction, Manufacturing, Sports, Transportation

ANNs & Deep Learning

Artificial Neural Networks (ANNs) are computing systems that are inspired by the layout of neurons in the brain. Neural networks have a long history, but only in recent times have they seen widespread adoption and success. This was mainly driven by the improvements in algorithmic approaches, computational speed, and the availability of large amounts of data.

Neural networks is one algorithmic approach to machine learning as opposed to more traditional algorithms. Most commonly, ANNs are used in classification tasks where ANNs promise increased performance over traditional machine-learning algorithms. But this typically comes at the cost of additional computational complexity, a need for large amounts of training data, lack of interpretability, and general “brittleness” (i.e. the ANN is very good at one task but fails or breaks if “out-of-the-ordinary” data is received).

For every given task that benefits from machine learning, it is important to select the right algorithm. One major difference between neural networks and other machine learning algorithms is the amount of “expertise” that is required before the model is set. Feature engineering involves manually manipulating the data for the best model — or providing that “expertise” manually.

For example, classical machine learning requires a human to input example characteristics of what cancer might look like in an MRI scan. This instructs the algorithm what to look for to determine if cancer is present, and can only be as good as the input data. The identification and engineering of these features can be time-consuming and typically requires significant domain expertise.

Neural networks have the ability to discover these features internally — i.e. automatically surface features that would find cancer without as explicit of inputs required. However, it requires far more data and processing power to be able to do so.

This foregoes the need for time-consuming feature engineering and allows the neural net to discover those features that are most informative for the task at hand. This may include higher-order or more abstract features that would otherwise be difficult or impossible to articulate.

To go one step further, ANNs can have one or more hidden layers. ANNs with two or more hidden layers are typically called Deep Neural Nets (DNNs), also known as deep learning. Deep learning is equipped to learn from what is referred to as “noisy” data — by sorting out what matters and is most relevant. Deep learning algorithms are algorithms that automatically engineer the features that are otherwise reliant on manual input in traditional algorithms.

While neural networks seemingly provide a short cut around the labor-intensive feature engineering, they also demand specific attention and care as they obscure their decision process. This is part of what makes neural networks hard to interpret or troubleshoot, and requires the machine-learning practitioner to pay close attention to the network’s performance.

Where is this used?

Jump to an Industry: Agriculture, Finance/Banking, Healthcare, Transportation

Image Recognition

Computer vision is concerned with the automatic extraction, analysis, and understanding of useful information from a single image or a sequence of images. Image recognition is the science of identifying objects from a given image as a strict subset of computer vision.

Image recognition identifies objects within an image and classifies those objects, whereas computer vision also encompasses functionality such as style transfer, facial recognition, and anomaly detection.

Computer vision and image recognition live at the intersection of Machine Learning, Linear Algebra, and Image Processing. These forms of machine learning have been used for applications in a large number of domains. Depending on the use case, different paradigms, such as machine learning and neural networks, have been used to implement Computer Vision.

This is an area that in itself and has been in existence since the 60s when it was meant to be an alternative to human vision as a system integrated into robotics. The idea here is to treat an image as a data source and annotate various aspects of the image that can be used for learning. The annotation can be unsupervised (eg. clustering images) or supervised (e.g. classifying images of cats vs. dogs, Facebook image tagging suggestion, etc.)

By using image processing and object detection, image recognition and computer vision have been successfully implemented in a wide array of use cases ranging from fraud detection to oncological testing.

Where is this used?

Jump to an Industry: Agriculture, Manufacturing, Healthcare

Machine Learning

Machine Learning is the concept of computers being able to learn how to perform a task without being explicitly programmed to do so.

Traditionally, automation was achieved by explicitly programming machines. This entailed telling the machine what to do given a particular scenario. Any scenarios that had not been specified would not trigger an action from the machine, or lead to undesired or unexpected responses.

Machine learning foregoes the need for explicitness and allows machines to become “smart” by learning expected responses based on the data itself. This is in particular true for supervised learning, the most common form of machine learning, where the machine learns expected responses based on examples. For unsupervised learning, the use of sophisticated algorithms allows the machine to automatically discover structure in data from the data set itself.

At the heart of any machine-learning model is an algorithm. An algorithm is a mathematical function that takes an input (data or observation) and converts it into an output (response or prediction). The nature of the response depends on the choice of algorithm. And the rules by which the response is shaped depend on the values of certain parameters within the algorithm.

It’s important to note that the heart of machine learning is data. Therefore, if the data is corrupt, biased or otherwise faulty, or inappropriate for the task, the machine learning application will fail to meet expectations. This is nicely summarized in the “garbage in, garbage out” paradigm.

Further caution needs to be exercised when considering the role that the machine learning solution is intended to play. Not all tasks lend themselves directly to the application of machine-learning techniques, or call for specific approaches.

In a nutshell, machine learning encompasses the parts of artificial intelligence where the data is key to discovery.

