How Do I Create a Successful Artificial Intelligence Strategy?
Here's what 22 executives who are familiar with AI and all its variants said when we asked them, "What are the keys to a successful AI strategy?"
Join the DZone community and get the full member experience.Join For Free
To gather insights on the state of artificial intelligence (AI) and all its variants — machine learning (ML), deep learning (DL), natural language processing (NLP), predictive analytics, and multiple neural networks — we spoke with 22 executives who are familiar with AI.
We asked them, "What are the keys to a successful AI strategy?"
Here's what they told us:
Identify the Business Problem
- Begin by picking a problem to solve with machine learning, deep learning, and AI. Potential to solve with neural network feature sets. Have a single platform for analytics.
- Identify a business problem to solve. We focus on bots driven by AI to provide a great user experience (UX) and customer experience (CX). To build a good bot, you need strong AI. Bots must understand verbal (NLP), emotion, content, context, and intent. A bot development platform must include intents unique to the domain, runtime, and connectivity. Consistent tooling, runtimes, and management. You need a framework to bring mobile, AI, and bots together.
- Understand the business problem you are trying to solve. Amplify human cognition. Thinking about the business problem with a cognitive assistant benefits employees and customers. In the chatbot/conversational agent space, no one is getting past answering the question to determining the bigger problem someone is trying to solve. Conversations go deeper than answering questions. Help people work through the problem — help them see through their biases based on real evidence.
- Identify the right business case regardless of if you are using AI. Have the right skills to get the job done — in-house or available, highly advanced statistical modeling. Have the right data. Determine the type of analytics needed — prescriptive or descriptive.
- Look at and understand what you are trying to do and what business problem you are trying to solve. Take stock of what people call AI. AI/ML algorithms are looking for anomalies. Creating distributed training to detect anomalies. Descriptive analytics allows you to describe the world around you. Diagnostic analytics allows you to describe why something happened. We are just moving beyond diagnostic and descriptive analytics. Use the most advanced AI outcomes and work backward to solve the problem. In security, you are always analyzing the adversary. To get ahead, you need to imagine what can go wrong, have a sense of the operating environment, identify the most valuable assets you want to protect, and identify your weakest links. This requires a different class of algorithm with different training and data to get to predictive analytics.
- Maintain an internal state to claim prescriptiveness — know what to turn red and what to turn green. Make tactical changes at endpoints. Implement strategic intersegment routing. We work to be predictive and prescriptive in our AI. This requires a different class of input and a different class of algorithm that can learn, have feedback loops, learning appropriate and inappropriate activities. Optimize the algorithm and internal base. We have an internal memory or algorithms and neural networks maintaining certain states.
- Know what kind of problem you are trying to solve. Apply machine learning to help identify where the problems are and then determine if AI is an applicable solution. Start with something you know, go to ML, then to AI.
- At Barclays, our strategy started with simple use cases (like checking your account balance) so we could focus on building solutions that are easy to use and intuitive. The CX remains our number one priority. Over time, we will move on to more complex solutions (like handling customer disputes) while we continue to develop an infrastructure to complement AI with capabilities like cloud, big data, and natural language processing.
- Firstly, organizations need to understand what the business problem is that needs to be solved. All successful projects start with a problem. AI is a tool that can be applied to some problems, but it should not be a means on its own. It's always secondary to the business problem. Some of the most successful case studies have come from companies who truly understand the limitations (positive and negative) so they know how to leverage the technology in the best way. Our technology will have the biggest impact on business processes that involves a high volume of routine cognitive labor. For example, Junior Lawyers used to review thousands of contracts and enter data into a spreadsheet. This can now be done, at least partially, by a machine, leaving them to focus on higher value work.
- Start with the business outcome you want to achieve — whether it is to improve the responsiveness of customer service or to reduce labor costs — and then see if AI is the solution that will get you there. Success is all about real world outcomes. You may be building some incredibly interesting models to do analysis, but how is it actually going to be applied to the business — i.e. what is the real-world application?
- Many applications go through churn, buying, and failure. Figure out what you want to predict. We chose to focus on industrial IoT (IIoT). We need ML to teach how to solve problems.
Have the Necessary Data
- Watson for the enterprise. Metrics and metadata. Non-domain expertise. Anomaly detection, great dataflow, and databases. Foundation for real-time data processing. Domain expertise.
- Don’t be afraid of exploring your data. Get all of your data together in a data lake. Linking data adds value to data. Think of graphs connecting entities and associations. Come up with hypotheses. Ask and answer questions.
- Start with data and learning. We help businesses transform starting with AI/ML and a business model that allows us to transform in three areas: 1) Better leverage data, the new currency, to make better business decisions. Predict and address problems rather than react to them. Relationships — interaction with customers based on structured and unstructured data. How to use unstructured data to improve the CX. Ability to customize everything for the customer thereby being more relevant. Tie to intelligent automation. Apply AI/ML to help automate routine and beyond. Use internal data to make predictions on how to go to market more effectively.
