September 22, 2024

Whitepaper: How to Dramatically Improve Chat Conversions with Lift AI

Don Simpson

Quick Summary of Findings:

Online web chat is one of the best conversion tools available to marketers and sales teams today, however it is not without challenges. Visitors have a complex variety of needs that are hard to anticipate (especially with chatbots), visitors expect immediate assistance or else they’ll move on, and visitors expect the right, knowledgeable resources to be matched with them in chat.

There are tools and tactics available today to try and combat these challenges, but none of them address the issues entirely. A new machine learning model was created which flipped these challenges around - rather than dealing with them reactively, they would be solved proactively using real-time intent scoring for every website visitor (even if they’re anonymous) which would then determine the optimal chat experience for them (live agent, chatbot, or self-serve) BEFORE they are engaged with chat.

The machine-learning model, called Lift AI, dramatically improved live sales chat performance, with one of their clients seeing a 168% increase in leads in just one month.

Lift AI solves the common challenges prevalent with chat in the following ways:

Visitors have a complex variety of needs
Lift AI’s Machine Scoring model knows visitor intent in real-time BEFORE inviting them to chat so that we can optimize live agents for conversions and productivity, while also optimizing chatbots to match the level of experience needed for lower scoring visitors

Visitors expect immediate assistance
Based on those same scores, Lift AI can quickly throttle chat invitations based on our staffing level and traffic volumes to ensure we provide a consistent, quick level of service with the best results

Visitors expect knowledgeable resources on chat
Because Lift AI knows which visitors need sales assistance and which are likely to only need basic help, it can direct visitors the optimal resource accordingly for the best customer experience

The results of using Lift AI on top of existing chat tools are:

Understanding the full conversion potential of a website
Optimizing conversions and revenue
Optimizing live agent productivity and job satisfaction
Better customer experiences matched to their needs

The following whitepaper unpacks the key challenges in chat, the nuances of each, the weakness of traditional tools, and how machine learning found a unique answer for significant results.

Introduction:

The widespread adoption of chat on smartphones and social media has led to the explosion of Live Chat across all digital content. In fact, messaging and chat have drastically outpaced other communication channels to become the primary tool for website conversion.

Smart brands are deploying live chat to help their customers shop. In fact, 85% of B2B and 74% of B2C companies have adopted chat for sales. However, these companies are not getting the ROI they want from their chat program.

When auditing these chat deployments, we see a basic misalignment between the company’s needs and what the customers expect from chat. Customers engaging in chat have expectations about how that chat experience should be delivered and many of those expectations are challenging to deliver. The gap between customer expectation and what brands can deliver is frustrating customers and brands alike.

Here are the top three challenging customer expectations that are tripping companies up.

Visitors have a complex variety of needs
Visitors expect an immediate response
Visitors expect knowledgeable resources on chat

While these seem like reasonable expectations, they come at a price. Chat tools are increasingly inexpensive and easy to use, but installing chat without a sound strategy can hinder sales resources with unproductive chats, forcing them to miss out on other good opportunities.

Delivering quality chat engagements for a large volume of site visitors on a variety of topics, all within seconds, requires the well-coordinated use of people, process and technology.

As a business that has been delivering chat for large ecommerce companies over 15 years, it was our business to solve these challenges. With over 1 billion profiled website visitors and 14 million chat interactions worth of data on hand, we are sharing what we have learned and developed to deliver on the full potential of live chat for sales.

Visitors have a complex variety of needs

Companies using live sales chat find out quickly that visitors need help with a variety of resources from employment to technical support, customer service, training, research, and ideally sales opportunities. To add complexity, each visitor has a varying degree of interest in the subject. On average, less than 50% of chats end up being good sales opportunities.

For websites with a significant amount of traffic, it is not affordable to chat with every visitor. As companies are onboarding their sales team onto a new sales channel, the last thing they need are those resources frustrated by the volume of non-sales chats. Worse still, their lowered expectations of chat from that experience result in lower conversion rates and a poor customer experience as the sales teams are not equipped or motivated to address the real needs of the visitor. As a result, we see many companies remove chat or hide it deeper on the website on goal pages which will reduce non-sales interactions but also reduce leads or sales overall.