By this definition it excludes parts of AI that are rules-based and where the rules are imposed independently and unchangeably. In these cases, the machine has no way of changing or adjusting these rules.

Where is this used?

Jump to an Industry: Agriculture, CRM/Customer Service, eCommerce/Retail, Education, Energy, Entertainment, Finance/Banking, Healthcare, IoT, Gaming, IT, Marketing/Sales, Media

Natural Language Processing

Natural Language Processing (NLP) is the area of artificial intelligence that focuses on interactions between computers and human languages. Machines need to be programmed to analyze big data sets of languages. NLP is exhibited through text classification, sentiment analysis, language translation, speech recognition, entity recognition, summarization, topic segmentation, and domain specific chatbots.

Issues arise due to the fact that human languages are complex and are not limited to a finite number of interpretations. Words are not binary, and it is difficult for NLP to respond like a human. This technology is improving though, and today’s NLP models can be categorized into rule-based (linguistics), statistics-based, and neural network or machine learning based.

Where is this used?

Jump to an Industry: CRM/Customer Service, eCommerce/Retail, Finance/Banking, Gaming, Healthcare, Hospitality, Insurance, IoT, Media, Travel

Natural Language Understanding

Natural Language Understanding (NLU) or Natural Language Interpretation (NLI) is a subtopic of NLP that specifically deals with machine reading comprehension. This encompasses one of the more narrow but especially complex challenges of AI: how to best handle unstructured inputs that are governed by poorly defined and flexible rules and convert them into a structured form that a machine can understand and act upon. While humans are able to effortlessly handle mispronunciations, swapped words, contractions, colloquialisms, and other quirks, machines are less adept at understanding unpredictable inputs, much less molding them accordingly for conversational continuity.

NLU is comprised of a set of models like relational analysis or semantic roles in which each component solves a particular problem to derive the structural information from unstructured text. Some of the most public “failures” of artificial intelligence have manifested through NLP and biased data sets.

Where is this used?

Jump to an Industry: CRM/Customer Service, IoT, IT, Media

Predictive Analytics

Predictive analytics lives at the intersection of supervised learning, data analytics, and domain knowledge. Predictive analytics involves mining data and modeling it accordingly with the purpose of creating assumptions about future events. These assumptions, or “predictions” are therefore based on insights derived from historical data that are then applied to current data in order to predict future actions.

Typically, historical data is used to build a mathematical model that captures important patterns. That predictive model is then applied to large data sets to predict what will happen next. Frequently, supervised machine learning techniques such as classification and regression are used to predict a future value set.

Similar to supervised learning, predictive analytics can be influenced by bias. If the inputs for analytics are biased, this can affect the integrity of the data and have negative consequences as a result. However, predictive analytics have been extremely successful for an array of use cases such as fraud detection, stock prices, health conditions and purchasing behavior.

There are three themes in particular that showcase the capabilities of predictive analytics. The first is personalization: an algorithm analyzes past behavior, such as previous purchases, of a subset of people with similar profiles in order to personalize suggestions to others who fall within that profile.

Another key capability is predicting potential ROI. An algorithm is trained to focus on a key metric across profiles and is therefore able to predict value of similar profiles — whether that be a person (such as an employee) or a service (such as an insurance quote).

At its core, predictive analytics is an increasingly accurate way of improving efficiency and optimization. For example, an algorithm can analyze food delivery routes and optimize them for both more efficient routes, and more accurate delivery estimates.

Predictive analytics can blend applied statistics, machine learning, and other advanced analytics with big data to provide highly accurate insights in regards to future behavior across industries.

Where is this used?

Jump to an Industry: Agriculture, eCommerce/Retail, Energy, Entertainment, Finance/Banking, Food, Hospitality, Human Resources, Insurance, Marketing/Sales, Sports, Transportation, Travel

Reinforcement Learning

Reinforcement learning refers to algorithms that learn to reach a specific goal (e.g., maximize a reward) via the experience they gain from exploring an environment.

Compared to other concepts of machine learning like supervised or unsupervised learning, reinforcement learning is largely independent of external data. Instead, the data is internally created through experience: once the environment and its rules have been established, the algorithm will explore the available event space to determine the best path to success.

The most successful application of the reinforcement learning paradigm may have been achieved when combined with neural networks and deep learning. These have enabled machines to play games (computer games or board games) and become skilled enough to even beat the most experienced human players.

Reinforcement learning is an approach where a machine-created Agent learns to optimize cumulative gain from Rewards based on whatever Actions are performed. The Agent operates in an environment with specific rules (such as a chess game). At a given time step, the Agent can choose which Action to take based on a specified set of available actions, where the Reward of the action depends on the State of the system. The State of the system and the Reward associated with the Action may depend on previous choices. The goal for the Agent is then to find the optimal path to maximize cumulative gain (winning the game).

Where is this used?

