AI and natural language understanding (NLU) is evolving quickly and made quite some progress over the last couple of years. Still, it seems that fully automated support via chatbots is not there yet.

Therefore, a human intersection is necessary for every successful customer care bot. There are many different ways humans and bots can work together to increase the performance of every customer service team.

Here are four examples of how human operators can successfully work together with bots in a clear and more-or-less seamless way.

1. Upfront choice

This is the most basic intersection between bots and humans. When starting the conversation, the user simply decides upfront if they want to chat with a real human or a bot.

A benefit for the user to use the bot over a human operator is to reduce waiting time. The bot might be able to solve the user’s issue right away without the need to wait for a human to be available.

Especially if your company is new to bots and want to run a pilot, this is the best option to get a good sense if your users are willing to talk to bots without forcing them.

ChatShopper is a fashion chatbot that tried that in the beginning. They let the user choose between the bot and a human operator, pointing out at the beginning of the chat that it may take up to 10 minutes for the human to respond. Around 80 percent chose the chatbot (even though many people asked what a chatbot is).

AI / NLU Level needed: None
Human availability: Low-Medium (waiting time expected)
Development costs: Low
Operating costs: Medium-High

2. Human takeover

This is pretty similar to the option mentioned above. Every user starts off interacting with a bot, but the bot always provides an escape hatch where the user can request to talk to a human. This can be done in different ways.

One way, for example, is to have the bot respond only to interactive elements like buttons in Messenger, which means every text message is automatically forwarded to human operators. This is technically speaking the easiest way to handle this, since there are actually two different interaction levers: UI elements and text messages.

Pure Cycle chatbot

Above: Pure Cycle Shopify shop, where text messages are forwarded to human operators.

Another way is to have the bot listen to all input by default, but provide an option where the user can request an human operator from any point within the conversation — e.g., through a persistent menu, a keyword, or an intent.

From a UX perspective, there are multiple ways to accomplish the handover. Regardless, during the human-to-human conversation, the bot needs to be paused and must not answer the user’s messages.

ChatbotConf Bot Human Takeover

Above: ChatbotConf shop via oratio Shop Assistant, where the customer requests human takeover.

In my opinion the second option is preferable; there’s a clear line between the chatbot and human interaction, because you can only chat with one of them at a time.

The number one benefit for the human takeover is that the bot doesn’t need to understand natural language at all. Simple tasks can be (but don’t necessarily have to be) handled through the bot’s UI elements or through simple keyword and pattern matching, but for more complex tasks a human can be requested.

AI / NLU Level needed: None
Human availability: Low-Medium
Development costs: Low
Operating costs: Medium-High

3. AI-assisted humans

The idea behind bot assistance is that the users themselves actually never text a bot. All messages are forwarded to human operators. The job of the bot is to analyze the content of each message from the user and extract entities and intents.

Accordingly, the resulting tags can be automatically added to the conversation, and the bot can suggest a canned response to the operator. If the operator decides that the response fits, they can immediately send it by just pressing a button. If the bot’s suggested canned response doesn’t work, the operator either can edit that canned response to fit the conversation or just type a new one. In both cases, the new or edited response is used to train the bot.

Depending on the size of the company or customer care team, the bot can already decide which team or operator is best suited to take care of an inquiry.

AI / NLU Level needed: Medium
Human availability: Medium
Development costs: Medium
Operating costs: Medium

4. Human-assisted AI

This is the natural evolution of human assisted-AI. Just like with AI-assisted humans, each message is analyzed and classified using AI. The main difference, however, is that if the AI achieves a certain confidence level with its suggested response (e.g., confidence >= 90%), the message is sent out automatically.

Only messages with a lower confidence level are forwarded to human operators to be reviewed, sent out, and then used to train the bot again. The downside of this approach is that high availability of operators is needed to shorten response times to a minimum, especially if the user thinks he or she is messaging with a bot. On the other hand, only a few operators are needed to supervise the bot.

AI / NLU Level needed: High
Human availability: High
Development costs: High
Operating costs: Low

A glimpse into the future

One of the issues right now is that many bot developers don’t put that much effort into designing and training conversations. Machine learning and AI is already at a very decent level, but many teams and developers fail to model conversations properly. It also helps to have a big set of actual customer data or to train your bot with new incoming conversations.

I expect that NLU and AI will evolve even faster than they did over the past couple of years, and bot developers will put more effort into training bots with the help of human operators.

In the near future we will see a paradigm shift in customer care teams. Operators will not be in charge of answering simple customer inquiries any more — their role will transition into a supervisor of bots.

This post appeared originally at the Orat.io blog.