How to Use Artificial Intelligence

How can I use artificial intelligence?

The central role of AI-powered solutions in the workplace is growing rapidly, changing the way companies approach business. Companies in almost every industry are recognizing the potential of AI’s many abilities and applications and are eager to harness it. In 2017, 81 percent of Fortune 500 CEOs listed “artificial intelligence and machine learning” as “important” or “extremely important” to their companies’ future. [1]

While the buzz around AI and the many stories of successful implementations are exciting, it is important to remember that this technology is not one-size-fits-all. Artificial intelligence powers many different cognitive tools, each with varying abilities—everything from machine vision to natural language processing. Different types of AI are designed to address specific tasks and problems. They must, therefore, be trained using different datasets.

For truly effective AI implementations, businesses must carefully consider how they will approach each of the following steps:

There are many companies and vendors that specialize in helping businesses adapt to AI technologies. There are also specialized management consulting firms that work with client companies and vendors to understand how AI might impact their business and develop a company-specific strategy for implementing AI.

Creating an AI project portfolio

The first step for a business planning to implement an AI tool is to evaluate the business’s current needs and capabilities: what is being done well and what can be done better. Identifying business processes that could be improved using AI is a good place to start.

To identify which business areas could benefit from AI, an assessment of process flow, bottlenecks, risk, and latency must be conducted. Subsequently, to determine if AI implementation in the identified areas is possible, the quality, quantity, and applicability of potential training data sets must be evaluated. This assessment is often done in partnership with a management consulting company, such as The Burnie Group, that is skilled in identifying areas of opportunity. Assessments typically focus on areas of the business where there is plenty of data from which to draw insights, but for some reason, it is inaccessible or unexploited. Reasons why existing data might be inaccessible or unexploited include:

1. Not optimally distributed
Think of the healthcare industry: hospitals collect a massive amount of data, but that data is siloed between different departments and practices.

2. Too costly to scale
Think of the financial industry: companies have enough knowledge to answer every client’s financial questions but may not have enough time or money to pay the number of advisers required to do so.

3. Not enough processing power
Think of the retail industry: companies collect huge amounts of data on customers’ spending habits but not possess computers with enough processing power to run analytics and produce insights out of this data.

Once areas of opportunity have been identified, the next steps are to prioritize projects by determining their potential value and the quality of the available training data.

Read about 29 cutting edge applications for artificial intelligence.

READ MORE

Determining project value

When businesses think of AI implementation, many ask themselves, “how can I use AI to discover patterns and unlock insights trapped inside all this data we have?” These types of insights can be hugely valuable for go-to-market strategies. However, AI offers many other significant value-add opportunities, including new customer interaction models (such as chatbots and intelligent virtual agents), smart devices, and business process automation.

After a business has created a project portfolio based on potential areas for improvement, it should evaluate the business value of each potential AI solution. This potential value must be weighed against business priorities, opportunities, and resource constraints. As AI has the potential for widespread impact, ideally all departments and processes of the business should be thoroughly evaluated and departmental feedback incorporated at this stage. It is very common that employees have insight into which low-value tasks could be eliminated and where processes tend to bottleneck because of human decision making or a lack of reliable prediction.

To further establish the value of each project, a business must establish how critical the identified problem is to the overall business strategy and success. This evaluation must then be weighed against the technically difficult to implement AI solution to determine if the solution should be pursued.

After the value of each project has been determined, projects must again be sorted. Certain projects may hold short-term value, others longer term value. Even if a project is deselected, a future opportunity to integrate AI should be considered and noted. The foresight of the potential deployment of cognitive capabilities is what will create a true competitive advantage. Once all of the preliminary evaluations have been completed, it is time to choose which AI technology and what training data sets will be best for the chosen task.

It can be difficult for businesses without industry-specific AI experience to determine the value of an AI project. It can be even more difficult to predict future effects, integration challenges, and the impact on labour and culture surrounding a project. To ensure your business makes the most educated AI project choice, contact The Burnie Group today.

Choosing the right technology

After determining which business areas would benefit the most and generate the most value because from an AI application, it is time to select which AI tool is right for the project. Choosing the right technology also means considering limitations. Many deep-learning applications can derive previously inaccessible and unknown insights from data, but they may require voluminous well-structured datasets and often labelling to do so. It is important to be aware of the abilities and limitations of each technology, and the specific data training sets required for successful implementation.

Choosing the right tool: The possibilities of process automation

Process automation is one of the most common and compelling use cases for AI. Robotic process automation (RPA), is a powerful and cost-effective way to improve process automation. Its traditional implementation has been limited to applying business logic to process flow in a high quality and efficient manner. However, traditional RPA is not well suited to dealing with more complex decisions that require judgment calls. As decision process nodes are common in many processes, an automated process flow often requires human intervention to make a judgment call. This can significantly slow process efficiency, creating bottlenecks. Artificial intelligence can automate many decision nodes previously immune to automation, thus improving overall process efficiency and cost-effectiveness while improving customer satisfaction.

Choosing the wrong tool: The downside of virtual assistants

While some companies race toward dominance in the virtual assistant field, such as Alexa, Cortana, Siri, and Google Assistant, many approach implementation of this AI application with caution. Despite the massive cost and time savings virtual assistants promise, there are a few horror stories of implementation gone wrong. For example, in 2017, Facebook announced high-hopes for their newest project: a virtual assistant known simply as M, built into Facebook’s messaging platform. The goal was to integrate Facebook Messenger with outside services. A user typing “can you call me?” would trigger M to open Facebook’s video and voice calling ability. As a stand-alone chatbot, M was designed to provide other ambitious services, like reserving a table at a restaurant or making changes to a flight.

Despite Facebook’s huge investment in the pilot project, it became clear that M had serious issues. The chatbot had a 70% failure rate, meaning nearly all inquiries were passed on to a human operator. M’s NLP capabilities were also seriously flawed; when two bots were tasked with trading items with each other, they quickly developed their own shorthand language, completely unintelligible to humans. Facebook scrapped the entire project shortly after.

Choose with caution

It is important to note that AI is not a hammer in a world full of nails. It is likely that, in the short term, choosing a technology that can manage simpler tasks will prove more effective than attempting to tackle the most ambitious project in your portfolio. Turning all customer interactions over to virtual assistants and chatbots is indeed a possibility that will only improve with time but consider the more feasible step of automating your internal helpdesk first.

As noted, identifying the appropriate AI tool is only half of the key to success; in order for the tool to be successful, it must be provided with high-quality data.

Learn more about how you can apply AI In your industry.

CONTACT US

Assessing the quality of training data

Artificial intelligence tools are only as powerful as the data they evaluate. The quality, quantity, and applicability of the data provided directly determines the quality of the results. Therefore, it is essential that data be prepared correctly, and the AI system trained adequately. Three main features must be carefully considered when selecting data: bias, labelling, and context.

There are many types of data bias that affect what data an AI application receives each of which can negatively impact the quality of the output of an AI system. The two most prevalent types of bias are exclusion bias and selection bias.

In exclusion bias, data is cleaned in some way. An online retailer, for example, may want to use an AI tool to gain insights about its customers. Let’s say the majority of its customers are from Canada, while a very small percentage are from China. The retailer may be tempted to remove the data about the Chinese customers as it wants to focus on how best to target its Canadian customers. However, insights from the AI application might reveal that the Chinese customers bought twice as many high-margin items as the Canadian customers and were more likely to be repeat customers. These important insights would not have been identified if the Chinese customer data had been excluded.

Selection bias occurs when the training data fails to reflect the actual target environment. A study by the University of Washington sought to illustrate how this type of flawed data produced inaccurate results. [1] In the study, a machine-vision system was trained to identify wolves and huskies in images. After training, the system still could not correctly differentiate between wolves and huskies. Although the deep-learning neural networks that power this type of problem-solving are considered to be “black boxes” (meaning the logic of their inner layers is unknowable), the researchers were able to identify the reason the system was failing: the data it was trained on showed huskies with a grassy background and wolves with a snowy background. The system learned to make identifications based on this correlation (see Figure 1). The data was flawed, or “biased,” because it did not accurately reflect the future data the system might encounter.