So how can you dispel the fog of chat while keeping your visitors, sales teams and bottom-line happy?

The simple answer is that you need to figure out who should be invited to engage with live sales agents and who to send to support channels or bots. To help with that, the tools available to help companies qualify visitors has been evolving over the years;

Evolution of chat qualification

Basic Qualification: Targeting Conditions

Depending on the chat tool you have, you may be able to use “targeting conditions” such as pages viewed, marketing campaign exposures, demographics, and so on. Once you have defined and configured the conditions, the chat tool will only display chat to visitors that meet those criteria.

Advantages:

This method is a good way to define target audiences or segments that you know have a stronger buying intent based on previous analysis. These segments can also be used to provide more contextual or targeted messaging to visitors which will have a positive impact on response and conversion rates.

Disadvantages:

It requires time and expertise to do the analysis needed to set the targeting conditions, and it needs to be done frequently because the market, human behaviour, and your website are constantly changing. The cycle of analysis-configuration-analysis can become time consuming and complex. It also requires a good technical understanding of how these visitor conditions can be read. For example, are conditions like campaigns stored in cookies or UTM parameters or both?

Our Experience:

We had a very rigorous forecasting process that analyzed visitor site navigation, content being viewed, engagement with the site, visitor brand affiliation, marketing exposures, and more. We then became very good at using site analytics data to identify visitors with stronger sales intent. However, we also found it challenging to adapt quickly enough to ever-changing environments. Plus, some of the best signals of customer intent were the most dynamic and complex to analyze such as visitor navigation, because there are literally thousands of unique visitor pathways through a website making human interpretation difficult.

Overall, these targeting rules enabled us to filter out a significant portion of the "noise". In the end, we still saw about 30 to 40% of the chats being non-sales related (which was better than the 60 to 70% without it), but we also missed conversion opportunities because we just couldn't move fast enough.

Basic Qualification: Chatbots

Like many organizations, we turned to chatbots to help by applying a layer of automation to shield our live agents from non-sales opportunities. Chatbots would intercept visitors through chat, then work through a series of automated questions to help qualify the visitor for further treatment.

Advantages:

Chatbots definitely help to deflect non-sales interactions away from sales teams, and they also apply a layer of automation to deal with simple, repetitive tasks so customers can still get the assistance they need while minimizing the volume of non-sales chat for your live agents.

Disadvantages:

Simply put, chatbots are not as good as live agents when it comes to serving customers - especially on the sales side. Because of this, the deflection approach is not infallible. Whether using an AI or decision tree style chatbot, customers find ways to bypass the process and get to a live agent because ultimately that’s the experience they prefer. Visitors make selections or ask questions that cannot be anticipated by the chatbot upfront, which then sends them on the wrong path. There is a lot of effort being poured into conversational design, but they’re still many years away from mimicking the nuances of a live agent.

In fact, only 9% of people trust bots to help them in any meaningful way and even the best bots will get bypassed intentionally by those visitors who just want to talk to a human. On top of this, visitors become very unhappy when they’re forced to go through an experience they don’t want, leaving a bad impression on the company’s brand.

Our Experience:

Once we accepted what chatbots were good at and where they were challenged with, we were able to successfully use chatbots to deflect a good percentage of our unproductive chats from sales resources. Because of their ability to assist with repetitive questions, they play an important role and they are here to stay. They do, however, take experience, funding, and effort to design, monitor and optimize - they are not a set-and-forget technology.

In the end, we found our best results came from a combination of using pre-chat qualification tools, like the targeting conditions approach outlined above, and chatbots as a second layer of defence. Despite that, given the effort that was required to manage both those tools, we felt we were always behind where we wanted to be and in the end, still dealing with too many non-sales opportunities.

Advanced Qualification: “Machine-Scoring”

We were committed to driving down our cost of sale further, so we took this problem into our R&D department to see how artificial intelligence could help. Using our learning from our forecasting process which identified the attributes that had the strongest signals of conversion intent, we created a machine learning model.