Jump to an Industry: Education, IT

Rules-based Bots

A bot is an artificial intelligence software designed to perform a series of tasks on its own without the help of a human being. A chatbot is a bot which conducts a conversation via auditory or textual methods. This is in contrast to traditional robotics that involve a machine handling a physically manual task or tasks based off of the rules programmed into the software-based bot.

As the name suggests, a chatbot is simply a software assistant that a person or a different bot can engage with via a conversation to accomplish a task or a set of tasks. The conversation can take place either through text messaging or voice commands.

A rule-based chatbot is not necessarily intended to sound human, make small talk, or pass the Turing test (this would require other forms of AI) — rather, it’s to rapidly lead a person through a streamlined channel of information as efficiently as possible. So unlike bots that use other forms of AI, these bots do not need to “learn” how to actually chat with a sophisticated level of comprehension; they just follow the deterministic set of rules laid out for them.

That’s not to say that these bots are simple. They can be simple, but they can also be extremely complex depending on the complexity of the rules and extensiveness of the decision tree.

Decision trees are laid out to send users down a designated path with allowing them a limited set of options at each juncture. Complex decision trees can be time-consuming to write, and it is also impossible to write rules for every possible scenario. This is both a limitation as these bots work well only in a narrow demain, but also a good option for these narrow domains as it removes potential for error.

Rule-based bots can be particularly helpful when it comes to providing binary options. This has been successful in collecting information for travel purposes, customer support, and sales, for example.

Where is this used?

Jump to an Industry: eCommerce/Retail, Finance/Banking, Gaming, Insurance, Travel

Supervised Learning

Supervised learning is the machine learning task of discovering the rules that map an input to a response by providing the algorithm with examples of the desired input and response behavior for it to learn from.

The majority of all machine learning tasks, historically and today, are based on supervised learning. The first step is to choose a machine-learning algorithm that is appropriate for the task at hand. This choice depends on the data that will be used to train the algorithm and on the type of response that is required. The actual response to an input (i.e. data point or observation) is governed by both the choice of algorithm and by the values of adjustable parameters within the algorithm.

By providing the algorithm with examples of the desired input and response behavior, the exact values of these parameters can be discovered. This way the machine-learning model learns to respond to an input with the appropriate output.

Since supervised learning relies on the example input-output pairs that are provided during the learning phase, the model will only be able to operate within the constraints of the training data.

This places additional responsibilities into the hand of the person providing the data.

Any bias, be it known or hidden, will carry over into the learned model. For example, training an image recognition model by providing a large number of images of cats will only allow the model to learn to recognize cats. It will not be able to differentiate between cats or dogs.

For example, at the post office, the handwritten (or printed) addresses are read by a machine which will extract and interpret the zip code for sorting and routing mail. It can perform these tasks at much higher speeds than any human. But it is limited to the training data that it has received.

Supervised learning includes functionality such as handwritten digit recognition, image recognition, and text classification.

Where is this used?

Jump to an Industry: Agriculture, CRM/Customer Service, eCommerce/Retail, Finance/Banking, Human Resources, Insurance, IoT

Unsupervised Learning

Unsupervised learning means that the machine discovers an underlying structure in the data only from within the data itself — there are no predefined inputs with associated outputs.

This type of machine learning is still only used for a small percentage of applications when compared to applications based on supervised learning. However, labeled data — especially in large quantities — is often difficult or expensive (in time and money) to acquire. In principle, unsupervised tasks can take advantage of the immense quantities of largely unlabeled data we currently have access to.

The inner workings of an unsupervised ML task depend to some degree on the exact algorithm that is being used. And the choice of the algorithm is often driven by the purpose of the task. Generally, the algorithms determine similarity between data points and allows grouping them together. This way unsupervised machine learning algorithms infer an underlying structure or pattern in the data. Unsupervised learning is largely an overarching umbrella that encompasses clustering, data compression, dimensionality reduction, and topic modeling.

While unsupervised learning can be used for discovering patterns, it oftentimes lacks specificity and reliability.

Where is this used?

Jump to an Industry: Agriculture, CRM/Customer Service, Finance/Banking, Healthcare, Transportation



Weather and Climate Forecasting

Agriculture is extremely dependent on the climate and seasons, and especially with climate change, there is heightened uncertainty in regards to timing crop cycles. In 2016, there were 772 weather and disaster events, triple the number that occurred in 1980. Farmers had to deal with significant losses that were largely unexpected. AI is now being used to help farmers overcome these adverse effects of climate change and unpredictable cycles. In India, AI is being used to increase yields by 30 percent by providing timely sowing advisories to farmers based on climate models.

A startup in India, Fasal, has gone one step further and uses IoT sensors to monitor micro-climates of the farm and provide recommendations to farmers for optimizing yields. In 2013, Monsanto acquired an AI startup called Climate Corp for $930M that underwrites weather insurance for farmers by using predictive analytics.

