Agile AI

Chapter 4. Summary

As you can see, artificial intelligence (AI) has many moving parts.  A practical Agile approach to AI focuses on the right team mindset of flexibility, the right set of tools, and the right set of team skills. Across the first three chapters, we talked broadly about those ideas as well as identifying the proper uses cases, organizing your data, and changing your business by teaching it to trust the products you’re building. We want to take a few last words to close on the significance of these points so that you can go forth leading your team to achieve its most ambitious AI projects.

Use Cases

Integrating AI into your business happens one use case at a time. First, work to identify the appropriate use cases: use workshops and design thinking, and ideate as a team to get many perspectives and insights. Focus on business problems; don’t get caught in the trap of building technology for the sake of technology.

After a project, conduct retrospectives on those use case implementations. What worked well? What could’ve been handled better? Identify patterns in your use cases, and share those across the organization. Ideally, create a Center of Excellence for AI patterns within your organization.

Data

You can’t deploy AI without first getting your data into shape. Those are table stakes, the absolute basics. Recent industry surveys about AI adoption in the enterprise have shown that the majority of firms become bogged down in the technical debt they must resolve before they can share data across divisions. Also recognize that data sharing runs counter to the inertia that tends to arise in large organizations. When people have complex tasks to manage, it’s generally simplest to break those into smaller, simpler tasks and then keep the parts separate. Organizations respond to complexity in much the same way over time, establishing silos between divisions. That’s especially true when it comes to data management and access controls: the easiest answer to a request for data access is “no.” However, the easiest answer is not always the best in the long run.

Some people have called data “the new oil”—an economic good driving enormous value—but it’s better to think of data as an investment portfolio. Your data science teams help manage those investments in data, and they help realize equity and yield from them.

One idea that’s been making the rounds is data democratization: an idea in which nearly everyone in a company can run queries, create analysis, generate reports, and so on. Obviously, not all the data in a firm will be made available to every employee, given that there will likely be confidentiality concerns to balance, as well. Data democratization has drawbacks: making claims about data analysis depends on having enough of a statistics background to get it right—although this is where data science teams can help coach and amplify results from the rest of the organization.

There’s a related term to data democratization: citizen data science. This is when people throughout your organization become involved in developing data insights; if you think of data science as a continuum in which people on one end are just beginning to “upskill” into these kinds of roles, and experts on the opposite end are continually learning new skills for their data science practice, citizen data science makes a lot of sense for a large organization. Drawbacks and critiques aside, both of these approaches can make a lot of inroads toward breaking down data silos.

Tools and Process

To get a job done well, you must have the appropriate tools and apply the appropriate process to them. Embracing open source is an excellent first step because there’s so much available and robust developer communities to learn from. Also, don’t limit which tools a data science team can use: let it evaluate alternatives and use what’s appropriate for the people and the challenges involved. Bring in new tools to support new talent while utilizing the existing approaches and tooling to support your existing talent. For example, some of your staff will need to use Python and machine learning frameworks, whereas other people might be more focused on SQL and reporting.

Be agile in your approach! This means keeping in mind that effective work is not about building the best possible model for one use case; it’s about the cadence you establish for creating many effective models for a range of use cases. Invest time to prove value; however, time is typically the most expensive resource in business, so don’t spend too long to prove value in a given use case. Time is not your friend. Joni Rolenaitis, Experian CDO, said it best, “Go agile or go home.”

Mindset

You need to foster the proper kind of culture for a data science team so that the members can develop an effective mindset for the challenges they face. Data and dogma don’t mix well. Data science teams are often in the “hot seat” because they’re the gatekeepers for analysis that disproves what others in the organization might take as givens. On a data science team, we’re fostering an analytic mindset, one that’s driven by curiosity. To support that, create an environment in which people can collaborate with their peers, where criticisms coming from outside the team are constructive, and in which opinions—both inside and outside the team—are respected but not worshipped. Ultimately, data and reproducible analysis, and an extra dose of curiosity, must rule over opinion.

Diversity helps. Diversity builds a wide range of skills and experience across a team, which combine to produce effective solutions. Data science is inherently interdisciplinary. Diversity also helps temper opinions so that they don’t turn into bias.

In terms of organization and planning, some aspects are crucial for establishing an effective mindset:

• Keep your sprints quick when you need to prove or disprove value, but give your people time to invest in a problem that needs to be resolved.

• Develop team deployment models that fit your business needs. Some data science practices use a hub-and-spoke structure to get their members involved in use cases while still closely communicating. In other firms, there’s an “embedded” model of rotating people into product teams, but bringing them back together periodically to compare notes. Experiment with what’s best for your organization.

Integration and Trust

Integrating AI into your business will probably have at least two dimensions. First, there’s the impact on your internal business workflows. How does the organization arrive at decisions? How should automation affect the decisions that people make? Second, there’s the decision of where AI fits into your roadmap for customer solutions. Develop your analytics and machine learning work with these end goals in mind: how will they be consumed?

In either of those categories—impacting internal operations and product decisions—trust is key. To build trust in AI, people need to be comfortable with automation systems and confident that they produce good results. Good results go beyond the accuracy of predictions; your organization and customers must also be comfortable with your data science team’s practices addressing concerns about fairness and bias, developing models and supporting workflows in which decisions can be explained and explored, and designing for security and robustness. There’s a growing body of open source projects to help remediate concerns about AI trust, and related foundations such as LFAI; for example, the Linux Foundation’s AIF360 toolkit provides a range of advanced statistical tooling for measuring and correcting bias in training data.

Conclusion

To close this book and pull together the components described throughout it, your challenge is to build a data science strategy for your organization. As Figure 4-1 shows, we approach this challenge from two directions. We want a culture that supports discovery, diversity of skills and perspectives; embraces new ideas and insights about your business; and moves quickly to prove or disprove them. We also want an effective data platform that supports a data science team that has diverse needs, along with scaling for customer use cases.

Figure 4-1. Strategies for accelerating AI

The upper-right corner of Figure 4-1 faces closest to your customers. AI trust is paramount there, and to reach that, you must build an appropriate culture for your data science team and for their relationship with the broader organization. Your data science team needs the skills, the diversity, the mindset, the mentoring, the curiosity, the domain expertise, and continuous learning to take advantage of machine learning and integrate data products to meet your customers’ needs.

The lower-left corner of the diagram is about your data and investing in its value. Break down the silos, collaborate across teams, and get the appropriate tools and processes in place to unlock the value of your data. Make these platform capabilities meet the needs of your data science team’s culture and skills, both for immediate purposes and as their curiosity and insights and organization learning grows. Also, make your platform capabilities meet your customers needs as your business grows.

Ultimately, these two approaches must meet in the middle. Throughout this entire process, there are feedback loops that you can recognize and nurture so that both your platform and your team improve continuously. With these practices, skills, and culture together, you can be agile in building AI in your organization.