Chatbot service agents and more responsive AI-powered processes become available to Salesforce Service Cloud customers today — provided they have enough historic data for the machine learning to work with. But even as little as a couple of months’ data is often enough to get started, according to Salesforce product managers who briefed diginomica in advance of the news.
The three new AI-powered capabilities round out the existing Service Cloud Einstein offering with a configurable chatbot, guided workflow and suggested next actions. All three functions are focused on providing a guided customer experience and thus improved customer satisfaction by combining intelligent automation with more efficient human interactions, says Bobby Amezaga, Senior Director, Salesforce Service Cloud Product Marketing. The three functions are:
- Einstein Bots for Service — a new chatbot capability that can be set up without any coding to automatically handle routine customer requests. Initially designed to work with webchat or SMS — mobile messaging and voice interaction will come later — the chatbot can hand off to a human agent whenever needed. It’s available now for customers who have Service Cloud Unlimited Edition or Live Agent licenses.
- Lightning Flow for Service — provides the ability to set up guided service processes for customers or agents that use machine learning to automatically determine each step, based on the context of the interaction. Available now with all Service Cloud editions including the entry-level Essentials.
- Einstein Next Best Action — drives a side panel in the agent’s screen with suggested actions and offers, selected using intelligent analysis of the customer context based on rules-based and predictive models. Currently available in pilot, the aim is to surface insights at the moment of maximum impact to speed case resolution, improve customer retention or maximize upsell.
The customer experience focus is in contrast to previously introduced Service Cloud Einstein functions, which have focused more on improved efficiency within the customer service team. These include automated prioritization of cases and field service schedule optimization, case routing, recommended responses and predicted case resolution times.
Machines to augment, not replace humans
One of the characteristics all of these functions share in common is the emphasis on helping service agents or self-service customers complete tasks more effectively. The aim is to augment the human element rather than replace it. So in the case of the chatbot, help from a human agent is there when it’s needed, says Amezaga:
It’s not just that the bot can solve an issue on its own, but if the bot cannot handle an issue or the customer wants to speak to a human, the bot is collecting information and seamlessly hands off to the agent.
That combination of machine automation and human intervention has been found to lead to higher customer satisfaction scores than one or the other alone, he says. If the bot can deal with the issue, the customer gets a faster resolution, but if not, the handover to a human agent ensures that the information the bot has already collected is passed on:
Speed is definitely a component. When you engage with a bot you don’t have to wait for anything or anybody. You can immediately engage and start to get your questions answered. If it can’t answer [and hands over to an agent], you’ve still had a valuable interaction.
AI etiquette and the customer experience
Customers will always know whether they’re interacting with a machine rather than a human, he adds, even during a web chat or SMS messaging session. It’s a matter of AI etiquette, he explains
We’ve done a lot of research around how to use AI responsibly and productively for services organizations.
Some of the best practices we’re sharing include being transparent upfront. Another example could be to provide a failsafe for a customer to speak with an agent if they choose.
One of the most sought-after AI capabilities is the new Next Best Action functionality. Customers are keen to use this because it can help generate extra revenue in the customer services environment through counteracting attrition or encouraging upsells. It’s able to do this by looking at the individual customer record and the context of the interaction, and then evaluating the most valuable next steps based on what’s happened in other similar engagements. These recommendations are then offered to the agent who makes the final decision as to what to do next.
How much data is enough?
Many customers already have some form of rule-based system in place, so applying machine intelligence is the obvious next step. While it’s important to have enough historic data for the machine learning to work with, that’s rarely a problem, says Marco Casalaina, VP of Product Management for Einstein.
You need to have some history of having made these offers to have a basis on which to predict them. Many of our customers often have enough data in the system to make these predictions or can collect it in just a couple of months. It hasn’t been a barrier up to now.
The prediction that you’re making can be about specific offers, but there are more general predictions that have to do with the customer. It might predict whether you’re going to attrit. That might drive whether I make a retention offer or an upgrade offer.
The historic data also has to be reliable enough, although interestingly Casalaina says the team has learned from experience that there’s “a certain pattern of noisy data we often find in Salesforce.” In a recent InfoQ article, one of his colleagues details how the team deals with the phenomenon of hindsight bias, which frequently occurs in enterprise datasets due to data being recorded close to an outcome.
How Einstein learns from customer pilots
For example, sales people are most likely to complete a ‘deal value’ field when a lead is close to being converted, which creates a false correlation between that field being filled in and the likelihood that the lead will be converted. Having encountered these and other examples, the team has been able to automate ways of detecting and mitigating hindsight bias.
Nevertheless, the team is being cautious how it introduces the next best action functionality, which is why it’s initially available only in pilot. Unlike most consumer and academic AI projects, Einstein is designed to work reliably even though its data scientists can’t see the actual data customers will use with it, for data privacy reasons. That means taking extra care to check for unexpected outcomes, as Casalaina explains:
These AI capabilities, we tend to have a fairly long pilot process on them. We can’t see the data. We’re not driving the car, we’re building the autopilot. We do have fairly long pilot processes to make sure we get this right.
Salesforce continues its strategy of offering pre-packaged AI capabilities that help customers achieve real-world business outcomes. The emphasis on an improved customer service experience in this latest batch of tools is welcome, but the vendor is also right to pace delivery so as not to raise unrealistic expectations as to what can be achieved.
[Updated final quote from Marco Casalaina per comment below, to correct a previous version which said “We’re not building the [autonomous] car, we’re building an autopilot”]
Image credit - via Salesforce
Disclosure - Salesforce is a diginomica premier partner at time of writing.