Machine learning has become one of the business and product world’s go-to buzzwords. Healthcare! Business intelligence! Personal assistants! If you were to take the tech press at face value, you’d come away thinking we’re barreling headlong into the future, with the whole universe ready to be turned over to benevolent automated machine intelligence.
Of course, the reality is far less impressive. But it is still true that machine learning has made very real impacts in a lot of very important industries. And that means there are some really important discussions to be had about how machine learning fits into products and, ultimately, how it fits into our lives as people.
But first, let’s talk about customer support hotlines.
A customer service apocalypse
Anyone who’s had the misfortune of calling a customer service or support one line for an airline, cable company, bank, or other big company can probably agree that it’s an incredibly frustrating and time-consuming process. To understand why it’s so aggravating, let’s take a closer look at a typical experience:
First, you call the number and choose from 4 or 5 options. It takes a long time to hear them, and they may or may not line up with exactly what you want. Sometimes you have to speak your answer, which sometimes works and sometimes doesn’t. Either way it’s usually annoying.
Once you’ve chosen your option, you’re asked to enter some sort of account number. This is long and cumbersome, and if you mess up you have to start all over. Sometimes you have to choose from another bunch of options – and, again, they may or may not line up with what you want. Sometimes you realize that you entered the wrong first option and, yes, have to start all over again.
Want regular news and insights about innovation and product development?
Sign up for our Product Hacker newsletter.
You’re probably pretty annoyed by now, but luckily the next step gives you some time to cool off — the waiting, accompanied by some selection of light jazz or classical. You may or may not be told your place in line.
Finally, you get to talk to someone. Hooray.
It’s deeply frustrating, and for a variety of reasons – but most significantly, it tends to give the impression that you’re annoying the company just by calling them, or that.you’re being forced to jump through hoops or self-screen on the off-chance that you have a problem that can be solved by a pre-recorded message. (When in reality you probably have some weird corner-case issue that a real person could solve in thirty seconds, if only you could talk to one.)
But it wasn’t always like this. There was a time not too long ago when you’d call a help line, wait around for a while, and then get transferred to a person who solved your problem – which is usually what people want anyway.
So why’d it change? The simple answer is cost.
The high cost of saving money
Human-powered customer support is expensive. It involves a lot of people answering phones in real-time, and the workload can be devilishly hard to predict. Enter basic automation.
Each step in a customer support automation system saves a little bit of money. If you can figure out why someone is calling before they talk to anyone, you can send them to a team that handles that specific problem – which means you don’t need to train every representative on every problem, and lets you send the bulk of the issues to less experienced representative while keeping the more experienced and consequently more expensive representative only for the calls that actually need them. And entering your account number saves probably 30 to 40 seconds per call, since the representative doesn’t need to collect it.
If you’re answering thousands or even tens of thousands of calls each day… well, the cost savings of automation add up pretty quickly.
There’s a downside to all of this, of course. While these automated phone systems probably do save consumers time on a statistical level, they definitely FEEL worse. They’re annoying. It’s a bad experience. Consumers notice when companies don’t seem to care about them, and they’re not afraid to say so.
What AI aficionados can learn from customer service
So, what does all this have to do with AI?
AI, like automated phone systems, is a way of taking an amount of work currently done by people and moving it to be done by a system. For product people, that’s a HUGE opportunity. It might be looking for lumps on radiology images, scanning through financial data for evidence of fraud, or – yes – responding to customer service issues. Anything that’s repetitive, data heavy, and at least somewhat structured can probably be turned over at least in part to machine learning.
But the challenge isn’t finding something you could use AI for – it’s determining what you want to optimize for.
Like any good project manager will tell you, you can build for scope, for cost, or for time. That’s the case with just about any project under the sun, and AI is no exception.
So, yeah, you could develop an AI-assisted system that handles certain types of diagnoses – but make sure you know why you’re doing it. Do you want to reduce the cost of getting the diagnosis? Deliver more diagnoses per day? Or do you want to build a more flexible system that can handle more types of diagnosis?
Each of those would necessitate a very different machine learning approach – and a different sort of outcome. And it’s important to consider the pros and cons of each.
Back to phone systems. The main drivers of the modern help line are clearly cost and time, not the scope. More automation means faster turnaround and reduced staffing costs – but it also means more user frustration, because there are plenty of things those system simply can’t do.
You can imagine a help line that uses automation not to reduce costs but to make customers happier – maybe you could give them more tools to solve their own problems without needing to wait for help, or improve the system’s ability to recognize and solve for niche use cases. But it would probably cost more, and it might be slower. There’s always a trade-off.
Now imagine the help line approach being applied to, say, healthcare – maybe you’re going to use machine learning to assist with front-line diagnoses. Is the intent to reduce costs, to get your results faster, or to build a more flexible and feature-rich system? Because each of those is going to look very different, from the user – or patient – point of view.
Imagine the frustration of having to rely on automated test results, and being forced to wait in line for a human doctor if you don’t have one of the handful of big conditions the system’s been designed to treat. You can even imagine a scenario where less expensive insurance only covers those cheaper automated tests, with human doctors reserved for a higher tier of coverage.
And it goes beyond healthcare. Machine learning is coming to retail, customer service, logistics, everything.
Make no mistake – as product people, we have to choose, and choose carefully, if we want it to be a tool for cost, for scope, or for speed.
Otherwise we risk building products that jettison functionality and user needs, even as they save money or get things done faster. And the last thing we want is a world where every experience is as frustrating as calling a help line.