With over 1 billion site visitors profiled and 14 million live sales chat interactions, we had a significant volume of data on top of diversified data which was ideal for AI training. Aside from the need for a volume of data, building an AI model comes with challenges related to wrangling (or “munging”) the data in the right way. As a pay-for-performance business, clean data has always been in our company DNA to help us prove results, and because we did this analysis manually for years, we understood the data well. This really helped us filter the noisy data which otherwise would have made it difficult for an AI model to properly detect patterns. Additionally, we had built the technology to collect data for the model in real-time from any website. We simply had to update that technology to stream the data to the AI model and provide that data back to the site in real-time so we could use it to further enhance our model - but to what end?

The model we built is called Lift AI machine-scoring, which is able to predict a visitor’s level of sales intent much more accurately and quickly. It assigns a score to each individual visitor based on their individual journey, not a segment of visitors with certain attributes, so that we can filter with much more precision. The best part is that our machine-scoring automatically adjusts to change, we don't have to analyze and re-configure technology to keep pace anymore. The scores are calculated in real-time, updating by the second based on the visitor’s behaviour. Now, armed with conversion intent scores for every visitor, we can find and direct high score visitors to live agents for instant sales assistance, while deflecting low score visitors away from them.

In our initial testing, we saw immediate results - our conversions increased significantly.

With Lift AI we could see and engage every high potential visitor on the site wherever they were in that moment. We realized that finding these visitors in the first place was imperative, as over 50% of visitors on any given website could be missed opportunities if not scored. By finding and scoring them, we could proactively engage them. It also reduced the workload of our program management teams because they were not wasting any time on the wrong (low score) visitors. In fact, lower scoring visitors were also being treated properly - being sent to support or chatbots automatically based on their score, and ultimately improving their customer experience too.

Visitors expect immediate assistance

Unlike customers with a specific support issue, every other visitor will likely go to a competitor’s website if they don't feel like they are being helped quickly enough. In our experience, chat abandon rates start to increase considerably after as little as 30 seconds.

Forrester’s Raising the Bar report found that 1 in 5 customers are willing to stop using a product or service for slow response times via online chat.

Today, average chat response times are still too long, which frustrates customers and reduces conversions. Aside from the fact that most live agents are demoralized due to wasting so many hours chatting with the wrong visitors, it’s also almost impossible to assign an agent to every single visitor and have them reach out within 30 seconds - you either don’t have the resources or it would be too expensive to possess those resources. This is compounded by the fact that chat volumes and your staffing levels will need to fluctuate based on various influences like marketing campaigns, world events and working hours/holidays/time zone coverage, all of which makes a consistent level of service that much more difficult.

Traditionally we used out-of-the-box tools in chat applications that would hide the chat invitations when agents were too busy or away, but that also reduced conversion and revenue opportunities. We wanted to maintain revenue on low staffing days and capitalize on revenue on high volume days. We tried optimizing this based on the tools and configurations available - turning lower performing campaigns off and on, adjusting visitor segment profiles, and adjusting the timings of invitations based on staffing hours, however the whole process just became time-consuming. We needed a way to more easily and precisely manage the volumes to our staffing without sacrificing revenue or the customer experience.

That’s where our machine-scoring model shined. Machine-scoring adjusts daily to visitor behaviors regardless of their segments, and we could select specific scores or ranges to suppress so it provided us instant throttling capability. Visitor volume analytics for each score range coupled with our staffing forecasting ability gives us a very precise way to suppress lower performing traffic. On days when volumes drop or spike unexpectedly, we disable lower scores to focus on the higher performing ones with our limited resources. This way we maximize our revenue and deliver a consistent level of service to higher value visitors with reduced wait times. Because of the granularity of the scores, we were more precise with the volume we turned off and we could very quickly do it on the fly so our sales operations teams appreciated how quickly it could adapt and the customers had a much better experience.

Visitors expect knowledgeable resources on chat

Nothing is more frustrating to a customer than getting into a chat only to find that the resource on the other end doesn't have the right information. A critical ingredient of a successful live sales chat program is maintaining well-trained agents and bots. Both require significant effort to deliver good, consistent customer experiences. Some tools are helping companies deliver more consistent experiences by either serving up “best responses” to agents from a knowledge base or using chatbots to directly provide responses. When used correctly, the ability to both automate the work and ensure customers are provided the correct information is significantly impacting a company’s ability to deliver higher quality experiences.