- Make software smarter and easier to use. Enable AI application development. Provide digital mesh to embed AI from the core to the edge. This enables DevOps to inject AI into the infrastructure for real-time analysis. Ability to do smart ingestion, quick data wrangling, shifting, pivoting, splitting, and shifting. Look for columns that are similar and need to be aggregated. As you are doing smart data wrangling you have a record of everything you’ve done. Ability to get recommendations. Suggested recommendation to use AI behind the scenes to help users. Smart ways of provisioning nodes in clusters. Identify resources to tasks. Allow injection of ML into AI on the edge can create apps for embedding or calling with REST interfaces.
- Be clear on the business outcomes you want to achieve. Have the right data and a sufficient amount. How do you know what’s sufficient? We employ NLP to the data and the queries to determine if we have a good chance of success. Data scientists, ML experts — we use independent software vendors as needed for our clients. We have an ecosystem of partners. Projects are experimental by nature and may fail. There are no guarantees of success but by testing you learn a lot more than you knew going in.
- Take a vertically-oriented approach and customize to the domain. Use different elements of AI based on the problem you are trying to solve: Attributes of the product category with NLP (i.e. three-fourths sleeve, V-neck, crew-neck, boot-cut); Once you have the attributes understand what drove the consumer purchase (three-fourths sleeve, brand preference, fabric preference) learn the relative importance of attributes with ML. As you understand consumer preferences you are able to present content based on the attributes preferred by the consumer. Different features and attributes are ranked based on consumer importance.
Have the Right Tools
- Having the right tools. Knowing what you want to accomplish. Having specific goals. Starting simple. Having the right data to accomplish your goals. Machine learning can learn to live with dirty data. For example, acoustic data varies based on if it’s coming from a mic, versus a phone, versus something with a lot of background noise. AI/ML has to deal with real data and real data is dirty.
- Problems are the same but the tools and the skills are different. What AI solution you use depends on the nature of the problem — gigantic data set to mine; build an autonomous vehicle; everything in between. The right tool and technology depends on what you are trying to accomplish. Think about how the application of technology provides a competitive advantage. Know what technology to use when then how to effectively use the technology without having a Ph.D. in statistics. Know how to leverage your domain knowledge. You cannot apply one solution to solve every problem. Empower developers with abstractions, tools, and platforms to solve problems with unique solutions. Understand how different versions of AI can be applied to different problem types.
Keep Your Models Current
- Do an assessment to figure out where AI can be beneficial. Identify what steps your most tedious and time-consuming processes are. Identify where you have the greatest risk exposure for failing to fulfill compliance issues. Determine how you will implement the resources you have. Continuity — continuous management to deal with all of the changes, evolution of skillsets, and new technologies. You must commit to ongoing support of your AI initiative or it will fall apart quickly given how fast everything is changing.
- It’s important to focus on constantly improving the infrastructure that powers AI. This will lead to more accurate and faster results. Having software run efficiently is a key benefit. We prune 95 percent of unnecessary processing threads so there is less data to analyze. We average results of about 99 percent on detection rates (vs. 80 percent amongst the competition).
What are your keys to a successful AI strategy?
Here’s who we talked to:
- Gaurav Banga, CEO, and Dr. Vinay Sridhara, CTO, Balbix
- Abhinav Sharma, Digital Servicing Group Lead, Barclaycard US
- Pedro Arellano, VP Product Strategy, Birst
- Matt Jackson, VP and National General Manager, BlueMetal
- Mark Hammond, CEO, Bonsai
- Ashok Reddy, General Manager, Mainframe, CA Technologies
- Sundeep Sanghavi, Co-founder and CEO, DataRPM, a Progress Company
- Eli David, Co-Founder and Chief Technology Officer, Deep Instinct
- Ali Din, GM and CMO, and Mark Millar, Director of Research and Development, dinCloud
- Sastry Malladi, CTO, FogHorn Systems
- Flavio Villanustre, VP Technology LexisNexis Risk Solutions, HPCC Systems
- Rob High, CTO Watson, IBM
- Jan Van Hoecke, CTO, iManage
- Eldar Sadikov, CEO and Co-founder, Jetlore
- Amit Vij, CEO and Co-Founder, Kinetica
- Ted Dunning, PhD., Chief Application Architect, MapR
- Bob Friday, CTO and Co-founder, and Jeff Aaron, VP of Marketing, Mist
- Sri Ramanathan, Group VP AI Bots and Mobile, Oracle
- Scott Parker, Senior Product Marketing Manager, Sinequa
- Michael O’Connell, Chief Analytics Officer, TIBCO
AI Machine learning Data (computing) Database neural network NLP mobile app Analytics
Opinions expressed by DZone contributors are their own.