As marketers, we’re naturally drawn to the shiny and new. Why? Because commercial competition demands it (to put our companies ahead of the pack), and on a personal level, our vanity and interest in doing something different drives us.
One such shiny, new example that’s looming large is Artificial Intelligence (AI), harnessing a conversational user interface (CUI). Many are speculating the demise of the typical GUI (graphical user interface) that characterises the websites and native apps we all know and love, in favour of ‘chatbots’ - a computer program designed to simulate conversation with human users, especially over the Internet. Potentially, even the death of the GUI altogether.
Hyperbole aside, discussions about AI, and chatbots in particular, have taken place for a while. But it’s only recently that the application of this technology has started to become more widespread, most notably since the launch of Facebook’s new Messenger Platform with support for bots.
To be clear, Apple’s Siri, Google’s Now, Microsoft’s Cortana, Amazon’s Alexa, and smaller companies such as X.ai, have been delivering functionality through natural language processing for some time. However, working either by speech or text, they have had fairly limited scope and very limited capacity for 3rd parties to attach themselves to those proprietary services.
With the advent of APIs from the likes of Slack, Telegram, and particularly Facebook, the door has been flung wide open for the rest of us to get in on the action and connect with consumers in a new way.
Some companies (e.g. CNN, Burger King, Domain.com.au) have started to deliver bots on top of those platforms, generally for fairly simple tasks and interactions. Other brands, such as North Face and 1800 Flowers, are among the first to go down the browser-based route, with the ‘intelligence’ element driven by sophisticated AI service providers such as IBM’s Watson.
Here at e3 we’re excited to be collaborating with IBM for one of our clients, to deliver one of the first projects of its kind, using Watson technology. But while working with such new technology is exciting, it’s not easy. There are various challenges we’ll all need to overcome if we’re to be successful in implementing an intelligent chatbot solution:
1. Understanding our audience’s expectations of the technology: for those of us (and that should be all of us) who like to take an audience-centric approach when designing new solutions to our consumers’ problems, we need to be clear on what our audiences expect from an experience with a bot.
This is crucial: it helps us set appropriate expectations early on and avoid potential frustration when the reality doesn’t meet them. It’s particularly important given not all bots are equal, with a similar level of domain-specific knowledge or natural language processing capabilities. As Jeff Bezos recently highlighted: "There will be a lot of intelligent agents in the world. You may not ask the same AI for everything. Some will be better than others [at certain things]".
But how will we know, as users, what we can and can’t ask of each bot before we speak to it? People’s ability to attribute mental states—beliefs, intents, desires, knowledge, etc.—to oneself and others is called Theory of Mind. Yet how can we, as marketers, clearly communicate what the mind of an ‘intelligent’ machine is?
In 2014 the Turing test was passed (somewhat controversially) for the first time, when a computer was mistaken for a human (a 13-year-old Ukranian boy named Eugene, to be precise) for more than 30% of the time during a series of five-minute keyboard conversations.
Now there’s still a long way to go before machines are genuinely, convincingly intelligent as humans in terms of conversation, but I believe we’re fast approaching a time in which we’ll regularly experience a similar phenomenon to ‘uncanny valley’ in the robotic world – that is, the closer a machine appears to being human, but ultimately not, the less comfortable we feel interacting with it. Concisely and accurately communicating bots’ ‘intelligence’ levels, capabilities and limitations, and reducing the potential discomfort of interacting with an apparently intelligent machine, will become increasingly important.
2. Designing a chat experience: designing the user experience for a conversation, and testing that design, is a very different kettle of fish to the mature design and test process for traditional GUI-based sites and apps. Instead of navigating content in a pre-defined, structured manner you now need to cater for a more fluid and less predictable overall interaction. And that’s before you get into niggly details around your bot’s tone of voice.
3. Overcoming the machine learning hump: The nature of AI technology such as Watson is to improve over time through machine learning. So on day one of launch, the experience is at its worst and rather unlikely to be very good. But you need thousands of people and interactions to persist with this somewhat crappy experience to eventually make it better.
According to the Senior Director for eCommerce at North Face, their Watson-based personal shopping assistant (referenced earlier and critiqued below) is at a ‘second or third grade level’ – i.e. the intellectual capacity of an eight year old. Tellingly, the only link to it on the website is currently tucked away in the footer. More pertinently, it will take “a couple of years to get the AI just right”. As marketers, we need to ensure we limit the damage of inevitably bad experiences, while continuing to encourage vast numbers of people to interact with our new solution, in order for it to grow. Not an easy balance to get right.
