This is an example of how Savv-e uses gamification to match the user experience with specific learning outcomes.
In this case, we designed an emotional 'chat-bot' to train staff in the finance industry in dealing with angry customers. The results were surprising and a lot of fun.
In a previous post, I wrote about the transformative power of games for learning design.
I argued that games, unlike simple quizzes or tests, put the learner in a position where they need to ‘transform’ to achieve the best outcome. And they’re encouraged to do this because overcoming failure is part of the fun.
In this post, I’m going to run you through an example of how we put this into practice at Savv-e, using gamification in business and learning, and some of the challenges we encountered along the way.
Recently, we were approached by a client in the financial services industry. They wanted to train new staff in how to deal with angry customers, in response to an increase in verbal and physical attacks when on the job.
The learning outcomes included safety procedures, information on how to read aggressive body language and how to de-escalate a threatening situation as they may encounter them.
For us, this was an exciting opportunity, not just to build an interesting piece of learning architecture, but to build a game that might transform the relationship between front line staff and their customers.
After all, in these situations, both the person working at the financial institution and the customer are playing parallel 'games'. They both have goals and might see the other person as their key obstacle to be overcome.
The staff member wants to de-escalate possible conflict and resolve the customer’s issue, and the customer wants their issue resolved, but in each case the agenda is different.
We would present the staff member with a game where the apparent goal was to 'solve' a problematic customer, but the true goal was to help that customer achieve their own outcomes. And they could only achieve that goal by transforming their perspective of the encounter.
In other words, we'd train them to empathise.
Our proposal was to present the learner with a gamified simulation where they would have to de-escalate a realistic situation. We'd present them with a virtual customer with an obvious financial problem. However, they would also have a hidden personal problem which was making them unreasonably angry.
The learner would have to choose between a number of preferable and less preferable conversation options that could bring them closer to resolving the financial issue and unlock new information about the personal issue which would provide context for their anger.
Getting Rid of Unimportant Choices
Traditionally, this kind of scenario might be built with a ‘branching narrative’, one of the key tools in an instructional designer’s arsenal.
A simpler cousin to interactive fiction games or the ‘choose your own adventure’ books, branching narratives allow the learner to choose between a few possible plans of action and see the consequences of those actions. Like this:
To make this manageable from a design perspective, they also usually have ‘choke points’: after one or two choices, the narrative will bring the learner back to the central path.
But in real life, the effects of our choices ripple out, closing off some doors and opening others, affecting the way other human beings choose to interact with us in return. For example, if you picked ‘choice a’, screen 3 should be fundamentally different than if you picked ‘choice b’.
Choke points reduce the 'customer' to a plot device, by robbing the player's bad choices of their lingering effects. We needed the virtual customer's opinion of the player to persist, accumulating good will or anger depending on the choices.
On the other hand, accounting for every possible scenario would be inefficient and frustrating to build.
An AI or open world scenario wouldn’t be fit for our purpose either. We wanted to keep the number of choices contained for the same reason that we wanted to keep them open: all choices needed to be impactful, even the bad ones.
Instead, we compromised with a fake AI who could hold an organic conversation, but couldn’t actually drive it.
Bob who would have to exhibit a range of emotional states. He would need to respond convincingly to the learner's actions, demonstrating his emotional state through visual and verbal cues. The learner would need to read these cues to understand whether their choices were escalating or de-escalating the confrontation, which meant context was as important as how he felt 'right now'.
In other words, Bob's level of aggression would persist through the entire simulation. He would be directly impacted by the learner’s choices and the order in which they made them. And that meant every choice the learner made would carry weight.
Mistakes would be more than points on a screen, they would have lasting effects that carried through the conversation.
Bob had to respond to questions on a number of topics, including himself, his account and his personal situation where it impacted on his problem with the institution. Those responses had to be modulated by his emotional state and the questions that had come before.
We knew that the learner’s success would depend on two factors: whether or not they could resolve Bob’s issues, and whether or not they had empathised with and calmed him down in the process.
To achieve this, we created an ‘aggression counter’ to track how angry (or calm) Bob got, on a scale of 1 to 10. We reasoned that Bob couldn’t be any calmer than ‘1’, and if he exceeded ‘10’, he’d be so angry that he’d storm out of the building, ending the simulation.
We also built a learning string, which would be the learner's (hidden) score. The client provided us with a list of positive and negative strategies for de-escalating conflict, so we assigned each one a letter.
The learner's achievements, missed opportunities and poor choices would be presented in a granular feedback screen as an itemised list, and the string would help us keep track of them.
To make the conversation feel organic, our branching narratives were actually made up of free-floating blocks that could be popped into the conversation at almost any time. But unlike a traditional branching narrative, these blocks didn't link to separate outcomes, they simply affected Bob's aggression score and the learning string, as well as the language Bob used to phrase his response.
For example, asking Bob about his account might be predicated on having asked Bob for ID. On the other hand, being clear with Bob about his threatening tone could be unlocked by Bob’s aggression level getting too high.
These responses were then scripted and recorded by a voice actor in five different emotional tones. The emotional states were also linked to a different actor, performing five different facial expressions giving Bob the functional sound and appearance of an angry customer.
Getting Up Bob's Nose
These two sets of variables interacted in surprising ways. Being helpful and practical with Bob would make him calmer, but not as calm as if you were empathetic and could demonstrate you were listening. Similarly, being officious and obstructive would make Bob angry, but being rude or showing a lack of empathy would make him even angrier.
These questions and their responses also had to be massaged so that each path felt organic and unique.
One thing we hadn’t expected was the way the learner’s experience could be transformed by the context of the choices themselves.
For example, a learner could take a constructive path through the conversation, resolve Bob’s issue and leave him feeling calm and reassured.
Or, the learner could choose those same options in a different order and come across as officious and functional, solving the customer’s financial issue while ignoring their feelings.
Alternatively, the learner could act like a sociopath, teasing Bob about his predicament, then guilting him into calming down, before goading him again. This would usually end with Bob ending the conversation, but could end with him remaining angry despite his problem being resolved.
All this with the same set of 12-15 questions.
Transformation in context
So, we didn't build an artificial intelligence. But by providing just enough material for the customer to feel organic, we created an effective solution for empathy training.
Bob's responses would take on new meaning depending on the context and attitude that the learner brought to the situation. They would fill in the gaps of Bob's behaviour by intuiting how their choices had inspired these changes.
And in recognising their impact on the character, eventually 'winning' the game, the learner would engage in a transformative understanding of their own role in an interpersonal conflict.