Figure 1. Raw data and the system explanation of a flawed prediction in the “Husky vs Wolf” task. [1]

Machine learning is capable of analyzing some data without labels. However, such unsupervised learning models are designed to produce more exploratory analysis than insights from data. Having carefully labelled data is usually the key to useful and effective AI applications.

Machine-learning algorithms are great at determining correlations. However, they are not as good at interpreting the surrounding data. For instance, an online retailer might use an AI adviser to recommend items to customers. The retailer may notice the system is over-recommending one particular item. Upon closer inspection, it becomes apparent that there was a spike in sales of the item a year ago due to a discount promotion. The AI system interpreted the high sales as a sign that the item was both popular and useful and therefore a good recommendation. To avoid this, the system would need to be provided with the context of the sales spike data in order to make a more accurate recommendation.

In selecting data, a business must consider all possible biases that might contaminate the data and thus affect any results from an AI system which has been trained by that data set. Labels must be accurate, and context provided before a pilot project can be launched. To learn more about proper data selection and preparation, contact the Burnie Group today.

Launching a pilot project

Pilot projects and proof of concepts are useful for testing out AI initiatives with high potential business value before implementing a project across an entire business. One business giant taking this approach is Pfizer. The pharmaceutical corporation is placing sizable bets on AI and has more than 60 on-going AI implementation projects, business-wide; many of these are pilots, while others have moved on to the active usage stage. As pilots graduate to implementation, new pilots are initiated. If a business plans to launch several projects or to test multiple AI technologies, like Pfizer, a centre of excellence (COE) should be formed to manage the projects.

A COE team should be formed of talented AI experts who can take initiative and develop effective AI systems. Management consulting firms, such as The Burnie Group, are often brought in to set up these COEs. An effective COE strategy ensures that the technical skills required to develop and manage AI applications continue to be developed and fostered within the company.

A strong COE also provides deeper insights into AI application implementation. Expert team members may be able to recognize how a small pilot might have a greater impact if moved into a broader application. The medical tech company Becton Dickinson operates a large COE that it calls its “global automation” division. The “global automation” team oversees all the company’s AI pilot projects, from identifying areas of opportunity to implementation and management of the application. It uses specially designed process maps to identify which business processes would benefit the most from an AI application; it then uses these maps to guide the implementation process.

This type of comprehensive business division, staffed by a team of experts, enables companies, like Becton Dickinson, to manage and introduce multiple AI applications to business processes seamlessly. However, before applications can be introduced, business processes need to be redesigned, and employees must be helped to adapt to their new, intelligent system.

Redesigning business processes

As AI projects are developed through pilot projects, potential changes to workflows become apparent. The division of labour between the AI application and humans will, inevitably, alter the traditional job tasks of employees. The AI applications themselves will require some level of human supervision, creating new employee tasks. Human labour will be liberated to perform more complex, judgment-oriented tasks and make ethical decisions. Regardless of the division of labour, it is important to remember that AI and human capabilities are not the same: each is required to leverage its particular strengths while compensating for the other’s weaknesses.

The mistake many companies make is avoiding process redesign entirely. They recognize the potential value of automating established tasks, but they ignore the opportunity AI?enhanced automation presents to improve processes themselves. This is particularly true of RPA projects. To avoid this short-sighted approach, businesses and their COEs need to evaluate the entire process. Each project design should be considered a draft, susceptible to change. By recognizing the needs of the user or customer, involving the employees whose work will be affected, and considering all alternatives and technological capabilities, it is possible to create a new and better method of doing things.