Here are some considerations to keep in mind when trying to get this piece “right”

Content Creation

First is the difficulty in anticipating both the necessary content that needs to be created to answer customer questions and the nuances of how to deliver that content. Studying conversations from sales and support teams will help you identify the questions that need answering but it takes time to understand the multiple ways in which customers are asking the same question and the context of where the customers are coming from in order to develop your content. This will be an ongoing, iterative process that will take time for the output to meet that optimal experience.

Being able to design content that will answer the question in a way that considers each customer’s use case and emotional context is a complex task. Expect that even the right answer given in the wrong way can miss the mark with some customers because predefined content cannot replicate the emotional intelligence a human being can provide.

In practice, companies are devoting considerable resources and time to training chatbots including conversational AI, but they’re a long way off from mimicking human behaviour and consumers can catch on quickly to when they’re speaking with a chatbot.

Use Cases

Understanding where chatbots are most effective and where humans are a better alternative is important when considering use cases. Chatbots are doing great work for certain types of engagements but there are conversations that are still too dynamic, have too many one-off inquiries, and require a level of emotional intelligence that chatbots aren't capable of yet. The most successful programs find the right blend of bot and human—using them for what they each do best.

“According to CCW Digital Research, 71% of consumers are open to using bot technology. Yet they are not seeing the value in practice - only 9% currently trust chatbots to solve meaningful issues."

Source;

Special Report: Omnichannel Chatbots.

To mitigate risk, we used chatbots for small simple tasks and gradually increased their sophistication. The trick, however, was understanding which customers need human intervention before they develop something called “bot fatigue”. Customers do not want a one-size fits all experience and from the brand perspective not all visitors will result in the same opportunity. To drive the best result, you want to roll out the red carpet for your high value visitors and those that need a complex sales dialogue. You don’t want to place hurdles for these customers to jump over in order to achieve their goal, rather you want to compress the process by getting them the best resource quickly, which is going to be your live agents (not your chatbots).

So, how do you sort visitors into those you want to engage live versus those you want to engage with a bot when most visitors are anonymous and we are faced with increasing restrictions due to privacy protection? By the time they engage in chat you have already routed them into an experience that they have to properly navigate in order to get where they want to be. We wanted to know sooner who is more likely going to need a human to facilitate a more intricate sales conversation, rather than running the risk of sending them to a high-friction bot experience first. By knowing in advance of offering chat, we could tailor the experience to ensure these visitors didn’t get stuck in bot jail by quickly sending them to live agents while visitors that had support or needed more qualification would be routed to the best bot experience for their needs.

Again, we turned to machine scoring and used the scores to guide this process. Knowing that visitors with higher sales intent required a "sales conversation" and visitors with lower intent more often have service, support or quick questions, we were able to use the scores to filter visitors into segments and then deliver different experiences to each. Our high intent segments have quick access to humans whereas low intent segments will always start with a bot. The ability to adjust the score ranges to what we consider to be high, mid or low intent gives us lots of flexibility so we could offer different bot experiences and maximize our conversion with the least amount of cost, even to the point of adjusting chat volume to meet real-time staffing capacity. The result is that all the customers have better experiences, those needing quick answers got them from bots and those needing human engagement got to a human quickly. The secret ingredient here which differentiates this technology from the rest is that the machine-learning scoring model is applied and acted upon before engaging the visitor with a chat experience.

“To maximize the conversion of our existing website traffic, we engaged Lift AI to optimize our Drift implementation. We chose Lift AI based on their proven machine learning model and extensive experience in sales chat optimization. The combination of Lift AI’s targeting model and chatbot optimization enabled PointClickCare to realize a 168% increase in qualified leads from chat in the first month!”

-John Walker, Director Demand Marketing, PointClickCare

Next Steps:

After perfecting the machine-learning scoring model and seeing real, significant results with our clients, we have now made the technology available as a SaaS product called Lift AI. It works with any chat tool and is simple to deploy and works out-of-the-box.

The results of using Lift AI on top of existing chat tools is: