In the past two years, both growth numbers and tech advancements in the AI industry have made it fairly easy for its outsiders to fantasize and panic over a fully automated future.  

Since January 2015, when Chris Messina, Uber’s previous developer experience lead, started preaching about the power of ‘conversational commerce’, tech startups have been flocking to the AI and chatbot space.

With investments into the chatbot industry reported to have grown by 229 percent between 2015 and 2016, and those involving conversational B2B bots holding the biggest promise for large enterprises, the working force has reasons to worry. McKinsey predicts that sales and related jobs in the US hold at least 40 percent automation potential by 2018.

Facebook’s Messenger chatbots have jumped from 10,000 to over 34,000 and allowed native payments by 2016. The networks also foresees a more lucrative and integrated future for chatbot developers, with new offerings such as group chats with businesses, with integrated third-party apps, and chatbot searches via the discover tab, a new section in Messenger where users can look for bots .

Slowly but steadily

“People fantasize about bots and AI being super complex and understanding our emotions,” said Serag Menassey, cofounding COO at Botler, a Dubai-based seed-stage startup modeling itself as the Wix for chatbot developers. “We are going in that direction, but right now, the simpler it is, and the narrower the use is, the likelier a chatbot is successful,” speaking on a panel at this year’s Step conference.

Currently, Botler’s sole focus is on automating level-one cases of customer service, and designing bespoke chatbots for individual clients. At this year’s Step, it piloted a Facebook Messenger bot for Pepsi, with an app-like, rather than a conversational, interface. For the most part, users are led to choose between answers preset by the bots.

Multitude of models

This level of conversational understanding is not particularly Botler’s business, which has already onboarded clients like Bayzat, Helpling, Carmudi and Netmarble. The company taps into the deep learning solutions of tech giants and smaller companies. It brings machine learning algorithms that allow pattern analysis, classification, representation and, at a more sophisticated stage, abstract concepts of unstructured data. Large tech companies like IBM, Botler’s largest partner, offer access to their APIs through solutions like Watson, which provides developers with natural language understanding (NLU) capabilities.

But since these involve a lot of rule-based programming, the result is a somewhat inaccurate meaning attribution to data, which then gets honed, restructured, and tailored for client verticals by developers like Botler.

Smaller players

On the other hand, companies like Deepmind, acquired by Google in 2014, focus on replicating the human brain. It has recently made an AI program that can actually learn, like humans, from past experiences. Following suit in the smaller player arena is Finnish company Ultimate.AI, which specializes in NLU and builds the brains of the bots so that they can understand human language - and is currently under the Dubai Future Accelerators program. “One of our business models [branching out of NLU in chatbots] is customer service automation,” explained Reetu Kainulainen, Ultimate.AI’s CEO.

Cogtalk, another Dubai-based AI startup, pitches itself as a ‘cognitive computing customer service solutions provider’. It has a team of ex-IBMers that are constantly comparing APIs from different tech providers such as Amazon and Facebook, and adding their own layer to data structuring. But Tarek El Azzouzi, Cogtalk’s cofounder, said that the company’s expertise in customer service translates into a more consultative approach with enterprises, from conversation flow design all the way to tone analysis with AI interfaces.

Cogtalk has a suite of products, none of which are off-the-shelf, and most of which are in developmental or proof of concept stages. Its first AI product is a sort of internal customer agent assistant. It aims at reducing the learning curve for customer service reps in a high attrition industry. Its second product is a real-time listening service that offers agents immediate responses to customers on call. The third product, built around voice and tone analysis, is a ‘quality auditor’ which listens to, and assesses customer calls, tone, keywords, in real time, detecting ‘whether or not they were handled properly’.

The data mined and patterns collected through these products form Cogtalk’s fourth product, is “more of a back office analytics” big data repository, explained El Azzouzi.

Layers and layers

Though voice tech is out of Ultimate.AI’s scope, its work with text-based language understanding is more than enough of a challenge. Ultimate.AI’s Kainulainen said that most companies use old-fashioned statistical models, which establish relationships between data variables on mathematical equations. This model is contrary to machine learning, which centers on self-learning algorithms that are not dependent on rule-based programming.

Ultimate.AI starts with publicly available data, like language, for instance, to build a baseline deep learning model.

Other layers of classification and understanding follow, when it comes to enterprise verticals, such as clubbing a telco’s chat history and Twitter feeds into industry-specific queries and words. A sample of 500 questions would then serve as the AI’s training base for the rest of the data sets.  

Another layer has to do with intent modeling, such as attributing the query ‘weather forecast’ to a question.

A third is context understanding, whereby bots or AI interfaces are particularly trained on sequential and connected questions. For example, a bot that is asked: “what will the weather be like tomorrow in Dubai?” Without ‘reinforcement learning’ by means of human training, it would still not know that the question “and what will it be like the day after?” is referring to the weather.  

This complexity of language and data understanding layers translates into highly inaccurate and human-dependent bots. In February 2017, Facebook’s chatbots were reported to record a 70 percent failure rate in understanding human requests, and more recently its M Suggestions feature in Messenger scaled back. Deep learning will inevitably get smarter, but for now, Kainulainen says his company’s current scope of work has a lot to do with managing enterprises’ expectations. He believes the AI industry is experiencing a bubble, with the likes of IBM Watson, and other smart personal assistants like Google Home, Amazon’s Alexa and Facebook’s M overpromising and under-delivering on self-learning capabilities. The value in AI is, contrarily, in understanding narrow client domains, such as data packages for a telco.

Save the data

Data shortage and sharing restrictions imposed by big enterprises also concede to the growth of commercial AI. In the Middle East, Arabic natural language processing is a ‘huge issue’, particularly given the multitude of dialects across Arab markets, said Botler’s cofounder Taymour Sabri. IBM offers the technology to tell different Arab dialects apart, ‘but that is pretty much it’.
Academic institutions such as Stanford and NYU Abu Dhabi are actively tackling the Arabic NLP gap. But for companies like Botler, looking at commercial models for specific industries, data restrictions by big enterprises pose a challenge, with open-source libraries like Tensorflow offering limited support.Ultimate.AI has been working around this barrier by layering up the data architecture to dissect a limited set of data, rather than sifting through larger volumes. Moreover, it is offering on-premise software integration and sanitation services for enterprises so that their big data is completely rid of personal details.

For now, “I see the most value in functional bots, which provide a layer before the human interaction,” Kainulainen explained. Bots will take over, but not so soon.