We have a tendency to blame technology when things go wrong. I’m the first to admit that after years of working in the technology industry I’ve become more and more annoyed with the technology I use. As artificial intelligence (AI) capabilities have emerged in my smart phone keeping me on schedule, telling me how to get somewhere, or generally keeping me in line, I’ve gotten conditioned to technology just working. Except when it doesn’t. That’s when I want to throw that phone, espresso machine, laptop or home security pad into a blender. (Yes, it was a rough morning.)
AI pioneers have provided us with a glimpse of and conditioned us to ambient AI making it hard to break up with each other. They have also set a very high bar on our expectations of what AI should do for our businesses. But, let’s understand, Google was able to do this after two decades of research, curating collections and observing our every move. Apple too has tracked our app usage, music preferences, and daily lives through its iCloud. And Facebook sees our public and private conversations, what we share, and our personal opinions. Creepy, yes, but that is another conversation.
The point is that enterprises embarking on AI need to radically shift their approach to technology adoption and analytics. This is not a plug and play and bolt on strategy. It takes work to go from POC to a capability that comes close to our expectations of AI based on our consumer experience.
AI begins its education through observation, not instructions. It needs vast amounts of data to establish domain expertise. Even if you purchase an AI solution that promises that domain expertise upfront so the data and training is reduced, it still needs YOUR DATA and YOUR CONTEXT to do the job. You aren’t programming a robot, you are building a relationship.
And so, it is not that AI is immature for enterprise use or value, necessarily. That was already proven by Google, Apple, and Facebook. The issue is that we, the enterprise, are the problem in the relationship for two reasons:
- We ignore the Observation Principle. AI pioneers put in listening posts in their solutions long before they turned on intelligent services. They took the time to understand intent, behavior, personalities, and expectations. Projects to recognize cats was the stepping stone for expanding use and interaction with image content. Companies like Crowdflower built intelligent data quality capabilities by watching humans augment and fix data on a gig worker platform. We need to orient AI implementation around what the AI system till do and allow it to watch and learn.
- We are really bad at data. Start any analytic project and the first question is what data sources are needed. The next question is what do I need to do to prepare the data. Where we get the data in some ways is less important than what data we get. If AI needs to observe, raw data without context is a really bad school book. The other issue is that most organizations are still in the throws of executing on data strategies. Forrester’s Data Management Playbook Benchmark study showed companies gave themselves low scores across the data management board. When they embark on AI, the deficiency in information competency becomes more evident.
These are not insurmountable challenges. They just require companies to:
- Define business milestones. Think about what the AI system should do and how to know it is doing a good job. That is always the goal with technology. The difference is understanding what value will be delivered and where. Even if the intent is to have a virtual agent directly interacting with a customer, the first phase may be only to provide insight or sit side by side with a live agent.
- Create an education strategy. Recognize the types of knowledge the AI system needs from recognizing objects, to understanding language, to being able to communicate, make decisions and actions to take.
- Create the curriculum. There is an order to how AI will learn that mimics the way humans take in information, understand, and determine what to do. To be effective at reading an X-Ray, part of learning may be to see an anomaly on the image and compare that to a CAT scan to detect a tumor accurately, just like a doctor would to confirm a diagnosis. Introduce information logically to educate AI from foundation to expert.
- Be realistic with result horizons. Poor early results is not necessarily and AI killer. It just might mean you need to add additional training stages. Promises of results in weeks might be true for insights where humans are still the gate keeper on if those insights are used. But, launching an independent intelligent agent can take up to 2 years as AI progresses through its education. This is where business milestones right size expectations for value delivery.
- Don’t stop training. As humans, we learn through additional experience but also by continuously sourcing information from blogs, articles, books, and conversation. Continuously add and expose your AI system to new scenarios, new insight, and allow it to collaborate more widely with other bots and people. This keeps your AI system a relevant virtual worker.