Thomas H. Davenport: The AI advantage, How to Put the Artificial Intelligence Revolution to Work

AI is slowly maturing and now we can see its use in many industries, even though they had some challenges in the beginning.

In health industry M.D. Anderson tried to implement huge project to tackle oncology issues with IBM, spending more than 60 million USD, but they didn’t get to phase of treating patients. On the other hand, in separate projects they use AI for:

Proposing hotels and restaurants for patients’ families,
Application to determine which patient needs support at paying bills,
Cognitive help desk for IT support.

In financial industry DBS based in Singapore, tried to introduce AI based robot to support customers in their wealth management activities, but wasn’t very successful. But they stayed with AI and implemented it in:

ATM – prediction of money supply.
Churn of employees based on time of holidays, medical leaves and even length of e-mails.
Detecting fraud in trading.
Algorithm-based lending model.
Chatbots in customer services.

Amazon, probably one of the most active companies in field of AI usage, have their challenges with drone delivery and self-checkout stores at Amazon GO. But they are strong users of AI in their voice lead assistants Alexa/Echo. But as Bezos said, majority is happening bellow surface, algorithms running:

Demand forecasting.
Product search.
Product and deals recommendations.
Merchandising placements.
Fraud detection.

We need to be careful at estimation of AI influence and spread- as Roy Amara said: “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”[1]

There are several different technologies that are all called AI. This can sometimes lead to terminological mess and can cause some complications in communication about usage of AI. In general, we can talk about seven technologies and their usage:

Statistical mechanical learning – that automates process of training and fitting models to data and it is used for highly granular marketing analyses on big data.
Neural networks – that uses artificial “neurons” to weight inputs and relate them to outputs and is used for identifying credit frauds and weather predictions.
Deep learning – which is neural networks with many layers of variables or features, and it is used for voice recognition and extracting meaning from text.
Natural language processing – that analyzes and understands human speech and text and it is used for speech recognition, chatbots, intelligent agents.
Rule-based expert systems – is set of logical rules derived from human experts and it is used in insurance underwriting and credit approval.
Physical robots – that automate a physical activity and are used in factory and warehouses for different tasks.
Robotic process automation – is automating structured digital tasks and interfaces with systems and is used for credit card replacement, validating online credentials.

When choosing proper machine learning approach first we check for the appropriate algorithm to be used. Then we check what kind of learning is appropriate for certain use case. Machine learning models are based on statistics and they should be compared to conventional analytics to establish their incremental value. NLP also uses two approaches – statistical or semantic.

Cognitive technologies are specifically aimed at knowledge work processes within organizations and these have been particularly important and problematic in terms of productivity. As Peter Drucker argued as far back as 1959 (when he coined the term knowledge worker), knowledge work productivity is the key to success in the twenty-first century.

Thus far, the most aggressive adopters of AI have been “digital native” online businesses, large enterprises and tech startups, many of which have some aspects of cognitive technology as the key differentiation for their businesses. There has been relatively little adoption by small and medium size businesses that are not in technology-intensive industries. One of the reasons why SMB is not strong in usage of AI is also lack of awareness and understanding of what is possible in AI arena. Because of that managers in SMB are missing opportunities.

“We need to get intelligent about AI and move from experiment to solving real problem at scale.”[2]Good explanation of today’s status of AI from GE CIO Jim Fowler. How committed are companies to AI can also be seen by where they choose to implement AI projects. If they are really committed, they will do it in customer facing and revenue generating activities. If they are not, they will work in supporting functions. Challenges of integration are also making big scale, real problem solutions hard to implement. In order to prepare companies for integrations, a lot of work needs to be done not only on AI, but also on IA – information architecture.

If we look at AI through the lens of business capabilities, we can say it supports three important business capabilities:

Automation of structured and repetitive work process (robotics or robotic process automation)
Gaining insight through extensive analysis of structured data (machine learning)
Engaging with customers and employees (NLP, chatbots, intelligent agents, machine learning)

Main area where companies are running their AI projects is automation. The technology of RPA is relatively easy to use, but the challenges of implementing it come primary from the business process. Companies should have a good understanding of both their existing business processes and the new processes they want RPA to enable before implementing the technology. When we talk about cognitive insight on top of machine learning used for providing insight, it can also be used for data integration. Connecting different data to same source. Regarding cognitive engagement we can say that there is a lot of cases of usage, but maybe because technology itself is not that capable yet, some real job replacements haven’t been done yet. Usually projects combine different categories that are mentioned above.

