The Truth About AI in Customer Service

Essay published on September 21, 2017

Note: this was originally posted on the Help Scout blog.

We’re in the midst of a massive AI hype cycle.

Artificial Intelligence is important technology, and there’s no escaping that it’ll change the way work is done in the future. But as with any hype around new tech hype, there’s a lot of misinformation floating around about AI.

We’ve talked to experts in the space, read hundreds of articles and reports, and invested in AI as a company, doing everything we can to establish best practices for customer support. Based on what we’ve learned, this is how business owners and customer support professionals should be thinking about AI today.

A brief introduction to AI

Whereas humans and animals exhibit what scientists refer to as “natural intelligence,” artificial intelligence is the ability of a machine to emulate natural reasoning.

As you can imagine, machines are capable of a wide spectrum of “intelligence” — and therein lies much of the misinformation about AI. Winning a game of checkers is quite different from diagnosing depression from speech patterns, but both are the result of AI at different ends of the spectrum.

Artificial Intelligence timeline Source: NVIDIA

The latest AI hype cycle is the result of two AI innovations: machine learning and natural language processing.

Machine learning is the result of more powerful computers being able to process very large amounts of data to “learn” from. You may not realize that you interact with machine learning AI every day — that’s how your spam folders work and how Facebook suggests which people to tag in your photos.

Your more obvious day-to-day interactions with AI, however, are the result of natural language processing (NLP), or the ability of a machine to understand and interpret speech. Siri, Alexa, and Cortana are all the result of incredible innovation over the last few decades in NLP.

The media more commonly refers to NLP, hence all the recent stories about chatbots. But both machine learning and NLP are instructive for customer service.

Will AI take customer service jobs?

People are quick to refer to reports like this one, which forecast a net loss of nearly 10 million jobs by 2027 as a result of innovations in AI. But that’s only one version of the story.

Customer service experts say that chatbots are only able to resolve 10-35% of customer inquiries without a human touch. Let’s assume we see a massive improvement over the next five years and will be able to resolve up to 40% of customer questions via automated assistance.

What do support professional think about all this? They think it’s fantastic. Any question a chatbot can answer is monotonous work for a support professional. The vast majority of support pros — 79% — feel that handling more complex customer issues improves their skills. A further 72% feel they have a bigger impact in the company when chatbots take on the easy questions.

AI elevates customer service professionals into roles that develop their skills, increase their impact, and improve their ability to participate in proactive, revenue-generating activities.

What happens when you take monotonous work away from support professionals? They level up. They create more value for the company and, in turn, for themselves. Those aren’t the people who will be losing their jobs by 2027 — assuming that prediction is remotely accurate.

Related: Chatbot Check-in: Are Robots Coming for Your Customer Service Job?

Should I implement a chatbot?

The first question to ask yourself is if the tool you are considering backs up to a real human. If the chatbot is unsuccessful, will the customer be able to talk with someone? A whopping 86% of consumers expect chatbots to always have an option to transfer to a live person, and a staggering number of tools in the market have conveniently ignored that.

Another important factor to keep in mind is that a chat-based user interface (UI) isn’t always best. It sets unrealistic expectations when the customer believes there’s a person on the other end to help, when in reality there isn’t. Chat UIs don’t have any visible features. It’s essentially trial and error to figure out what a bot is capable of. A UI with a search box and a couple other options, however, is a more intuitive experience for customers. This is why you won’t see Google removing their search box on the home page in favor of a chat UI anytime soon.

Many products that promote themselves as “AI-powered” are nothing more than natural language search, which can only return relevant help articles. If your chatbot does nothing more than search your knowledge base and return help articles, a chat-based user interface isn’t the right choice. When we designed Beacon, we intentionally did not make the interface look or feel like chat, because we didn’t want to set unrealistic customer expectations for the tool’s capabilities.

The last consideration for using a chatbot is whether you have extensive help content to support it. Remember, AI is only as smart as the data it’s learning from. If you only have 25 support articles in your knowledge base, chances of success with a chatbot aren’t high. If you have 250 support articles, your chances are better.

Consider using a chatbot if you can check these three boxes:

✅ Your chatbot backs up to a real human, for the more than 60% of the time it won’t be able to help your customers

✅ The UI features are intuitive to your customers, and they perform better than an old-school search box

✅ You have an extensive library of help content, and you estimate it can answer a minimum of 15%-20% of customer questions

Can AI help our team with email support?

The ideal machine-learning scenario is that when a customer emails support, AI technology can make suggestions based on what it has learned — anything from suggesting a help article to writing a draft reply that suggests the precise automations the customer needs to run.

For most small to medium-sized businesses — Help Scout included — AI can’t help with this scenario, because it requires an extremely large dataset to present meaningfully accurate suggestions. AI may be able to suggest a help article using NLP (like chatbots do), but learning from previous interactions and answers requires a ton of data.

According to our friends at SmartAssist — a leader in the AI and machine learning space working with companies like Mailchimp, Twilio, and Thumbtack — you’d need to be answering a minimum of 5,000 support conversations per month for the learning to be accurate enough.

If your team is managing that kind of volume on a monthly basis, it may be worth investing in machine learning AI that can enable your team to help customers more efficiently.

What to expect from Help Scout

If you’re a Help Scout customer, you may be asking the obvious question: How is Help Scout investing in AI to help my business?

We’ve always believed in building the most thoughtful product rather than the first product in the market. After doing our research and talking to hundreds of companies, we’ve been methodical in creating tools that set proper customer expectations and enable companies to operate at scale without having to hire a bunch of new people.

Help Scout customers can expect us to offer a product that delivers on the principles I’ve outlined here. That’s all I can say at the moment, but customers should know that we’re putting everything we’ve got into getting it right. We’ll be talking more about AI and product updates on the blog, so be sure to subscribe if you haven’t already.

Nick Francis headshot
Thanks for reading! If you have any comments or feedback, I'd love to hear from you on twitter.