Managing employee adjustment

Having employees that both accept, understand, and even embrace AI is a long-term competitive necessity for any company. Many businesses have recognized this and are investing heavily in employee education. In 2017, 82 percent of companies said they planned to implement AI within the next three years. Of those, 38 percent said they were providing their employees with relevant reskilling opportunities. These opportunities may be provided to employees in the form of organized workshops, on-site training, or compensated online courses. These options should be tailored to fit the company’s needs. Just as not all AI fits all situations, not all AI education courses fit all businesses. Courses and workshops should be focused on the size of the company, its industry, and the type of data with which it deals.

Investing in employee AI education guarantees a smoother implementation process and an easier learning curve. It may also stimulate innovative thinking in a workforce. Proper education ensures employees feel empowered by the technology, not threatened by it as business processes are redesigned. An investment in employee education is also an investment in the success of the AI project, and, therefore, the success of the company. Increased employee understanding leads to higher data quality and improved results; it also paves the way for a smother scaling transition between pilot project and company-wide implementation.

Learn more about managing employee adjustment through technological transitions.

READ MORE

Scaling a project

A successful pilot project does not guarantee a successful company-wide rollout. To ensure success, business executives, management consultants, center of excellence team members, and other technology experts must work together to develop a detailed plan for scaling up the project. This scaling up almost always requires careful integration of AI with existing processes.

Scaling challenges

To scale up a project, a business must consider the project’s limitations. For instance, if the new AI application is highly dependent on a technology or skill that is difficult to source, it will be difficult to scale. The challenges to scaling up should ideally be addressed before or early on in the pilot project.

Even giants like Amazon face scaling challenges. In 2016, Amazon announced the opening of its first brick-and-mortar stores—Amazon Go—in Seattle. These stores would be part-convenience store, part-grocery store and be fully automated. A machine-vision system would identify each shopper and what they placed in their bags; the Amazon app on customers’ mobile phones would be charged automatically. Amazon promised “no lines, no checkouts, no registers.” However, Amazon was forced to delay the launch of the project for over a year, citing technology glitches and problems with the quality of the AI engine’s pattern recognition. The first Amazon Go store finally opened to the public in January 2018, with the second opening in August. Amazon has said it intends to open as many as 3,000 cashier-less stores by 2021.

Addressing challenges

Amazon intends to scale its Amazon Go AI project massively. However, they are exercising the appropriate amount of caution and planning. Too much, too fast, could easily cause the project to fail. Despite public pressure and expectations, Amazon took an extra year to develop its pilot project before rolling out the scaled-up model. Chief Executive Officer Jeff Bezos says Amazon is still experimenting with the form of the project. Many alternatives are being considered to determine what works best: a convenience store, a grocery store, or a grab-and-go style café similar to Prêt à Manger. In this way, Amazon is demonstrating a strategy for careful implementation and scaling of AI projects, while also paving the way for the cognitive-powered companies of the future.

Read more about how AI will shape the companies of the future.

How will artificial intelligence shape the company of the future?

Artificial intelligence technologies are slowly but surely transforming the way every industry does business. Companies that implement AI technologies now and develop the infrastructure to support more comprehensive AI systems in the future are setting themselves up to reap the benefits of AI’s abilities.

Even the most basic of AI tasks—extracting information from lengthy documents or answering repetitive, routine customer inquiries—have the potential to significantly increase value and reduce expenses across many domains. In the company of the future, most basic, time-consuming, and repetitive tasks will be performed by AI-powered applications. Human workers, rather than being replaced by machines, will be able to focus on completing much higher-value judgment-oriented tasks. The widespread collection and analysis of data—including the invaluable and voluminous data produced by IoT devices—will provide insights never before available, changing the face of business as we know it.

The path to this company of the future will not be without difficulties. Addressing public fears about workforce displacement, debating the ethics of intelligent machines, assessing risk, and assigning responsibility will undoubtedly affect the pace and direction of AI development. These issues will only be compounded by our inability to ever really understand the decision-making processes of “black box” AI systems. However, companies with the foresight and proper planning to implement an effective AI strategy early on will not only increase process efficiency and reap the benefits of untapped analytic potential but will help to bring about a new era of technological prosperity and productivity.

Develop your own AI strategy with support from Burnie Group.

CONTACT US