“Do things that don’t scale.”
— Paul Graham

That quote from the British computer scientist says it all. At an early stage, startups need to do things like recruiting users manually — one by one — and going door to door to promote a product, which is what AirBnB did.

Chatbots were being hyped as the next big thing — but then the backlash hit, as disappointing experiences sent them plummeting from the App Store charts. But it wasn’t the concept that was the problem — chatbots really are a valuable way to drive scale within operations teams. In fact, my company, Mezi, has used chatbots to achieve 5-6x human leverage in just two years, while our user base has grown tremendously during that time as well. So yes, chatbot-enabled mobile apps can work for both your customers and your business. You’ve just got to build them the right way.

Here’s what you need to know.

1. Start (and end) with the human element

You can’t use automation effectively if you don’t know what you’re automating. Your first step has to be to understand human processes. Remember, the role of bots isn’t to run your business — it’s to allow you to become more productive. Whatever you’re doing, from outbound sales to customer support, there will be ways that bots can help you scale operations without increasing staff. Keep in mind that human involvement is still an important supplement to bots for the foreseeable future. Human agents still play an important role in our business, ensuring that we build lasting relationships and always deliver the best possible user experiences. In fact, this kind of thing is exactly what Graham was talking about when he said, “the unscalable things you have to do to get started are not merely a necessary evil, but change the company permanently for the better.”

When Mezi started in 2015, our plan was to offer a portfolio of bots designed to make shopping fun and easy across categories such as retail, travel, and dining. It wasn’t about programming bots at that stage — the first few months were all about understanding human behavior. By cataloging interactions between customers and our own human agents, we learned a lot about how people use messaging to make a purchasing request, from the steps they follow to the language they use. At the same time, we led time studies on end-to-end user experience and human operations flows, and we identified repetitive, low-value activities that were ripe for automation. For example, we found that our agents were spending more than half of their time gathering requirements from the customer, doing shopping research, and presenting recommendations in an elegant way — steps that were easily addressable by technology.

With this knowledge in hand, we then brainstormed solutions, pairing these manual processes with available AI and machine learning technologies to form an initial roadmap for our company to follow. Along the way, we continually gathered real-time feedback from our operations teams to refine the solutions we implemented.

2. Analyze users, ops, and market data

Once we’d established a baseline in step one, we dug deeper to gather additional data and insights on the customer experience we were delivering. While bot analytics and KPIs helped us assess and troubleshoot our performance, we also developed internal metrics and dashboards to benchmark ourselves against ambitious target KPIs. How long did it take our agents to complete a travel booking, place an order for a piece of clothing, process a request for a dinner reservation, send a gift, and so on? How could we make our messaging-based purchasing and checkout process as seamless as possible, yet fulfill the customer’s request as quickly as possible? How did sale conversion patterns differ between bots and humans? Did our customers even realize they were speaking to a bot? The ability to clearly answer these questions gives visibility into where we’re ahead and where we’re behind so that our next engineering hour can be optimally spent.

Our benchmarks also helped us understand which verticals could gain the most from the application of technology. Realizing that travel was our most requested, stickiest, and most profitable category, we narrowed our focus accordingly, and in Q2 2016, we evolved our business model to a bot-assisted travel agent platform.

The operational data and trends we captured have significantly informed and influenced our technology and team development efforts. Without this, we’d be flying blind and without a destination in mind.

3. Educate users and agents to embrace bot technology

It’s human nature to prefer engaging with people to dealing with a bot or computer, and that’s as true for your team as it is for customers. Still, for your business to scale, you’ve got to earn acceptance on both sides. Many bot-enabled businesses overlook the importance of developing the right voice, and they end up with an experience that’s about as warm and engaging as an IVR system. (“Touch 1 to repeat this message. Touch 2 to give up and go to a store instead.”) We did a lot of work to train our bot voice to be likeable, even charismatic, using humor, excitement, and emojis in ways that make it hard even for us to tell whether we’re interacting with our bot or one of our human agents. And it’s paid off; our top customers refer to Mezi as their BFF for travel, and we’ve got a great net promoter score (NPS). Even when people know there’s a bot on the other end, they feel the kind of emotional connection that makes them enjoy the experience and want to come back.

The same is true for our own team. We conduct relentless training and development for our operations and customer service teams to help them work with our bots effectively, and they enjoy engaging with our internal interface. In fact, they treat our bots like colleagues. As a result, not only has Mezi become more efficient and scalable, we also see very little attrition compared to the typical customer/sales service center. With bots taking over mundane and frustrating tasks, individual job satisfaction is very high.

4. Drive continuous improvement through feedback loops

While we’ve come a long way in just two years, we’re still early in our journey with much more yet to accomplish. We’re continually focused on two goals: 1) improving satisfaction for users booking travel through Mezi, and 2) enhancing the efficiency, accuracy, and job satisfaction of our agents.

For customers, we survey both our most-satisfied and least-satisfied users to learn what we’re getting right and where we need to do better. For agents, we use Slack channels, weekly team meetings, and individual surveys to learn what they enjoy about their jobs and what they’d prefer to avoid. The typical call center job involves a lot of low-level tasks people just don’t like doing; the more of these we can automate away, the more our agents can focus on the things they do enjoy, like building relationships with users, offering more personalized service, being thoughtful about their roles in the company, and making a higher-level contribution to our business. To satisfy both groups, our engineers conduct persistent testing and validation, internal and external, of new features so they’re rolled out as flawlessly as possible.

So, what’s the outcome of this process? For Mezi, we’ve seen high rates of customer growth without increasing the size of our team.  Our customers love our app, and our agents love working here. That’s a strong validation of the potential for chatbots to transform service centers and business operations — as long as you go about it the right way.

Brian Li is the VP of business operations and partnerships at Mezi, the travel concierge service.

Above: The Machine Intelligence Landscape. This article is part of our Artificial Intelligence series. You can download a high-resolution version of the landscape featuring 288 companies by clicking the image.