If we talked about AI usage in companies, it is important to see it from two angles. First is applications and we talked about it so far, second is capabilities. If we want to work on projects of AI implementation with proper capabilities perceptions, we need to:

Understand which technologies perform what types of tasks
Build on current strength in big data and analytics
Create a prioritized portfolio of technology matched to processes and tasks
Create a series of pilot or proof-of-concept projects
Engage in cognitive work redesign using design-thinking principles
Focus on scaling and achieving productivity benefits

There are three area of assessments that company should do before starting AI projects:

The domain assessment – when companies look at AI projects and are estimating which ones to address first, one of the main tasks is to determine which business domains are the ones where company can benefit by using AI – the domain assessment. Sometimes problems can arise from knowledge bottleneck, scaling issues or inadequate firepower.
The use case assessment – evaluate use cases in which cognitive applications would generate substantial value and contribute to business success.
The technology assessment – it is important to know if current AI capabilities can much use cases defined. If not, small steps approach should be used with long-term planning in mind.

If we want to achieve all benefits that use of AI can bring, we need to re-design work processes. It is easy to just use AI to improve productivity of existing processes, but if you only use it for such scenarios, limitations of old processes will keep AI usage benefits low. Scaling is a challenge today, since companies can run quite few proof-of-concepts but are not able to implement full size projects. If scaling is to succeed, firms must also work to improve productivity. You basically gain benefits, not with people reductions but with growth of business at the same level of peoples.

Through the application of AI, information-intensive sectors such as marketing, healthcare, financial services, education and professional services could become simultaneously more valuable and less expensive to society.

Strategic thinking is, or at least should be, broad, high-level, and aspirational. But for AI this is a bit of a problem, since existing cognitive technologies perform relatively narrow tasks and augment the work of human beings. That generally means that they bring about incremental change not transformational or disruptive change. But it is possible to aggregate a variety of shorter-term projects toward a larger, more strategic objective. Cognitive technologies can support so many different objectives that it’s important to make strategic choices about which ones to emphasize.

If we are looking at areas where AI can be used, it is important to understand what kind of strategies companies have and are their business models oriented on cost effectiveness or improved customer services – revenue-oriented companies. Since first are more likely to use AI projects to improve cost optimization and that means more internally oriented projects and second are more likely to use AI projects to improve customer experience and product improvements and that means more externally oriented projects.

OpenMatters is company that uses AI to analyze strategies and suggest improvements. First, they diagnose their business model. There are four types of business model:

assets and things
people and services
software and data
platforms and networks

Even BCG is now trying to use AI to improve their consultancy. AI will have influence on management also, not only blue and white collars. But we still see that old business models are persisting and that sometime even startups, that don’t have any existing business models to disturb, have hard time to find their position on a market.

Nine factors that limit AI-driven business model change are:

Technologies aren’t quite there yet.
Partial solutions are all that’s available.
AI picks off the easiest parts of process.
No common sense.
Startup processes are required, but startups don’t have customers.
Big companies buy startups.
Startups don’t have resources to wait out change.
The installed base will take a while to disappear.
Most cognitive applications are standalone but need integration.

Candidates that are able to develop cognitive strategy inside companies should have:

They should know something about the major types of cognitive technology and how they are used in business.
They should be effective at communicating to managers in nontechnical terms.
They must understand the key issues of the business and its current strategic direction.
They should have facilitation and process skills.

There are certain elements of strategy that needs to be address or can be used to set up proper process for strategy to develop:

Content – One of parameters of content is knowledge graph. Knowledge graph is a set of entities (people, places, objects) and facts about them and their relationships. When vendors work on AI projects, sometimes they insist on sharing intellectual properties because of usage of their knowledge graphs.
Talent – most of the companies find it hard to source people with AI knowledge, that they need.
Partnership or acquisition strategy – this is very popular in IT and automotive industry. Also, some retailers are either going into partnership with AI companies or acquiring them.
Ambition – it is approach based, it depends on every company goal and it should be suited to company abilities and potential for development.

The task of creating and applying highly detailed statistical models is forte of machine learning. If your organization has a lot of fast-changing numerical data and you want to make sense of it, this method is probably your best bet. This approach is most commonly used in marketing, sales, prevention of fraud and money laundering in banks, precision medicine, approval of claims in insurance and intelligence and military application in government. Possible obstacles for broad implementation of this technology are data availability and that many managers and business persons still don’t know about all possibilities of machine learning.

RPA uses a combination of capabilities to perform structured, information-intensive digital tasks. Returns of these projects are quite high, since they can be run autonomously and can accomplish useful functions in business. It can be used in identifying and reconciling records that don’t match, transferring data from one system to other, issuing receipts, confirmations, preparing standardized replies to e-mail and text, creating reports. With broad application of RPA it is important that it’s done with proper process reengineering and on a proper IT architecture.

Manipulation of information to accomplish a business objective is one of most valuable cognitive capabilities. One of variations of this approach is OCR. Another one is extracting information from contracts and check if they match with reality. It can also be used for combining similar data records across multiple databases.

Understanding human speech and text is really advancing since technologies are getting better and better. Much of its advancement is based on deep learning. This technology can be used for vehicle operations, online and mobile shopping, consumption of music and other content, operations of industrial machines, online education, e-commerce ordering, online support functions.

Plan and optimize operations were always done on a very limited scale, but with cognitive technologies this could change dramatically. Big River – steel producing companies is using technologies for: demand prediction, sourcing and inventory management, scheduling optimization, product optimization, predictive maintenance and outbound transportation optimization.

Ability to perceive and recognize images has advanced considerably during last years. Underlying technology behind is deep learning. Some applications today are: monitoring of construction projects for completion times, usage in autonomous vehicles, recognition of bought products in retail, recognition of product defects in manufacturing production lines, facial recognition in retail and technology, apparel identification and classification, medical images, online photos classification. Technology itself still isn’t developed enough to be fully utilized in production environments, so it still needs human assist in checking and controlling.

Autonomous vehicles and autonomous robots are category where high levels of automation are already achieved, but fully automated technology is still not in use. With use of sensors and cognitive technologies they can be included into human world and work in collaboration mode. Some applications are: delivery of goods, driving humans, drones, manufacturing applications, warfighting applications, warehouses, care for patients and elderly, cleaning duties, autonomous farming.

Assessment of human emotions is one set of tasks that is not usually associated with AI, but some of the companies are already working on it. It is based on facial recognition and some micro expressions that machines can pick up on. It is primary used in sales and marketing. Soma applications are: assessing customer reactions to online content, virtual market research for potential new products and services, evaluation of driver moods, identify boredom in online education, making “social robots” more empathetic, reaction to emotions in online gaming, understanding emotions in health status.

Cognitive technologies are the current equivalent of disruptive technologies for processes. First BPR way (business processes reengineering) started when America was looking for ways to improve productivity in order to compete with Japanese. BPR can be seen as an instance of design thinking. Design thinking according to Tim Brown is “a human-centered approach to innovation that draws from the designer’s toolkit to integrate the needs of people, the possibilities of technology, and the requirements for business success”.[3]

One of the main questions around AI in future is how will AI influence jobs? Two main ideas are:

Full automation, which is probably hard to archive, but it is very appealing to management because of cost reduction and operational improvement.
Second is broad augmentation of human workers by AI, that can bring more flexibility and better long-term results.

Here again we should look at AI usage through automation of tasks and not whole jobs and in this case, we can really understand power of AI. We also need to consider other factors not just ability to automate certain tasks, but also costs, potential cost and skills of humans that would be engaged in automated tasks, benefits beyond labor substitution, regulatory and social-acceptance consideration. If we consider them all, numbers of potential automation that is around 50%, could come down to 5%.

Organizations need to classify their jobs, determine the appropriate level of human and machine activity for each, identify the skills most likely to be needed in the future, and begin preparing both the people and machines for their future roles. Classifying jobs is important, because some jobs don’t even exist today. Author categorize five steps that we can assign jobs to: step in (checking machine works), step up (overseeing machine work and upgrade it), step aside (creating jobs in non-specifically machine jobs), step narrowly (jobs in very narrow specific areas, done by only humans, that create their own support tools), step forward (to construct next generation of AI tools and software). Also, when we estimate jobs according to levels of automation, we can use this classification: level 0 – no assistance from technology; 1 – human assistance from technology; 2 – partial automation; 3 – conditional automation; 4 – high automation; 5 – full automation.