Predicting Crop Yield

Crop yield prediction is a vital part of the agriculture supply chain process that helps the companies involved in the multi-billion dollar corn industry such as grain elevator operators, ethanol producers, commodities traders, hedge funds, insurance companies, and even the farmers plan activities based on supply forecasting. This time consuming and labor intensive activity is typically carried out by the USDA where agents physically survey the farms and painstakingly create a forecast for the year.

A startup called Descartes Lab is using computer vision and deep learning to analyze the satellite images of farms to make those same predictions without any humans involved, and is consistently beating the USDA’s predictions. By analyzing the satellite images to measure chlorophyll density daily, Descartes Lab makes a prediction that is updated every two days and is granular down to the county level.

Agricultural Robots

Agriculture is still a very labor intensive industry worldwide. That’s why AI-powered robotics is being widely used for functions like precision weed control, watering, crop health, harvesting, planting and seeding. Drones are even being used to recreate bee pollination to create more reliable crops year over year. The impact of precision farming techniques is that it greatly reduces waste and makes food security a reality in a very sustainable manner. Robotics are also used to grade and sort agricultural goods based on the quality of the product. On an industrial scale, this is done by robots that save the producers a lot of time and money by turning a manual process into an automated one. Tomra produces a range of equipment that can sort and grade products for the food industry.

Robotics are also used to grade and sort agricultural goods based on the quality of the product. On an industrial scale, this is done by robots that save the producers a lot of time and money by turning a manual process into an automated one. Tomra produces a range of equipment that can sort and grade products for the food industry.


Bricklaying Robotics

In 2015, several robotic bricklayers were developed: Australian company Fastbrick Robotics announced an end-to-end robotic bricklayer called Hadrian X, which uses advanced AI algorithms to leverage a multi-axis arm to lay bricks on a foundation, apply adhesive, and detect discrepancies between digital plans and the actual building site. In the U.S., SAM (short for The Semi-Automated Mason) uses similar technology to lay 3,000 bricks a day, as opposed to a human bricklayer’s average of 500.

SAM delivers real-time construction analytics through its AI. However, SAM and other similar robotic bricklayers still require human workers to load brick, refill the robot’s mortar and clean up the joints of the brick.

CRM/Customer Service

Problem Classification

It goes without saying that one of the most important components of any business is customer retention. One way businesses ensure that their customers keep coming back is by keeping them engaged and well-supported when they require assistance. This is achieved by optimizing customer service, and one of the ways to do this is by using machine learning to categorize a large volume of problems into different buckets so that they can be addressed efficiently and keep customers happy. Customer Service teams can leverage a machine learning algorithm via supervised learning to learn how to classify incoming issues based on an existing ticket database. This algorithm incorporates feedback to become increasingly accurate over time.

Automated Answers

As businesses automate more and more of their customer-facing processes, adding in a layer of artificial intelligence can help increase efficiency levels to a new degree.

Natural language understanding and natural language processing can be used to help customers find solutions to their questions by analyzing incoming inquiries and suggesting relevant answers that help customers continue along their paths. Similarly, NLP and NLU can also be used to suggest answers to the customer service teams as they work on these issues, which can be especially helpful in the onboarding stage for new team members.

Problem Trend Spotting

In business, putting out fires is often par for the course. New problems are always emerging, and new solutions are constantly being developed. The gap between new problems emerging and resulting solutions can be narrowed if the problem is recognized as it surfaces — something that is achievable with unsupervised learning.

Unsupervised learning and pattern recognition algorithms can be implemented to detect new patterns showcasing problems as they emerge, and direct customer teams to these problems as quickly as possible.

Customer Service Bots

Bots are used to automate various elements of communication, particularly during the pre-sales and post-sales aspects of business-to-consumer transactions. Bots can use natural language processing to understand customer communication and converse accordingly, and bots can also function purley based on decision trees.

Bots in this fashion are often programmed to lead customers down pre-determined workflows to collect all the information necessary to solve issues for customers — often without ever needing to involve a human. The benefits of decision tree-based bots are that they never go rogue based on user bias (like NLP bots have done from time to time), and they always provide an instant response.



Companies including Sephora, LEGO, eBay, and Aerie by American Eagle Outfitters have launched chatbots on Facebook Messenger, Kik, or on their apps to help users discover products, access customer service, and make returns. These chatbots use NLP, Machine Learning, and Deep Neural Networks to classify requests, suggest items or FAQs, or lead customers down step-by-step paths, such as making a return. Some products, such as Shopify Messenger, additionally allow users to make purchases through the bot.


Personalized Online Education

EdTech is one of the fastest growing technology sectors in the world. In 2017 alone, investors including Amazon, Google, and Goldman Sachs staked $8.1 billion in education technology companies. Many of these emerging EdTech companies, such as Carnegie Learning, use real-time data and AI algorithms to customize learning syllabi, provide feedback, and offer analytics for parents.

