The AI Organization

AI will transform every dimension of business, from operations to customer engagement to products.

The digital transformation started decades ago, when companies used software to create “systems of record.” Keeping records digitally extended into “systems of engagement,” expanding into customer relations. These systems generated enormous amounts of data, which now must be optimized with new “systems of intelligence.” Well-designed and deployed, such systems should provide insights and make outcomes as efficient as possible so businesses can make better decisions. No one called them artificial intelligence until recently, however, when they became more sophisticated. Today, AI applications recognize images, transcribe speech, translate languages and interpret texts – faster and with better accuracy than humans. However, AI has limitations. It hasn’t achieved general intelligence, which requires abstract reasoning, emotional intelligence, conceptualization and self-awareness. For instance, supervised learning algorithms require immense amounts of data to provide meaningful outputs. AI capabilities fall into three categories:

1. Perception – Vision, Audio, Speech and Natural Language.
2. Cognition – Regression, Classification, Recommendation, Planning, Optimization and Pattern Recognition.
3. Learning – Supervised, Unsupervised and Reinforcement Learning.

Over time, AI will improve its capacity not only to process information, but to build on perceptive and cognitive abilities, “sensing” the world around it and deriving meaning from that stimulus. While many improvements exist in natural language processing systems, machines do not yet have true understanding of what they hear or say. AI will transform how organizations operate, connect with customers, enhance products and engage employees. 

AI remains experimental and iterative, but still has applications in flexible systems.

You can deploy AI across your organization’s existing applications – specifically, systems of record, which form the foundation of most enterprises, and also systems of engagement, which are usually bespoke or modified out-of-the-box applications. “Customer-facing” applications can use AI to track number of users, time spent, conversion to sales, and others. In “employee-facing” applications, AI can streamline and automate tasks. 

“AI opens up new ways of interacting that can make our applications more engaging or accessible to more users.”

Assess which applications in your organization will most benefit from AI infusion. Determine relevance, control, impact, and the amount of time and effort it will require. Two common and popular infusions are in recommender systems and personalization. Users engage with applications that observe and filter their activity to provide recommendations that they base on their preferences. Augmented reality is another way to engage users by providing remote assistance and training. Finally, “cognitive search” can allow users to “extract additional attributes” from their applications – using image and speech recognition, for instance.

Conversational agents (or chatbots) are often the first step companies make into AI-assisted systems. They are a popular way to drive customer engagement and reduce human-to-human interaction for mundane or repetitive requests but also should be able to provide recommendations. Start with narrow chatbots for simpler interactions, then integrate them under an “omni-bot” that will sort through requests and direct them to the proper chatbot. Deep learning can make these agents more resilient by training them using past customer interactions, improving their predictions and outcomes.

Train business units to implement AI solutions horizontally and vertically in their enterprises.

Business units and technology departments should coordinate their practices to rethink their organization’s primary functions. Create inventories and categorize processes under a framework that includes the following zones:

• Incubation – Evaluating opportunities to expand business using technology. This is long term and therefore more agile.
• Transformation – Choosing opportunities to develop and scale that will push the organization beyond “business as usual” practices.
• Performance – Generating revenues that rely on ROI and investment priorities.
• Productivity – Optimizing effectiveness while reducing inefficiencies.

These zones will be the venues in which tech and business units can determine where and how to deploy AI. Consider carefully the advantages of bringing AI into various business-use cases, especially in the incubation and transformation zones. Because AI is iterative, its life cycle should have metrics to assess its impact, its quality and its actual implementation to improve future iterations.

“AI is more than just a technology; it’s a whole new approach to applications, business and people that requires a cultural shift in the organization.”