Jobs that will change the most are:

Jobs with a high degree of structure and repeatability.
Digital jobs that don’t involve direct contact with customers.
Jobs that make heavy use of quantifiable data or codifiable knowledge.
Entry-level jobs.
Jobs that don’t generate revenue or profit.

Skills needed for jobs of the future:

Being conversant with how machines think.
Having an understanding of analytics and data structures.
Becoming familiar with different types of AI.
Having domain knowledge of the business and industry.
Possessing a strong ability to communicate.
Having high levels of emotional intelligence

Some of the main technological challenges in AI implementation are: integrations into existing infrastructure, issues with data – not enough or poor quality, managerial issues – not enough knowledge and vendor lobbying, expense issues – not enough budgets.

Easiest way to get into AI is to implement cognitive capabilities from transaction software vendors. Usually transactional systems like CRM, supply chain or HRM from strong vendors have AI already integrated into it. Some of the companies that offer AI integrated into their products are:

Salesforce – Einstein offers functions like predictive lead scoring and prioritization, automating data entry, ad personalization, social media and sentiment analysis, personalized product recommendations and image classification.
SAP – Leonardo offers solutions like cash management in finance, video analysis in brand management and trouble ticket analysis in customer service.
Oracle has strong focus on chatbot-like interactions with its system.
Workday is focused on employee retention analysis.

RPA is another entry-level strategy. It can be used for digitally structured processes. But RPA are not very smart, maybe vendors will try to add some additional AI capabilities into their systems (on top of automation of existing processes), but it is not happening yet. In effective digital organization, smart machines should be able to:

Eliminate process steps or processes.
Create intelligence.
Act intelligently.
Perceive its environment.

Some companies could turn to platform vendors like IBM and their Watson for use of their cognitive application to solve business challenges. Other companies are Cognitive Scale, RAGE Frameworks, Veritone. Another approach is to build multi-vendor and open-source capabilities. Taking the multivendor, open source approach to cognitive software requires that companies be very familiar with the different types of cognitive technologies, that they have sophisticated data scientist on staff and that they willing to expend considerable effort on integration of existing processes and systems.

Precondition of efficient AI is good data. High volume and high quality. High quality means clean, consistent and well-integrated through organization data. Machine learning can do great job in master data management. It should focus on probabilistic matching and not rule based according to Mike Stonebaker. BMO – Canadian bank wanted to introduce cognitive technologies but realized that they need to get their data in place first. They run Smart Core project, improving their data capabilities that will leverage future analytics, data science and cognitive activity. The Smart Core encompasses provisioning of data records, reference data, data governance (seventeen governance communities across bank) and a metadata hub. They use data-driven approach to improve customer analysis and their journey analysis.

GSK – was another company that realize that they need to get their data in order, before starting their journey into AI. They use help from company TAMR and their “probabilistic matching” approach and create single data lake based on Hadoop with three different domains.

Another challenge in data handling is that more and more data is coming from external sources and you need integrations into existing systems. Companies that do B2C were traditionally better at this, but B2B companies are catching up. Neural networks and other machine learning methods enable data scientist to extract important data from digital formats. These AI methods involve advanced search techniques that identify, categorize and gather user-defined data elements corresponding to search criteria. One of those companies is EverString.

Primary failing of AI in technology and the online content industry has been the inability to prevent some unfortunate tendencies in social media. Another big question is about AI fairness and algorithmic bias. If AI is to gain trust and wide usage vendors and companies should make sure that they don’t fall into “hype” mode and overpromise on AI abilities. If AI is to be transparent some sort of explanation for outputs is necessary but it should be done smartly, since if it is done all the time, benefits of using AI will go away. Development can actually go into directions that we would have some government or private sector trusted authority that could certify AI algorithms – for let’s say autonomous vehicles, AI diagnosing patients, using AI as marketing slogan.

Implementing AI can lead to loss of knowledge and skills of people using it. This is one of the fears about use of new cognitive technologies. In order to also address about fear of lost jobs and working with new technologies, management need to be careful about implementing proper change management procedures in proper time (best is to do it when POC is successful on a small-scale use). Three important roles of employees should be involved into implementation of AI: experts, decision makers and recalcitrant learners. But also other change management practices are needed. Good environment – country and company – is also beneficial for good implementation of AI. Mercadona – Spanish retailer has a moto that their people should not do, what machine can. They want their people to use their skills and knowledge for company to improve.

[1]In the book on page 7

[2]In the book on page 32

[3]In a book on page 127