Yuanfudao, China’s first EdTech unicorn (a startup valued at over $1 billion), uses an AI-powered matching algorithm to connect students with live tutors.


Forecasting and Efficiency Optimization for Clean Energy

AI can autonomously optimize home and city energy consumption. Siemens released software that can operate energy grids autonomously and tweak different energy loads to increase efficiency levels. When new energy sources become available (like a solar park or wind farm), the software adjusts responsively. On a larger scale, companies such as Nnergix use machine learning to forecast atmospheric conditions to predict when and in what volume sustainable energy sources will be available. Numerous technology companies have also released products that allow consumers to optimize home energy consumption. Nest, a smart thermostat, adjusts temperatures according to inhabitant occupation habits, and claims it helps customers save between 10 and 12 percent on heating bills and 15 percent on cooling.


Personalized Content Suggestions

One of the defining features of the streaming behemoth, Netflix, is its “Recommended for You” engine. Known as “The Netflix Algorithm,” it tracks viewing habits by genre, director, and content length, as well as other factors such as which cover images are most likely to result in the user clicking on the content — and then uses this data to offer targeted content suggestions.

Similarly, “Spotify Discover” uses a personalized algorithm largely based on implicit feedback (such as stream counts) to quantify users’ tastes, and then recommend music that the user is likely to enjoy. These AI-powered recommendation algorithms have proven to be extremely successful.


Wealth Management

AI has impacted wealth management primarily through robo-advisors. Robo-advisors allocate, manage and optimize assets with minimal human intervention. The most popular Robo-advisor, with billions under management, is called Betterment, and was released in 2010 after the financial crisis.

The primary value add of a robo-advisor is that it offers lower management fees than human financial advisors and fund managers. One such robo-advisor, EquBot is powered by IBM’s Watson, and uses proprietary AI-powered models to optimize investments.

Chatbots (Finance)

Banks have adopted chatbots with zeal; HSBC, Capital One, JP Morgan Chase and Wells Fargo are just a few of the banking behemoths that have adopted AI-powered chatbots for customer service, bill pay, and wealth management. Some banks are also using bots for loan calculations, applications, and refinancing; OCBC received $7 million in loan applications in the first three months that it launched its chatbot, Emma.

These chatbots use varying degrees of AI, but in general they rely on NLP for understanding customer requests, predictive analytics for loan calculation and payments, and machine learning to become more precise and accurate over time.

Fraud Prevention

Companies including Stripe, Airbnb, and Uber, as well as most banks today use Machine Learning for their fraud management solutions. AI can sift through big data related to fraudulent activities, analyze fraud traces, and develop predictive algorithms for fraud threats. Because AI can analyze data in real-time, it can detect anomalies and alert users and companies to the likelihood of fraud instantly, thereby protecting the bank, company, or customer before the potential fraudster is able to execute any activity. Additionally, AI can detect false login information by analyzing session, device, and behavioral biometrics to build a profile for “normal” user login behavior.


Real-time Analytics for Food Delivery and Inventory

Food delivery is a highly time-sensitive and circumstance-dependent industry — and it’s an ideal use case for AI. Using real-time analytics about delivery location, GPS, traffic, and food preparation time (dependent on factors such as inventory and the number of workers available), AI can find the most efficient delivery route and keep customers updated in real-time about their orders. DoorDash, an AI-powered food delivery service, is one of many examples of companies that rely on predictive analytics and real-time analytics to keep customers happy.

Additionally, many restaurants and fast-food chains rely on predictive analytics to calculate how many drivers, cooks, and inventory are needed based on customer demand.


Bot-based Account Recovery

One of the most common customer service queries in mobile gaming is account recovery. Re-granting access is a rote process that does not require human verification, which is why many mobile gaming providers use NLP-based chatbots to help players get authorized access to their accounts.

The chatbots can respond instantly, process the user’s data in real-time, and use NLP to understand the nature of the player’s request. This expedites the process for the player, and minimizes the amount of human intervention needed from the gaming provider.

Bot-based Self-Service

Keeping users engaged and in-app is paramount for any mobile game’s success. Because of this, many mobile gaming providers use chatbots to expedite customer service inquiries that can be answered through self-service. These inquiries can include questions about advancing to a new level, purchasing currency, or saving a game.

The chatbot uses NLP and AI-powered auto-classification to answer the customer queries in real-time, allowing them to return to the game as quickly as possible.


Dark Warehouses

Dark warehouses refer to warehouses populated by robots instead of humans. These relatively small and compact warehouses eliminate the need for freights or forklifts, and instead are populated by robots that unpack pallets, zoom to and from different stations on conveyor belts, and communicate data such as product inventory through sensors. Companies including Kroger, Coca Cola, Target, and Walmart have all invested in Dark Warehouses — and cut labor costs by up to 80 percent. Alibaba, the Chinese e-commerce giant, automates 70 percent of its warehouse work through robots that can carry up to 500 kg.

