Machine learning is fast becoming a reality for forward-thinking organizations. But for most businesses, the best way to take advantage of the capabilities of machine learning technologies remains something of a mystery. Still, the drumbeat to experiment keeps getting louder.
And the truth is, your competitors may already be laying the groundwork. IDC forecasts revenues for AI systems worldwide will almost double to $12.5 billion this year, and keep growing at a similar rate until they hit $46 billion in 2020. Some of that spending will go on hardware to run machine learning systems, but even if you don’t have the budget and the data scientists to build systems from scratch there are still plenty of tools and services that will let you use machine learning in practical ways that help your business.
Here are nine IT projects that almost any organization will find useful in getting started experimenting with machine learning technologies.
1. A customer service chatbot
If you have a list of frequently asked questions for customers to look up, you can turn that into a chatbot that can answer support questions using the Microsoft QnA Maker. It doesn’t have to be customer support, of course; you could create a bot to answer questions from new employees about HR benefits or how to contact the help desk.
Feed in the URL of your FAQ or upload spreadsheets and documents that have questions and answers and QnA Maker creates pairs of those questions and answers that you can review and train, and then call as an API. If you want to have a more interesting interface than just text replies, you can use the .NET SDK and the Microsoft Bot Framework to create a bot that shows pictures and rich content.
If you prefer the serverless approach, QnA Maker is one of the templates in the Azure Bot Service, so you can create a bot that works on email, GroupMe, Facebook Messenger, Kik, Skype, Slack, Microsoft Teams, Telegram, text/SMS and Twilio.
In the longer term, chatbots will evolve into intelligent agents more like Amazon Alexa and Microsoft Cortana. But rather than just answer individual questions, agents create a “goal-directed” conversation that works through the customer’s problem to help them solve it, which is what you need for ticket sales or diagnosing why a projector can’t connect. Microsoft has just added a customer care solution to Dynamics 365, in which a virtual agent suggests solutions, passes the customer on to human support, along with conversation details and the suggestions it made, if it can’t resolve the issue and learns what to do next time. HP, Macy’s and Microsoft’s own support service are already using this agent for online support.
2. Marketing automation and analytics
Marketing is often the first department to experiment with new technology, which is why marketing services like Adobe Marketing Cloud, Dynamics 365 and Salesforce are starting to offer machine learning predictions for everything from recommending related products for customers, to showing personalized search results, to classifying sales leads, to warning you when a deal is going cold, to finding alternate contacts at a potential customer company, even suggesting how and when to reach out to them. After all, predictive models for customer churn can help you with forecasting and planning.
If your marketing team isn’t already looking at these tools, this is a good way to apply machine learning directly to your bottom line. If they are, find out what’s working and look for other departments that could benefit from similar analytics. AXA is using a TensorFlow deep machine learning model with 70 variables to predict which customers are likely to have accidents that will cost the insurer more than $10,000, so it can optimize policy prices. Older models weren’t accurate enough to be useful, but with prediction accuracy improving from 40% to 78%, it might be good enough to consider when targeting potential customers.
3. Fraud detection
Spotting fraudulent and anomalous transactions is a classic data analytics problem, but if you’re doing it at a large scale, machine learning helps spot problematic activity such as scammers making multiple payments just under a trigger limit, new merchants exhibiting unusual behaviour and apparently legitimate customers who are connected to a network of scammers. Fraud.net uses Amazon Machine Learning to train multiple machine learning models to spot a range of fraudulent activity rather than trying to create a single model to score every possible kind of fraud; on any given day the merchants they protect might be facing a hundred different fraud schemes, each with dozens of variations.
Machine learning isn’t just useful for catching fraud by existing customers — insurers want to spot new applicants who plan to claim for a car that’s already been damaged before they issue a policy. And don’t just think about blocking bad transactions. Ford’s credit division is using machine learning tools from ZestFinance to predict the likelihood of specific borrowers paying back a loan so it can lend to people with lower credit scores. With car sales in the U.S. falling generally (and a slightly larger decline for Ford itself), finding buyers they’d otherwise turn down could be a big help to the business. Machine learning can help you tell good customers from bad risks more quickly.
