The Evolution of the Chatbot

The very first chatbot was not very friendly

When I was a kid, I read The Argonautica — a Greek epic poem by Apollonius Rhodius, written in the third century BC. It tells the myth of Jason and his crew (that included Hercules) having an epic journey full of heroic adventures in order to obtain the Golden Fleece, held in Colchis (modern Georgia — the country in Europe). The Golden Fleece was some kind of a cool artifact or a metaphor for it. Back then, there was no direct flight from Iolcus, Greece to Sukhumi, Georgia, so argonauts had to make a few stops in different places on the Mediterranean and Black seas. One of these places was Crete. That's where they met the first chatbot ever known to humanity — Talos.

Talos was a giant bronze automaton — protecting Europa (the princess, not the continent named after her) in Crete from pirates, hackers and card fraudsters. He circled the island's shores three times a day — obviously, he had a part-time batch job. Talos had Crete laws engraved in bronze and was able to adaptively interpret the situation against those laws and assess if somebody was an intruder or a good guy. A bit like the good RoboCop in RoboCop 2.

Clearly, Talos was a chatbot implemented in hardware (literally — bronze). He could recognize end-user intents, match patterns and apply decision trees and other machine learning techniques to produce responses to externalities. You would never hear from him: "I'm sorry, I don't understand, can you try rephrasing?". Talos was a very powerful chatbot but not very friendly.

Rage against the machine

Since Talos and the Argonauts, the field of artificial intelligence has had multiple ups and downs, its periods of romantic optimism followed by AI winters of low interest and funding cuts. However, the recent advancements in hardware designs and computing removed constraints and lowered the barriers impeding the development and commercialization of machine learning solutions in the past. As a result, in recent years, we've been observing a significant, growing interest in the domain of machine learning and artificial intelligence, and chatbots have been a big part of it. Billions of dollars were invested in chatbot companies and natural language processing research and development. Nowadays, you won't find a person who has never had to interact with a chatbot.

However, this wave of interest seems to have reached its peak and may fade again unless the industry aligns the expectations with reality and pitch decks stop competing with Jules Vernes for the most optimistic vision of the future.

The promise of human-like interaction with bots fell flat again. To date, no artificial intelligence application was able to pass the Turing test — a test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human, where "intelligent behavior" was reduced to "having a short conversation.” Google Duplex voice AI booking an appointment with a hairdresser may seem like an incredible leap in AI voice technology. However, it's an achievement contextually relative to the baseline. Similarly, when my two-year-old daughter learned how to put her shoes on, or my cat learned how to shake paws, high-five and lick my forehead for a treat. — It's an achievement relative to their personal baselines but not a groundbreaking objective change for humanity.

As I mentioned, you won't find a person who has never interacted with a bot. However, a lot of those interactions were frustrating. My most recent conversation with a chatbot was so infuriatingly maddening that, while I was on the phone waiting for a human operator, I wrote a little JavaScript in the browser console that was feeding the chatbot with random obscenities every second in an infinite loop. My rage was so through the roof that my smartwatch showed a spike in heart rate. The playful, intimate friendliness of that chatbot also was a contributing factor to my rage.

The problem with the pursuit to "exhibit intelligent behaviour" is exacerbated by a misplaced accent between customers and providers of chatbots solutions. Providers promise chatbots to become more friendly, intimate, more humanized and promise customers to replace humans with bots, thus, reducing the cost and streamlining operations. Customers, predictably, nod that it's a good thing. The problem, however, is that this is not what the users want.

Brave new world of bots

The Theory of Jobs To Be Done can give us an insight into what users want and don't want from a chatbot.

First of all, the jobs-to-be-done of an end-user remain the same. Whether there is a chatbot or not. One can imagine a person logging in to online banking to check their balances, review transactions to apply payments, make an account-to-account transfer, etc. If you added a chatbot — the user does not get a new job-to-be-done (to talk to the chatbot). Neither the user is relieved from reconciling their bank statements with accounting records.

Second, the chatbot can be the main user interface or a support tool for an existing interface like regular web-based online banking. While the applicability of conversational interfaces for banking and accounting jobs-to-be-done is a topic for a separate debate, I'll just leave you with my opinion that nobody wants to type or say out loud: "Hey, Alexa, what is the checking account ending on 7689 balance?" or "Hey, Chase, can you read the bank reference number for the transaction on that account with the amount of $32,443.93?". This is so 2015, and I hope we never come back to this. The fact that you can have a haircut done with a flamethrower doesn't mean it's the right instrument for the job.

It leaves us with the chatbot as a support tool. As a support tool, a chatbot can support operations and sales. If we go back to Talos The Chatbot, he was really good at what chatbots should be good at — detecting end-user intent and mapping the intent to available operations.

Detecting the intent is the core of the natural language processing field. Mapping the user's intent to solutions, so that the user won't have to call the call center, is the chatbot's job-to-be-done. Not becoming friends with users.

At this point, it should be self-evident that the tone, breadth of vocabulary and friendliness of a chatbot serves the same function as a nice loading animation on a website. It's a nice, but expensive cherry on top of a cake, but the cake has to be good in the first place.

Chatbots will not be able to replace humans in sales and support, but they can reduce their load. Not by being nicer and more friendly, but by being more efficient in guiding a user to the right path in the customer journey to solve their issues. If the chatbot doesn't have the right path — a human support person will find it. The necessity to fall back to human support is just a symptom of a broken process. A new/different tool won't fix the process. For conversational banking to be a success, it should stop being treated as a shiny toy and be built on top of solid operational processes.