4. Who watches the watchmen? It’s unlikely Juvenal had chatbots in mind when he penned this famous line two millennia ago, but the problem remains the same: if we transfer increasing control of interactions between our organisation and our consumers to machines, who’s going to know when it’s gone wrong?
We can, of course, put various checks in place and you could argue this happens anyway with human-to-human interactions, but one only need look at the flash crash of 2010 or the $23m book listing on Amazon to realise things can go seriously awry if we leave algorithms to their own devices. And what if the problem is not dramatically noticeable, like the examples above, but far more insidious than that? Hopefully it’d just be a few missed sales of widgets or unsatisfactory customer service interactions before someone noticed. But what if a medical chatbot was repeatedly recommending the wrong treatment to thousands of people before anyone realised?
5. Normalising internal understanding of the technology: both from a technical (i.e. developer) standpoint but, more pertinently, from a planning perspective. A variety of stakeholders with vastly differing levels of understanding, using new terminology incorrectly can potentially lead to confused and frustrating discussions, with little progress. To put it simply, it’s important that everyone involved swats up on the basic principles before beginning work to share a similar frame of reference.
Of course the real challenge for many will be greater than any of those listed above: what do we actually want our bot to do?
It’s so easy to get caught up in the buzz and attracted to the shiny and new that we often start acting before we really understand the problem we’re trying to solve.
Take the North Face XPS example referenced earlier (try it quickly now if you haven’t already) Fundamentally I believe it doesn’t work. Not because the Watson technology needs to learn and the experience is consequently a bit rubbish. But because it’s wrong as an idea. It’s a nice piece of PR, but wrong (for North Face right now anyway). Let me explain: the tool is designed to help you select the perfect jacket for you from the 350 in North Face’s inventory, based on a short conversational exchange.
Firstly, that conflicts with both the shopper’s desire to browse, either because they enjoy that experience or more likely want to eliminate buyer’s remorse by checking other options. Secondly, it conflicts with the commercial desire to increase customer dwell time – to upsell or cross-sell complementary products.
Consider the real-world equivalent: when we go shopping and enter a North Face shop on the high street, we don’t march straight up to the shop assistant without looking at any of the items and demand they provide us with the perfect jacket before we scurry out of the shop, jacket in hand, a minute later.
Admittedly, I’ve been a bit sneaky by suggesting the tool delivers just one result (‘the perfect jacket’) after this conversational exchange. It doesn’t. But, presumably, if you’re still going to encourage browsing a selection (as they do), I’d challenge this illusory assumption that typing fractured, prose-based responses is preferable to clicking/sliding a few UI elements in a well categorised and designed online store.
Now, that’s not to say that a tool similar to this could never be useful. I can imagine this working well (once it’s evolved past its third grader level) for a retailer with a truly vast array of stock keeping units (certainly not just the 350 jackets North Face currently offers), or for a search provider that’s indexed all the jackets in the world across multiple retailers. e.g. “Google, show me all dark blue men’s jackets that are lightweight and waterproof but have a retro 60s style, between £80 and £90, with free delivery to the UK” etc. This makes more sense.
Presumably, in time, North Face will expand the functionality to include its entire product portfolio. And perhaps this initial stage of development will help North Face’s agent interact with Amazon’s Echo, or make recommendations based on your Last.fm profile, your Fitbit profile and your geography, etc. in the future too. But I still question what number of North Face’s customers will want to shop this way. I could be entirely wrong of course.
In conclusion, I think many CMOs will soon be asking, or get asked the question: “If chatbots are the solution, what’s the problem?” It’s highly likely that chatbots will be the solution to many problems in the future. But, unfortunately, I imagine there’ll be a lot of wasted time and energy from marketing teams the world over on unsuitable, expensive and ultimately transient applications of this technology in the near future – with the second movers reaping the rewards of their mistakes. Personally, I’m excited to see how the smartest of the first movers take this technology forward, particularly in combination with augmented or virtual reality tech and who’ll win big when they do.
Agree with my North Face critique or think I’m way off the mark? Maybe you’re working on a related project and are experiencing completely different challenges not mentioned here? Tweet me @alexstansfield or @e3_media