4. ERP inventory planning
Supply chain automation isn’t new, but machine learning is making it much more common. Instead of just historic sales data, machine learning lets you use data about the way customers research purchases on line, the impact of weather on shopping habits and other internal and external trends to manage inventory by forecasting demand. Amazon claims it can predict exactly how many shirts of a particular color and size it will sell every day; Target credits machine learning predictive models with 15-30% growth in revenue. Online German retailer Otto uses machine learning to predict what will sell in the next 30 days with 90% accuracy, reducing the amount of surplus stock by a fifth and lowering returns by more than 2 million products a year; the automated purchasing system orders 200,000 items a month from third-party suppliers, choosing the colors and styles that are predicted to sell.
5. Logistics route planning
The travelling salesman problem is a computer science classic: What’s the shortest route between all the places your sales team needs to go on a round trip? Whether it’s getting salespeople to prospects, deliveries to customers or picking the business location that will attract the most customers, routing and travel planning has a big impact on your business. You can use the predictive traffic services in the Bing and Google Maps APIs to create isochrone maps that show you not just distance but travel time, to compare how many customers an engineer could reach in a 15-minute drive from various starting points, or find the best time of day to make deliveries. (Use the preview Bing Maps Truck Routing API to get routings for commercial and service vehicles that are larger than the average car.)
Add in asset tracking and location triggers and you can create your own logistics solution. Or you can make shipping more profitable by quoting rates that accurately reflect your costs, rather than losing margin by underpricing or losing business by quoting too high. Business communications giant R.R. Donnelley used R and Azure Machine Learning Studio to lower the cautious estimates that kept it from winning freight bid by combining historical data with variables like the weather, fuel costs and market conditions to develop a better pricing model. The automated system that generates real-time quotes for a given route is more accurate; the company is already winning 4% more of its bids and expects to quadruple the size of its truckload brokerage business. The same kind of predictive analytics would be useful for any contract bids where you have enough data to build a good model.
6. IoT predictive maintenance
If you wait until machinery breaks to fix it, you have downtime and unhappy customers; if you take systems offline to do maintenance too often, you reduce your production yields. When ThyssenKrup started analyzing the maintenance records from the 1.1 million elevators it installs and services, it discovered that the maintenance window could be quite a bit longer than it was. When the company used Microsoft’s Azure IoT Suite to remotely monitor sensors, predict failures and pre-emptively service equipment, it didn’t just increase customer satisfaction by fixing problems before they caused a breakdown; they reduced costs by fixing more issues on the first visit, and by being able to predict better what spare parts they needed to carry in inventory. Do the same thing with a manufacturing line and you can improve production yields. According to Accenture’s 2016 report on industrial IoT, predictive maintenance could reduce the cost of scheduled repairs by 12%, bring down maintenance costs by 30% and reduce breakdowns by up to 70%.
7. Machine learning for security
In the complex world of security, machine learning isn’t a silver bullet, but it can help you spot attacks that might otherwise be lost in the logs and alerts triggered by normal activities. Despite the name, Windows Defender Advanced Threat Protection isn’t anti-virus software; it’s a machine learning service that analyzes the behavior of PCs on your network running Windows 10 Enterprise and tells your security team whether an attack is a malicious process, social engineering or a document exploit. You’ll still need to dig into logs and deal with the consequences, but machine learning security tools can help cut through the noise.
8. Unbias your recruitment
There’s a growing push for diversity in business, but the way your recruitment team words job postings can actually discourage a wider mix of applicants. Try the Textio service which uses AI to flag corporate jargon, clichés, cheesy stereotypes and other offputting phrases in job postings and recruitment emails to help you get a wider pool of people applying. SAP SuccessFactors has a similar tool.
9. Image recognition for manufacturing safety
Building sites and manufacturing lines are full of equipment that’s dangerous in the wrong hands. Using cameras and sensors, you can use image and facial recognition to detect when that equipment is being used unsafely, or by someone who hasn’t passed their safety training. Hitachi has built a deep learning system with DFKI, the German Research Center for Artificial Intelligence, that uses wearables and eye-tracking glasses. Microsoft demoed a similar solution at its Build conference using Azure Functions, Microsoft Cognitive Services and Azure Stack. A full workplace safety solution might be challenging to build, but you can start with smartphone apps like The Safety Compass, which works with Intellect SEEC’s machine learning Risk Analyst to let workers mark hazards in a workplace by snapping a photograph and filling in the details; other workers will get a warning when they get close to the hazard.