Robotics for Assembly Lines

Using computer vision, sensors, and real-time analytics systems, robots have become a vital component of assembly lines in manufacturing today. Robotic assembly is widely used in the automotive industry, consumer electronics, medical devices, and household appliances. Assembly line robots come in three configurations: six-axis arms, four-axis “SCARA” robots, and the “Delta” configuration, which uses motors in the base to move three arms. The largest US-based industrial robot manufacturer is Omron Adept Technologies, which offers mobile robots, industrial robots, other automation equipment, and machine vision systems. Other large robotics providers include FANUC, Kuka, and Mitsubishi Electric.


Image Recognition for Diagnoses

Most image-based medical diagnoses consist of pattern recognition — whether or not a mole has irregular or notched borders can indicate if it is cancerous, for instance. AI has proven better than humans at this type of pattern detection, and as image recognition technology improves, using AI for diagnoses has become of increasing interest in the medical community. Machine learning algorithms have been able to detect the presence or absence of TB by looking at X-rays with 96 percent accuracy, a higher success rate than any human radiologist.

SigTuple uses AI for intelligent screening to perform blood smear analysis, urine analysis, semen analysis, retinal analysis, and chest x-ray analysis. Google’s DeepMind is currently developing an AI to detect breast cancer by looking at mammograms, and Google researchers found that the company’s AI could detect diabetic retinopathy, a diabetes-related complication, as accurately as ophthalmologists. A startup called Optellum is working to commercialize an AI system that diagnoses lung cancer by analyzing clumps of cells found in scans. AI-powered diagnoses promise to cut costs for hospitals and consumers by billions of dollars, and improve accuracy in the process.

Drug Discovery

Diseases emerge and change with every generation due to new pathogens, diet, lifestyle and the environment. Tackling these new diseases as they surface while advancing research on existing diseases is a very expensive and arguably insurmountable feat — the process of discovering new drugs is extremely laborious as it involves combing through mountains of data and identifying patterns. AI with Machine Learning is very proficient at this, and can help shave time and cost off of this process. On Jun. 12, 2007, a robot called Adam identified the function of a yeast gene by searching public databases and deriving hypotheses — and then tested them in the lab with a robot. A group of researchers at Berg, a pharmaceutical company based in Boston has applied Machine Learning to discover a new drug: BPM31510, which is in phase II clinical trials for treatment of advanced pancreatic cancer.

This method has eliminated the trial and error approach of the past and is instead using data to derive more precise hypotheses based on studying new molecules and their effect on cancer metabolism. The future of biomedical research will rely on AI and those adept at coding who can apply algorithms to the process of drug discovery. Here is a comprehensive list of startups applying AI in drug discovery.


Personal Assistant Bots

The hospitality industry uses bots via Facebook Messenger, WhatsApp, or even built into the brand’s own app for 24-hour customer service, directions, bookings, restaurant and activity suggestions, and room service requests. Both Marriott and Hilton have chatbots that provide immediate guest services, and smaller hotels are following suit. For instance, Cosmopolitan of Las Vegas has an AI-powered concierge named Rose that can be either voice or text message activated.

Chatbots for hospitality, powered by NLP, enable hotels to offer guests personalized recommendations and instantaneous service at all hours and with no human input.

Human Resources

Matching Algorithms for Candidates and Jobs

One of the best use cases for AI today is matching needs to solutions, such as Uber’s AI algorithm that matches riders with drivers. Similarly, recruiting and professional development platforms including LinkedIn, Indeed, and startups such as Woo.io, use AI to algorithmically pair candidates with jobs.

These algorithms go beyond simple keyword pairing, and include elements such as relevancy (title, seniority, job description), location, and salary prediction. Predictive analytics, as in this case, can be very useful for marketplace-driven innovation.

AI-Powered Performance Management Systems

The performance management software market was worth $2.3 billion in 2016, and has grown exponentially over the past two years. IBM uses its AI technology — Watson — internally for performance management, and claims it can predict future employee performance with 96 percent accuracy.

Companies such as BambooHR, Nexus AI, and Zenefits have developed products that give employees continuous feedback, not only on performance but also on context-specific strengths and weaknesses — e.g. who works best on what projects, and with whom. AI for performance review can reduce biases including recency bias, demographic bias, and contrast bias, and give real-time feedback, to incentivize and intervene based on employee performance.


Chatbots (Insurance)

The insurance industry contributes over $500B a year, accounting for over three percent of the U.S. GDP. Competition between insurers is stiff, which is why the industry has adopted chatbots, AI, and automation with zeal, using them for quotes, payments, customer service questions, and even claims. For instance, Geico’s chatbot, Kate, can answer questions about policy coverages, billing information, and FAQs. Allstate’s AI-powered chatbot guides small business clients through the insurance process.