Horizontal processes (HR, marketing, sales, IT) and vertical processes (manufacturing, health care, retail) require different approaches. Horizontal processes span industries and are more amenable to standardized applications. AI can provide great insight into marketing activities, for instance. It can use natural language processing to listen to customers, observe the competition and analyze feedback. It also can use pattern recognition to follow trends on applications and optimize targeted customers. Beyond just customer engagement, AI in horizontal systems can transform sales, finance and HR with exceptional classification and forecasting capability, as well as improved communications between internal and external stakeholders. Vertical processes are unique to their industries, and AI has more potential for disruption in them. They require “deep exploration” between tech departments and business units in organizations. For instance, the Internet of Things (IoT) will transform manufacturing. AI perception, which sensors deploy, will help with maintenance, quality control, industrial safety, automation and supply chain management. Similar varieties appear in applications in finance (fintech) and retail sectors. Where AI might have the biggest impact, however, would be the public sector, where efficiency remains elusive and affects other sectors with tedious bureaucracy.

Employees are the heart of the AI organization.

Technical departments and business units benefit from AI in their processes, but an organization doesn’t really undergo transformation until employees experience empowerment with AI. Past technological revolutions follow this “democratization.” Soon, employees will become “citizen data scientists,” using simple tools to participate in the AI revolution. It requires fulfilling three criteria:

1. Democratization of knowledge – Currently, data are disseminated in complex ways across organizations, and are inaccessible to most employees. They need knowledge that is structured (graphs), semantic (clearly defined attributes) and consolidated. 
2. Democratization of AI consumption – Build AI into existing programs, such as Excel, to augment employee workflow. For instance, a sales consultant could use AI capability to identify customers who might buy a certain product.
3. Democratization of AI creation – Using “transfer learning,” a nontechnical user can customize a model that a data scientist trains.

Already, tools and interfaces exist that will help more nontechnical users build AI into their processes. Automated machine learning (AutoML) takes a dataset as an input and identifies the “best machine learning algorithm, parameters and data transformations” to prod1uce a desired output. Azure Machine Learning Studio gives users more control, though it does require some knowledge of machine learning. Finally, the “machine teaching” approach allows the user to transfer knowledge to the machine by breaking down processes into small units, generalization and examples. For instance, automation could be accelerated with Robotic Process Automation, where the machine observes and imitates users interacting with applications. This could improve productivity massively in fields such as data entry.

AI development teams must be agile and independent, yet maintain strong connections with the business.

It is tempting to use the project-based management approach to AI development in your organization, where you identify goals and set up teams with deadlines and deliverables. However, because AI is evolving, more relevant would the product-based approach, which focuses more on the needs of the user. Even more salient would be a “platform-based” management process, where you consider not only final users, but how every product serves as a “foundation for other products in your organization.” This helps with sharing and integrating processes and information and reduces redundancy.

AI development teams, using the platform-based management configuration, must necessarily have some autonomy, but also operate with and understand the organization’s business goals. These Machine Learning Ops (MLOps) manage the “continuous integration process” and use telemetry to track and trace issues in real time. The MLOps Loop demonstrates this integration and dissemination process:

1. Definition – Business stakeholders working closely with developers to determine business needs, experiment with models and acquire data.
2. Development – Implementing the process (preparing data, modeling and training). Solutions are incrementally integrated and monitored.
3. Operations – Managing the system in production and operations is part of the continuous deployment in its various iterations, and provides feedback via telemetry.

The AI team should be leaders in all aspects of AI strategy and deployment in an organization. They serve many functions beyond the technical, such as research, AI architecture, AI production, services to the MLOps teams, and training across the organization. It is vital that the AI team have a strong relationship with the business, even while maintaining autonomy in rolling out AI processes. The CEO must also be fully committed to the transformation and reinforce the AI vision.

AI is only as good as the data that firms use to train it, and the data must be diverse and rich to be meaningful.

Data are created everywhere in organizations – which need strategies to manage that data, particularly when cultural change is occurring. The “data estate” includes operational data (CRM and ERP), but also the following:

• Unstructured information (scanned documents, photos, contracts, etc.).
• Usage data which systems of engagement (SoEs) generate, such as websites and mobile apps.
• Data from IoTs (sensors, controllers, etc.).
• Employee data (emails and minutes from meetings).
• Global and third-party data.

