Four intelligence gaps that’ll kill your AI initiative

Posted by Joey Greenwald

Whatever your profession, you can learn a lot by studying failure.

Naturally, our team spends a lot of time thinking about AI failures — and how to continually improve our CX automation platform so that our customers experience short-term and long-term success.

The sad truth is that outside our sphere, most companies today are failing with early AI and automation attempts. Analyst firm Gartner suggests that AI failure rates are as high as 85 percent. There’s a term our CEO Antony Brydon coined that encapsulates this challenge: “AI Project Mortality.”

Through our own external research, we’ve distilled the most common causes of AI project mortality down to four causes. Here are four intelligence gaps likely to kill any AI project — and how we address these gaps with Directly.

1. Lack of Content

AI simply doesn’t work well without a lot of data and content, which serve as the fuel for virtual agents to fulfill their intended purpose.

For companies trying to deploy AI, there are two sides to this content challenge. First is a lack of raw (structured or unstructured) data, which are the nuggets that answer questions people ask AI. This data can live in public or private databases. Second, is the specific language that an AI will use to frame a response to that question.

For contact centers trying to deploy AI, the data is likely limited to internally-created assets, such as historical help desk tickets and/or enterprise knowledge bases. And resource-challenged internal teams tasked with writing language libraries will have a hard time generating enough written copy to account for all the “edge case” questions that may come up.

AI powered only by internal content simply won’t scale.

How Directly can help support teams scale content: Our platform enables a company’s expert customers to create content and then train AI on resolving support issues by surfacing the right content. With the expert networks that power our AI, we’ve seen our clients grow content in enterprise helpdesk libraries by 10-50x in just a few months.

2. Low signal

In the age of big data, there may be no bigger challenge than finding the right data. Us humans are constantly “Googling” and clicking through seemingly endless web pages to find answers (and we often don’t).

One reason AI raises so much hope is that algorithms can process millions of pieces of information in an instant — where our human capacity for searching and learning is limited by time.

But even for AI, there’s also a lot of noise in data. AI needs signal — via human training — to know that it’s finding the right information and responding correctly to different questions. As we wrote in a recent post, AI needs human intelligence to be truly smart.

The first time an AI-powered assistant answers a question, that answer may not be very good. So, it’s super important that the AI is told when it's gone awry — by experts — so that it improves the next time a question comes up. Otherwise, it may head down the wrong path, like Facebook’s recent debacle.

How Directly increases signal: Continuous feedback, or what we call “Expert-in-the-loop” AI, is what makes our AI get smarter. Our platform helps our clients incentivize their networks of expert users to train the AI at scale, one response at a time.

3. Maintenance (and lack thereof)

Creating lots of content is one part of the AI challenge. Maintaining it is another.

Organizations sometimes have a wealth of digital resources — but how often have you searched a company’s website only to find an answer to your question that was wildly outdated? This, sadly, is especially common in the realm of online customer support centers. A knowledge base may have started off as an organized collection of useful documents and files. But over time, it can quickly lose its structure and relevance.

Companies looking to AI to automate customer service as a means to reduce costs neglect to consider the following: AI can only be successful if it has the proper fuel. An AI is only as smart and relevant as the content it has at its disposal.

How Directly helps with maintenance: Through Directly’s platform, expert users not only create content but maintain it. And they train AI to promote the timely content that’s most relevant. AI can also help companies identify outdated content, so it can be retired.

4. No empathy

As we try to create more intelligent and relatable AI, we often are trying to model some of the very characteristics that make us human. One of those very human characteristics: empathy.

Creating an AI that’s empathetic, and able to respond in a way that shows true understanding of what someone may be thinking or feeling, is no small task.

And lack of empathy can show up in many ways. At a base level, many chatbots simply don’t have enough command of language to offer relevant responses. Take this Zork bot from Facebook Messenger (source Medium):

Facebook Zork Bot

Why Directly AI is more empathetic: Empathy is more likely to come from someone who has experienced the same issue themselves. By having AI that’s powered by content from customer experts — rather than limited to company representatives — the AI is much more likely to be relatable. For example, if you’re a driver for DoorDash, who better to answer a question you may have than an expert driver like Jon Gallez.


Want to learn more about the state of AI in customer support?

Check out our latest research, produced with our partner ContactBabel, The Inner Circle Guide to AI, Chatbots & Machine Learning. And if you’d like to see our AI in action, contact us to set up a demo.