BACKGROUND IMAGE: stock.adobe.com
Today, AI is the Wild West of IT projects. It started with business intelligence, where data was collected across the enterprise and shipped into a big database. Reports and dashboards were then built on top to help shape our daily decisions.
Currently, enterprise adoption of AI includes looking at operational data hiding in every nook and cranny of the enterprise and making that data actionable. Organizations are now focused less on the technology and more on business outcomes and what can be derived from placing learning algorithms on top of the data we have.
People have also become skilled at computer vision and speech analytics to create best practices on converting speech to text and text to speech, as well as identifying objects and actions in images. These advancements, coupled with the idea that machine learning in business processes creates market advantages, have pushed companies toward AI.
When a colleague and I researched the use of AI in real-time communications, we interviewed companies working with AI -- from AI-first startups to large enterprises that needed to start AI initiatives. The survey uncovered some challenges that organizations face when adopting machine learning and AI in communications.
1. Finding the right talent
One of the main challenges for AI adoption is finding talent conversant enough in machine learning. The talent pool is relatively small, and large cloud vendors, like Amazon, Google and Facebook, are attracting developers and data scientists with high salaries, bonuses and highly rewarding work.
Enterprises can attract talent in one of three ways:
- Open offices in smaller technology hubs around the globe. The further away from San Francisco you go, the less of a fight over machine learning talent you will face.
- Compete in the market by offering higher salaries and benefits to lure experienced AI developers.
- Train the existing workforce. One organization we interviewed said all developers were given the opportunity to take online AI courses, which led to 10% of their developers becoming more familiar with AI.
2. Making business sense
Adding machine learning to a healthcare use case is different from adding machine learning to a social network interaction. While both these examples may use similar algorithms and techniques, you'll need some domain expertise in the market itself.
In contact centers, for example, the dynamics of a sales conversation are quite different than a partnership discussion. In these instances, AI could gauge who speaks when and for how long. Knowing these differences and nuances of the business is just as important as knowing statistics and machine learning.
This is doubly true in product management roles when people need to make decisions on product roadmaps, priorities and features. Filling these roles is critical to the success of such initiatives.
3. Collecting the right data
In many cases, machine learning algorithms need to analyze data that is not collected or not easily accessible. This includes data in on-premises databases, on different machines and systems and in legacy proprietary applications or data that never gets stored anywhere.
Data doesn't need to be collected at all times in order to get effective results, but it's important to have an understanding of available data and how accessible it is.
4. Cleaning data
Operational data can be quite dirty and ill-equipped for machine learning because of errors or inconsistent or incomplete data. For example, background noise, packet loss or other interference on a phone call can affect how a speech-to-text algorithm captures spoken words.
Such data needs to be cleaned before it can be used by machine learning algorithms; otherwise, the models created will not reflect reality or the optimization you are trying to reach. Cleaning data varies from manually going through the data by filtering inputs and adding conversion rules to changing the data used to create the model for a specific scenario.
5. Iterating on experiments
Once you have a team in place and start to experiment with machine learning, you'll probably ask the following questions: Do you have enough data? Is the data clean? Is the data representative? Do you need machine learning, or would a simple rule suffice? These questions will require more data for the machine learning algorithms, but the needed data may not be readily available if it's in another database or needs to be cleaned.
These changes in data requirements can cause delays in the creation of models that can take days to months, depending on the nature of the data and how it's collected. With each experiment, you will find new insights. These insights will spark more questions, which require the collection of additional, different data. Adding AI in communications to an enterprise means having the patience to collect, experiment, examine and iterate.
6. Deciding how to use AI in communications
With all that can be done with AI, one of the biggest questions is where to start.
You can use AI in communications to improve and optimize your internal efficiencies, from finding root cause failures before they happen to routing decisions for voice calls throughout your infrastructure. You can use AI to add direct value and new user experiences for your customers, such as providing more accurate suggestions or using chatbots to handle support issues.
You can also create new products and features that couldn't exist without AI, such as real-time transcription and translation of conversations or authenticating people using facial recognition.
As a starting point, you need to pick AI projects that bring high ROI with little risk and costs. But sifting through the options and knowing which projects are more feasible are not always easy.
7. Deciding on the organizational structure
Where do you place the people in your workforce who know and understand AI? Some enterprises create an excellence group in the chief technology officer's office or R&D teams that drive AI-related initiatives.
Other enterprises go for a bottom-up approach -- first, identifying an area or two where machine learning experiments can take place, fostering them over time and seeing where that leads the organization.
Some enterprises take a top-down approach, starting with a specific business need and use case and then figuring out how to proceed. Often, this approach leads to outsourcing to external consultants and vendors that specialize in AI to assist in the definition and implementation of the use case.
8. Not all data needs machine learning
When starting on the road toward machine learning, most of the effort is to collect and make meaning of operational data that was lying around idly. That data can be put to good use without resorting to machine learning.
Since the people looking at the data and making most of the decisions are data scientists, they might skip such opportunities to improve the business and try to find other, more challenging tasks elsewhere.