Predictive Analytics for Quotes

The insurance industry was an early adopter of predictive analytics, and at this point every major provider uses AI for quotes. AI can process hundreds of gigabytes of data in seconds, and generate a quote almost instantaneously after receiving data. This has revolutionized conversion for insurance companies; studies have shown that the faster the digital quote process is, the higher the policy purchase conversion rate is. for instance, a popular startup called Lemonade provides renters and homeowners insurance that is powered by behavioral economics and AI, which allows customers to complete the policy sign up process in 90 seconds, and have claims and payouts processed in three minutes. Many insurance providers today have integrated AI-powered quotes with AI-powered chatbots to automate and expedite the policy purchasing process.



Wearables, or internet-connected devices that are incorporated into items of clothing or accessories, use AI to track health, classify exercise routines, and track key movements, such as removing the cap to a pill bottle. Today, the most popular form of wearable technology is smartwatches, with the Apple Watch leading the market, and Fitbit coming in second. Other products integrate sensors and AI to provide personalized customer coaching and feedback.

For instance, Sensoria Fitness, a producer and vendor of smart sport apparel, provides customers with an AI in-app coach to improve exercise routines using performance analytics. Similarly, Atlas Wearables, a fitness band, measures heart rate and calories burned, and uses a machine learning algorithm that automatically classifies exercise routines in a 3D vector. The wearables market today is largely centered on health and fitness, as the primary benefit of a wearable is that it can track movement.


39 million Americans currently own smartspeakers, with Amazon Echo holding the largest market share. Smartspeakers integrate with other connected devices including doorbells, locks, lights, smartphones, and even coffee machines.

The speakers act as the nexus, or command point, for other IoT devices, and require only voice commands to execute tasks to other machines via the Internet. Understanding these commands and performing the correct corresponding actions requires advance NLP. Currently, the market is dominated by the Echo, Google Home, and Apple Homepod, which currently owns four percent of the market, but only launched four months ago.


Contract Review

Contract review systems today include Kira, Luminance, RAVN, and eBrevia, among others. These tools use pattern recognition, information retrieval, and NLU to organize contracts, give snapshots of relevant information, and generate custom reports. One study found that a human lawyer takes 91 minutes on average to read five NDAs and extract the relevant information, while an AI can do the same — with 94 percent accuracy, as compared to humans’ 85 percent accuracy — in 26 seconds. The majority of contract review software uses machine learning to improve accuracy and thoroughness over time.

Reporting & Classification

One of the initial AI use cases was real-time reporting and data classification. AI can process big data rapidly, classify it according to predetermined rules, and generate reports that continue to update as more data comes in.

The applications of this technology are extremely far reaching — classification is used in everything from customer service for auto-classifying incoming tickets, to contract review technology that sorts contracts based on NLU (Natural Language Understanding) instead of simple keyword classification, to sales and marketing software that classifies leads and prospects.


Content Personalization

Marketing and sales rely heavily on customer data to personalize interactions. This can range from predictive lead scoring, to tools like BuzzSumo that identify trending content, to actual content personalization via products like RichRelevance.

Additionally, email marketing tools such as MailChimp and Sales tools such as Salesforce Einstein utilize AI to optimize content elements like send time, message length, and subject line. These tools use customer data to analyze which customer cohorts react best to different types of messaging and content.

Ad Personalization

Facebook, Google, Twitter, and other web-based platforms use AI to personalize advertising by leveraging search history, preferences, data from cookies, and hundreds of other behavioral and demographic data points. This allows them to deliver advertisements to the right people — tech gadgets to people who tend to buy lots of consumer electronics, luxury retail to those who frequently purchase clothing, cookware for those who search for recipes and kitchen items, etc. AI allows for near real-time personalization, and optimizes targeted ads for people who are considering buying in a brand’s space.


Roughly two percent of e-commerce site visitors actually make a purchase. To lure these customers back, advertisers use “retargeting” — targeting consumers who previously demonstrated interest in the brand but did not make a purchase with advertisements about similar products or the same products that the customer had previewed.

Machine learning can be used to automated this process by accessing the brand’s traffic and user behavior data in real time, and then learn from it to become better and better at targeting cohorts and increasing conversions. The algorithm learns through an ever-increasing corpus of data about individuals and cohorts to become increasingly sophisticated and accurate.

E-commerce Chatbots

E-commerce chatbots are being used to serve in various roles: some help shoppers find an item and personalize search results through Machine Learning; others use facial recognition to personalize makeup suggestions; and others answer customer service inquiries about topics such as shipping, inventory, and payment. The H&M chatbot on Kik asks users a series of questions about their style and then creates a “fashion profile” for each user, complete with outfit suggestions and purchasing options. Sephora’s chatbot offers makeup tips and product suggestions, while eBay’s chatbot on Facebook Messenger uses the chat interface to help users find certain items.