For years, data have resided in system siloes, inaccessible to other systems. AI requires a very different approach: as much data from as many diverse origins as possible. Operational Data Stores (ODSs) and data warehouses are good resources, because they store data “as dimensions and facts,” making it easy to aggregate and train AI models. But that approach lacks scope. Enormous data “lakes” have “raw” data, ideal for feeding AI but difficult to interpret in an unstructured state. Data hubs can act as conduits to these different data storage systems, and because they are a strategy, not a technology, they are very flexible. A data hub can add semantics to unstructured data, can copy or just refer to data storage systems, and centralize governance and security.

“No matter what scenario you are targeting, chances are that without relevant data, you won’t be successful in delivering an AI solution.”

No amount of data in the world will help a company that isn’t data-driven. Organizations must embrace three core behaviors: generating, sharing and leveraging others’ data. They form a loop, and telemetry provides feedback to make the loop robust. With cooperation at every level across departments, data estates are gold mines.

Upskilling employees with AI training across departments will reinforce its relevance.

Many call AI the fourth industrial revolution, believing it has the potential to transform and redefine processes in organizations. Managing this transformation requires skills a lot of companies don’t have, and they have to be deployed in technical departments, business units and among employees. But you don’t have to train an army of AI experts. “Lower-friction” solutions exist through vendors like Microsoft and Google Cloud that can assist tech departments – though eventually, they will need to learn to build AI models themselves. Set up relationships with universities and research institutions.

Organizations will need multidisciplinary teams to facilitate symbiosis between tech departments and business units. Business units need “AI champions” who don’t need to build AI, but must know what to do with it. Finally, employees need training to absorb AI knowledge and apply it to their work. They will need to understand basic concepts even if they aren’t building systems.

Most companies start with upskilling their technical teams, and many technical training resources are available, including Microsoft AI School. Business units can also access Microsoft’s AI Business School to get the knowledge necessary to optimize AI applications. “Train the Trainers” programs create “catalyzers”: employees who learn about AI and share their learning with their peers. Most importantly, when hiring for an AI organization, leaders should seek employees who are “open to an agile environment” and are comfortable with the iterative process AI demands. They should be diverse, collaborative and passionate about what they do.

Follow the six principles of responsible AI: fairness, reliability, privacy, inclusiveness, transparency and accountability.

AI is a powerful entity, and you must deploy it responsibly. Organizations must get on board, developing AI that employees and customers alike can trust.

“Infusing responsible AI into the development life cycle requires the involvement of the entire organization.”

Microsoft created its six principles very early in its AI journey. They have guided the firm and its partners since. They are:

1. Fairness – Training data may have human-generated biases that affect how AI functions. An AI model trained on mostly Caucasian faces for facial recognition might not work on other races. To avoid biases, domain experts should assess and make certain the data are representative and fair.
2. Reliability and safety – AI has great potential to improve people’s safety, but it is not mature enough yet to be reliable in high-stakes situations. Testing and monitoring systems prior to implementation is crucial, especially since it uses probability in its algorithms. Domain experts can identify when systems are not performing to standard, and human intervention is allowable when failsafe occurs.
3. Privacy and security – Like other data-driven technology, AI will need to ensure safe collection, use and storage of personal data. It is also vulnerable to hackers because of its automatic learning processes. For instance, Microsoft endured an embarrassing incident when its virtual assistant (“Tay”) fell victim to hacking and “learned” hate speech.
4. Inclusiveness – Access to technology for people with disabilities has been an ongoing challenge, but with new developments in making access “context-dependent,” the solutions will be more general. Inclusive design starts with identifying mismatched experiences, fostering empathy and extending accessibility technology to everyone.
5. Transparency – AI is a unique technology because deep learning relies on algorithms that make decisions according to “inferred rules.” Their models are “black boxes,” with no one knowing how they made their decisions. Domain experts would have to monitor AI decisions to “debug” errors. It is not good for end users if AI can’t explain its conclusions about why, for example, a loan application was denied.
6. Accountability – As AI systems become more autonomous, it is tempting to offload responsibility for errors on its processes. This is unacceptable. Furthermore, if developers and users fall into “automation addiction,” they will cease to recognize problems over time because of complacency. Every decision AI makes must be traceable and subject to continuous monitoring.