The Chatbot Market: Cutting Through the Noise

By now most people are familiar with chatbots. They’re these conversational agents that you interact with through dialogue in order to perform tasks or get information. Getting a chatbot up and running, however, requires navigating a confused marketplace of managed-services-dressed-as-tech and smoke-and-mirrors-dressed-as-ai.

I’m going to attempt to break down how this market is constructed and try to rationalise a highly fragmented array of offerings.

First, let’s look at how chatbots actually work

For a chatbot to function the following four components are needed:

1. Natural Language Understanding
2. Dialogue Management
3. Chat processing
4. Hosting and deployment

In reality these four things address the very basics of what chatbots need, but to truly thrive you can include strong analytics to measure how a chatbot is performing and the ability to integrate with other systems, which requires code execution abilities in the chatbot that go beyond simply responding with a verbal message based on the user’s written or spoken query.

When these four components are offered by a single provider either as Software-as-a-Service or as Platform-as-a-Service, that’s a Chatbot Platform. There are many of these out there, offering vague or no differentiation, one from the other, and mostly targeting customer support or marketing use cases.

The Chatbot Platform market is mainly populated by startups but is also serviced by large enterprise software providers like Salesforce, SAP, Oracle and Google.

This market exploded as a result of the democratisation of Natural Language Processing technologies — much of which are open-source — and the decreasing costs of cloud computing, which is critical for more heavy duty machine learning operations as well as handling high volumes of chat. Of course, these emerging chatbot providers also latched on to the general hype around AI, despite many of them using little or no actual AI in their solutions (more on that some other time).

Natural Language Processing is the means by which software receives a text (like a user query) and deconstructs it semantically, grammatically and can even provide a measure of sentiment from the user’s tone. NLP is a precursor to NLU (Natural Language Understanding) which is the method by which a Chatbot Platform will act on this deconstructed language.

While these Chatbot Platform providers mostly use either open source NLP technology, or licensed commercial NLP from vendors like Google or Microsoft, the dialogue management and chat processing parts of the platforms are almost always entirely proprietary.

AI Snakeoil?

Every one of these Chatbot Platforms jumped on the AI hype train to push the innovation around their offerings but early consumers of these solutions largely encountered a harsh reality check. While these Chatbot Platforms promised (and could, to some extent) automate conversations at scale, they still required significant human intervention in one way or another to get the bots to the expected level of quality.

Anyone buying into chatbots was in for a rude awakening. Notions of a plug-and-play artificial intelligence that would simply start talking to people independently were quickly dispelled when the terrible performance of early chatbots was made publicly known.

Not only did it transpire that significant human effort was required to make these bots good but that the constituent parts of the Chatbot Platform required very different disciplines in order to make the bots good, namely:

Natural Language Understanding:

Technical people like engineers, data scientists or computational linguists.

Dialogue Management:

Creative people like UX copywriters and subject matter experts.

It’s confusing enough to understand how any one of these Chatbot Platforms works technically, but to cloud the issue further, there are at least three different ways in which Chatbot Platforms are capable of pursuing the same outcome. These are broadly summarised as:

Rule-based NLU

Fairly rudimentary method of matching a user’s input using keyword patterns and other linguistic ‘rules’.

Pros:

1. Unlike machine learning it doesn’t depend on the availability of labelled training data to work. It’s ‘deterministic’, meaning it doesn’t calculate a probability of what the user said matching what it’s been taught to do. It rather just looks at whether the words in the user’s input match the words it’s been given to understand and progresses accordingly.
2. More control over esoteric terminology. Because the rule-based approach is matching keywords rather than doing statistical calculations on complex linguistic patterns (like word vectors), it’s easier to teach it synonyms for certain words that might otherwise be a little too specific to a context. Knowing that a “cellphone” might also be called a “handy”, for example, is something easily resolved.
3. Conversation Designers (the people actually telling the bot what to understand and how) have a lot more control over improving the chatbot’s understanding because it doesn’t depend on volumes of data to train machine learning models.

Cons:

1. In cases where more generalised language is being used, it could take more human effort to achieve through a rule-based chatbot what a pre-trained NLU chatbot might be able to do.

Machine Learning NLU

This is what most Chatbot Platforms use. There is a base NLU model that calculates how close a user’s input is to an ‘intent’ defined by the developer or Conversation Designer. Based on that probability the bot can then act according to how it’s been configured to respond.

Pros:

1. Accuracy of the bot’s understanding can improve over time if more data is trained into the NLU models. Similarly, domain-centric models are possible if (and it’s a big if) enough domain-centric data is available to train from.

