11 min read
Don't be fooled by the hype over AI.
Just because there is a lot of hype doesn't mean there isn't huge potential value in AI. It would be a shame to miss out on the very real benefits of using artificial intelligence in your business, simply because the idea has become fashionable.
How can you put AI to work to your advantage? There's value in leveraging AI when you:
Turns out, it is just as important to have good domain knowledge as it is to be able to work with fancy algorithms, statistics, and advanced programming. To get real value from AI and machine learning, you need to recognize the points in your business where machine-based or machine-assisted decisions can make a real difference in efficiency, cost savings, speed of response, or delivering a new line of revenue. And to do that, you not only need to know how to build models to analyze data and derive insights, you also need to know your business. That's where domain knowledge pays off. You may not be a data scientist yourself and so you plan to work with a data science team within your organization or go outside to hire data science as a service. In any of these cases, in order for the project to pay off, someone in the process must know not only the specific business goals but also the challenges, pain points, and bottlenecks in your business. These often are the situations in which AI can make a positive difference. And remember: the people on a team who implement AI applications may not be experts in the business itself. You need domain experts as well as modellers.
Target class #1: Things people don't do or don't do well. You may recognize things that you'd like to set in motion, but people don't do them yet because they're too complex or there's too much data to analyze accurately or to analyze in a reasonable time window. Predictive maintenance in a large industrial setting is an example of this class of AI targets. Monitoring various parameters via data from many sensors placed on large equipment, such as drills, pumps, wind turbines, or manufacturing equipment, not only provides in-the-moment alerts for troublesome behavior, it can also provide the basis for finding problems before they happen. Often there is a “failure signature” in data that may occur days or weeks prior to a potentially catastrophic event, and if this can be recognized, corrective action can be taken before the failure occurs. But identifying a failure signature in huge amounts of IoT sensor data is just too much for a human to do, and that's where AI becomes valuable.
AI and machine learning models may discover a predictive failure signature in IoT data that can be used to prompt maintenance and thus avert a failure event. (Image courtesy of MTell, used with permission.)
In other situations, AI may be used to augment human decisions, making them more accurate and/or more timely. AI-assisted analysis of medical imaging is one such instance. The final diagnosis may still be made by the physician, but with a much faster response time and possibly a higher degree of accuracy if image data is first analyzed by a machine-based model. We have customers in healthcare who have applied such techniques and, in this way, reduced the time required for test results from days to hours. This, in turn, changes the way a doctor can plan therapies and interact with patients.
Target class #2: Things people do but would prefer not to. Another way AI can be a benefit is to automate decisions that are currently done well by humans but are uninteresting or unnecessarily expensive. Automating these steps through a correctly implemented AI program can cut costs and free up the person who would have done the boring task to do something more engaging and often more valuable. Some examples are putting chatbots to work for the first level of handling customer calls. If done well, this first-contact automation can improve customer experience by avoiding long call delays and by providing a response for mundane requests faster, while efficiently moving more complicated requests to a human support specialist. Of course, the key is to do it well. Another example in this class of situations of good targets for AI is quality control along an assembly line or in produce delivery.
This requirement is essential for putting artificial intelligence to work effectively, but there are less clear-cut guidelines about how to do it correctly. The challenge comes, in part, because you don't always know ahead of time exactly which data will be useful. What's the best advice for good data habits? The best way to succeed is to address the issue on two levels.
First, set up your organization to have the right data resources and data access. In order to do so, work with data infrastructure designed to make it feasible (in terms of cost and effort) to save a lot of data in relatively raw form, that has the capability to version data, such that you can go back to specific sets of training data with confidence and have infrastructure that lets different data science teams and applications directly access data without having to copy out to another platform for modeling and analysis. It's obvious why having a data platform that makes it possible for applications to run directly on data in place can improve the productivity of your data scientists. The reason you want to have access to raw data may be a little less obvious, but here is why: data scientists will extract features desired to train and use a particular set of models – that's in part why you want a practical way to version data – but who knows which features may be valuable for the next project? You don't want to come up with a clever and potentially valuable AI application only to find you've discarded the data that would have held the insights you're looking for. You may not know what specific data you'll need until long after it's been collected and even possibly overwritten. For example, it's one thing to detect a fraud event soon after the event occurs, but there may have been a time before the event when the fraud could have been prevented – that's much more valuable, but only possible to learn if you've still got the data.
The time at which fraud might have been prevented is different for each fraud event. In order to develop an AI application to discover these prevention points, you need retrospective data, such as an event-by-event history for each user. With the right technology, a stream can serve as a replayable system-of-record. [Image based on Figure 3-2 in the book AI and Analytics in Production by Dunning & Friedman © 2018, used with permission.]
Second, recognize the importance of having the right data for a particular application. Successful AI is not just about asking the right question and building high performance models; it's also about making certain that the data fits the situation. Here, once again, domain knowledge can pay off, both in terms of working with people who know the topic and the business but also with people who are data-aware.
The main lesson here is that unless you have a practical way to take action that ties into real business goals, you won't get value from AI, no matter how good the model is. And keep in mind: a report is not an action. The organizations that are reaping the benefits of AI and machine learning are those who recognize how to tie the results of these machine-based decisions into effective actions. Examples of actions tied to practical business needs are to automate adjustments of an automated manufacturing process based on feedback from an AI model, to provide personalized discount offers or recommendations that improve customer experience or reduce churn, or to have an AI-driven web page that reacts and adapts to user behavior patterns. Even in situations where a report is generated as the result of AI or machine learning – such as prioritizing items for human business analysis, according to predicted value – the value of AI depends on connecting its results and insights to the next step in business process, whether that's automated or human-based.
To dig deeper into best practices for getting real business value from AI and machine learning applications in production across a variety of business settings, read our new O'Reilly book, AI and Analytics in Production: How to Make It Work by Ted Dunning, Chief Application Architect at MapR, and me:
And to find out how to handle the logistics of model and data management, using a stream-based microservices architecture, read our earlier book Machine Learning Logistics.
Stay ahead of the bleeding edge...get the best of Big Data in your inbox.