Today marks my one-year anniversary of doing regional business development at KeyReply, an AI enterprise chat automation company. Earlier this year, one of our engineers shared about his first year at KeyReply and this is my business development take.
For starters, we all know that ‘business development’ is pretty much just euphemism for sales, though to give my team credit, I think we do a lot more than just cold-calling and hounding decision makers to buy our products. I like to think of us as the purveyors of AI technologies, driving change and supercharging our customers’ and their teams’ productivities (read: yes, we don’t just do chatbots).
As a sales person, I’ve met the gamut of customers; had such delightful and enlightening conversations with C-level executives on the AI and data initiatives their companies are running… and my fair share of clueless managers and associates who are just looking to fulfil their three-quote procurement requirements.
I’ve had countless discussions/calls with business users which went something like this:
Me: Chatbots are like humans, they take time to ingest the data and make sense of patterns, to return the correct answer to users. We need to teach the chatbot and bootstrap it with an initial data set.
Customer: What do I need to do to teach the bot?
Me: The usual process for us entails teaching the bot intents and corresponding variations to build the first version of the NLP model.
Customers: Variations meaning?
Me: Think of variations simply as the different ways the same question can be asked.
Customer: Which means I still need to manually teach the chatbot (lah). (Singaporean readers, you should be able to read the tone of the customer here)
Customer: So there is no AI (lah)!
Me: … *whimpers internally*
AI is an amazing technological advancement, but it is no magic. Without prior training, it is not possible for it to conjure its own answers just by feeding it with a simple FAQ. Recall what happened with Microsoft’s Tay, are you certain you want your ‘intelligent self-learning’ AI chatbot to be answering your customers and calling them slurs? I’m not saying that all unsupervised machine learning agents will result in a racist, bigoted artificial intelligence, but it most likely will return unsatisfactory answers if left completely untamed.
Enough of my backstory, here’s my list of the top 5 questions I get asked all the time, which hopefully also answers some of yours.
1. Can the chatbot learn by itself? Why do I still have to do so much manual work?
Drawing parallel to an infant; can your newborn baby learn by itself? In all seriousness, my CEO was the first to introduce this analogy to me. A fledgling AI is exactly like a baby, he/she must learn from scratch, and it really is up to the parent, in this case, us, or the business user, to teach the AI how to make sense of languages and questions that are thrown at it. We’ve sort of circumvented the issue of a cold start (where customers have completely zero data to work with) by creating our own pre-trained models for some industry verticals, but there’s still a process of teaching the AI company-specific jargons you simply cannot avoid.
2. Do I need to maintain my chatbot after it goes live? Can you do it for me?
Same analogy applies — do you or your child’s teacher mark his/her work (well I sure hope someone is 😊)? Without correcting and supervising your child’s learning, it’s impossible to teach him what’s right and wrong. This is where we use a combination of supervised machine learning and human reinforcement to ensure a recurring positive feedback loop. To answer the second part of the question, yes, we can, but as with company and/or industry specific domain knowledge, we do recommend having one of your subject matter experts weigh in on the monthly AI training.
3. How accurate is your chatbot?
This question is perpetual — even existing customers ask me this all the time. I’d like to introduce a concept that most of us are probably more familiar with: call center agents and their productivity. You don’t hear of managers asking how accurate are your call center agents, do you? Chatbots are (typically) deployed to solve a business problem, and our bots aim to help to solve said problems. A solution to business problems such as high customer attrition rates, low conversion rates or low customer satisfaction scores should measure relevant metrics, such as number of sales closed by the bot, or CSAT scores before-and-after the bot has been launched. Similar to the idea of call center agents, bots should be more fairly measured on whether or not they help to resolve the business issue on hand, and not merely on how ‘accurate’ they are. Of course with that being said, we do have our own tools, such as batch testing, to periodically measure the accuracy of the AI against a set of questions.
4. Do you do voice chatbots/assistants?
According to CES 2019, 2019 is the year of Virtual Assistants and 5G. Let’s park 5G aside for now and focus on Virtual Assistants, such as Apple’s Siri, Amazon’s Alexa and Google Assistant. Indeed, it’s easy to hook up one of our chatbots to Google Assistant or add it as a skill to Alexa, but I’ve noticed most customers who ask this question don’t really know what they are looking for. Listing your chatbot on a voice assistant store sounds great, but it only makes sense if your bot can help your customers perform specific tasks through these voice assistants (think booking an airplane ticket or checking your credit card outstanding bills). FAQ bots don’t find themselves very useful on a platform like Siri or Alexa, because customers don’t go to those channels seeking answers. It’s also possible to link up your chatbot to your native mobile application, but that involves effort from both sides, us, and your mobile application development team.
5. Can I get a free trial?
Here’s the truth, AI solutions (bot or not), are not created like typical SaaS platforms. There’s usually a steep learning curve, which tapers off after the initial on-boarding phase. This on-boarding involves a product specialist (we call them bot automation developers) teaching you how to create your own intents, and even developing complex business logic to handle the questions. And there’s a high dependence on the quality and quantity of the data sets you input to train the AI. We’ve had experiences with customers in the past who have requested for a trial, but after getting their hands on the product, they had no idea how to use it. Even with a hands-on training session, they struggle to find the data in the right format to input in the bot because the reality is, much of their data (if any at all) isn’t ready for machine learning. And so, they do a half-hearted job at it and call it a day, saying that the solution doesn’t meet their expectations.
These days, I avoid giving my customers a trial account. Instead, I would offer to create a dedicated demo for the company, that shows them exactly how data (based off their FAQ) should be labelled and stored for the AI to understand. Customers definitely see more value in a demo that has been properly constructed to meet their business needs, more so than a free sandbox for them to dabble with.
I know I’ve said a lot about these questions and might sound overtly passionate, but I find it heartening to see an increasing interest in AI solutions in general. These questions show the thought process behind implementing a solution, especially when it comes to the one-time setup and ongoing maintenance. I understand that different people have different levels of experience with AI; chatbot jargons like “intent”, “entities”, “variations”, “NLP”, can sound threatening to most business users but it’s nothing that a 10-minute explanation cannot solve. After all, I am happy to share more and would love to educate more businesses around us to become more confident and literate with AI enterprise solutions, and the technology that powers it.