Cons:

1. It’s no less work (possibly even more work) than a rule-based approach in order to create it in the first place. Someone needs to provide a number of phrases to the bot for each ‘intent’ in the same way that someone would need to provide all the phrases to teach the bot keyword rules in the rule-based approach.
2. Critically, it is a black box. So if something doesn’t work as planned it’s not straightforward to understand why. Ultimately a pre-trained NLU chatbot is only as good as the data its models have been trained on. Anything beyond that may be in the lap of the gods.

Agent-in-the-loop reinforcement

This method has a kind of base machine learning model in place that doesn’t require someone to pre-train the chatbot to understand certain things but rather makes calculated guesses based on its own model and then recommends these guesses to a human agent to either confirm as correct, deny as incorrect or provide a correct answer.

Pros:

1. Doesn’t take much time to set up initially.
2. Gets reinforced with use, rather than having to routinely train models as in the pre-trained NLU approach.

Cons:

1. Requires a good amount of clean data to set up initially
2. Cannot reliably function without humans being available, meaning your chatbot might not be usable outside of human working hours for a long while.
3. Sucks the time of your human agents and could be highly distracting and slow while agents try to make sense of what the chatbot is proposing.
4. Usually depends on human agents to create or curate content. Conversation Design is a design discipline as valuable as any other and if human support agents are required to provide the answers that the bot needs, in real time, the result will almost certainly be degraded, poorly written content.

As far as Chat Processing and Hosting and Deployment are concerned, these are largely already-solved infrastructural challenges that are important (because you want your chatbot to respond quickly and be reliably online) but most Chatbot Platform providers will be using the latest cloud-based methodologies to keep things performant, scalable and resilient.

The snakeoil in all this is the highly liberal application of the term “artificial intelligence”.

Perceptions in this market have been badly mismanaged in the name of fake it ’til you make it. Both sides of the Chatbot Platform space (business consumers and investors) got sucked into the hype and now perceptions are having to be rebuilt in order to re-establish confidence that automated conversation is a solution for customer support automation challenges.

For businesses looking to consume the technology there has been the necessary awakening to the fact that a good chatbot is a longitudinal process, an evolutionary process, rather than a plug-and-play solution. What’s more it takes significant human effort — and associated costs — along the way.

For investors the realisation has arisen that these Chatbot Platform businesses don’t scale as rapidly as they would like to, making the prospect of a ‘unicorn’ in this space difficult to imagine at this time.

Two of the bigger success stories (at least in terms of investment raised) are Clinc and Ada, both of which have opaque pricing and no self-service SaaS offering. This is reflective of the greatest inconvenient truth in Chatbot Platforms: “We have a solution for you. How much does it cost? That depends…”

Creating predictable, standardised and transparently-priced commodities is critical for a VC-funded software business to scale, and standardising anything that depends on dramatic variances in human language is a challenge that will take some time to resolve. While some Chatbot Platforms may have focused on industry verticals to generate impressive initial revenue growth, the exponential scale is hard to see — especially if it depends on global markets where language becomes even more challenging.

What next?

Chatbot Platforms aren’t going away but there needs to be a fragmentation of the whole space into focused parts of a value chain. Here are some things to look out for over the next 12 months:

Moving towards shared standards

Hundreds of Chatbot Platforms and almost entirely proprietary methods for defining chatbot structures for each of them. This is unsustainable and the industry will need to collectively move towards further standardisation of terms and definitions.

Further verticalisation around esoteric industries

Some providers will likely dominate specific industries where there is either esoteric domain subject matter, data sensitivity challenges or both. Think banking and healthcare, for example.

The rise and rise of Conversation Design

Like the early days of web and mobile before it, conversational experiences cannot continue to be disproportionately controlled by developers. UX design will increasingly make incursions into this space to drive better quality chatbots.

A split between rapid-scale businesses servicing the value chain and slow-grow businesses offering holistic solutions

The workflow required to get a chatbot from inception to live, and being improved upon, requires focused solutions that constitute entire products in themselves. Content management is an emerging challenge, for example, and already analytics has been subject to focused solutions like Chatbase. One single platform cannot continue to try to solve all of these simultaneously.

The early-mover Chatbot Platforms of the move-fast-and-break-things philosophy may be finding themselves having bitten off more than they can chew as they try in vain to provide standardised, out-of-the-box solutions for problems that are not that easily solved and are having to bolster their solution with managed services to get things done.

Conversation as a channel is not going away any time soon, so it’s safe to assume that a desire to enable language as a user interface with machines will continue to burn strongly. However the next 12 months should see a more considered and realistic approach to how quality is accomplished in conversational experiences.