Automated Articles

Outlets including the Associated Press (AP), Yahoo!, and Fox have been using AI to automatically generate data-backed stories since at least 2015. Typically, the only stories written by AI are related to sports news or finance news — stories that consist primarily of data. The technology works by processing structured data, mining the data for meaning, and then using natural language generation within certain linguistic parameters. The two most popular providers of this technology are Narrative Science and Automated Insights (used by Yahoo! News and the AP).

News Aggregators

News aggregators such as Flipboard, Google Reader, and Apple News use NLP and NLU to “read” incoming articles, tag them by topic and keyword, and weed out duplicate articles. Additionally, Flipboard’s AI, for instance, blocks spam domains that try and appear as legitimate news outlets.

Personalized news aggregators also use Machine Learning to learn user tastes and provide increasingly relevant and interesting content to the user. According to a Reuters Institute poll, 60 percent of publishers use AI to improve content recommendations.


IoT and Predictive Analytics

Most professional sports ranging from baseball to basketball to hockey have embraced a combination of IoT (Internet of Things) and predictive analytics to optimize plays and make strategic draft picks. In the NBA, teams have computer vision systems to track how the ball and the players travel around the court.


The team then uses this data by reviewing machine learning-based analysis to identify the optimal position for each player on specific plays. IoT sensors on shoes and other equipment, such as a bat in baseball, provide crucial data on player strengths and weaknesses. One company that provides this technology, SportLogiq, uses AI to help teams make strategic draft picks and optimize plays, and is currently used by two-thirds of NHL teams, as well as teams in other sports including soccer.


Autonomous Vehicles

Most major car manufacturers and a number of technology companies have invested in autonomous vehicles, including Ford, Mercedes, Apple, Google, Intel, Volvo, Uber, and Delphi Automotive. Waymo, a subsidiary of Google’s parent company, Alphabet, has been testing its self-driving car technology on roads since 2009, and claims that its cars have driven over seven million miles on city streets.

Self-driving cars rely on complex AI and IoT technology to “see” objects on the road, process them correctly, and react in the safest way possible. Currently, self-driving cars have backup humans in the driver’s seats to refine and correct the car’s algorithms.

Hyperloops/Autonomous Public Transport

Hyperloops are a proposed method of transport consisting of a sealed system of tubes through which an autonomous pod filled with passengers can travel without air resistance at extreme speeds. The city of Dubai recently unveiled a full-scale model of a passenger pod for a hyperloop at Dubai’s City Walk mall. Plans for a 12-minute hyperloop between Abu Dhabi and Dubai have been in the works since 2016, when Hyperloop One announced an agreement with the Roads and Transport Authority.

Elon Musk has long been a proponent of hyperloops, citing benefits including minimized safety risks because they’re built either on columns or underground, as well as reduced risk of collision, reduced impact by weather, minimized energy expenditure because there’s no wind resistance, and lower prices due to contactless tracks, minimal engineering, and conductor employment needs. Any autonomous transportation method requires advanced AI for algorithmic scheduling, computer vision, and deep learning to parse all the incoming data, such as whether an object ahead is a person, debris, etc.

Last Mile Delivery

Last mile delivery, or the movement of goods from a hub or warehouse to a final destination such as a home residence, is a major logistical challenge for many companies. For that reason, Amazon and other companies including Boeing, Dominos Pizza, and Chipotle in conjunction with Google have invested in autonomous drone technology for short deliveries.

Additionally, AI solutions have been in use for last-mile delivery logistics systems to make real-time route predictions and suggestions, adjust arrival time predictions, and provide accurate risk metrics for potential delays. For instance, ClearMetal, a supply chain visibility tool, uses AI for real-time updates and predictive alerts.


Bot-based Pricing and Suggestions

Expedia, Kayak, and Skyscanner, among numerous other travel and leisure providers, give users recommendations for flights, hotel rooms, and activities, all through a chat interface on Facebook Messenger. Additionally, many airlines and hotels have their own chatbots to help users find the ideal flight and accommodation.

Icelandair’s Facebook Messenger chatbot allows users to book stopovers in Iceland through the chat interface, and can also answer customer service queries. Similarly, the Kayak bot asks users to set parameters in terms of dates, location, and budget, and then intelligently suggests hotels and flights that the user can book through Facebook Messenger.

Peak Travel Time Analytics

Most airlines today use AI algorithms to determine flight destinations, times, and flight frequency. These algorithms look at peak travel time to different destinations based on time of day, week, and year, and then schedule flights to optimize them for maximum occupancy. Skyscanner, for instance, used an algorithm within its Travel Insight data platform, that looked for patterns of over 50,000 origins and destinations during one year to plan for the following year.

AI has also been used for peak travel time planning on the consumer side. An online rail booking service in the U.K. called Trainline has a bot that advises passengers where they’re most likely to find a seat, depending on location and the journey direction. It also has price prediction based on passenger demand, and identifies the cheapest available ticket for every day.