Will AI kill the customer support industry?
Is AI on the brink of obliterating the customer support industry? Klarna's latest news shows it replaced about 700 jobs. One may wonder, rightfully, what the future of this industry is.
Last week, we learned that Klarna built their own AI assistant using OpenAI, and it has been live for a month with spectacular results1. They announced that their AI is handling the workload of approximately 700 full-time agents across about 2.3 million conversations (two-thirds of their customer service chats).
It is not surprising that support roles are somewhat at risk due to AI, even today. However, this announcement has transformed something abstract into something very tangible. In this article, I would like to share some thoughts on what I believe may or may not occur.
It's noteworthy that AI is still in the early stages of development. Besides Klarna's case, major use cases demonstrating its ROI are yet to be seen. One thing is certain: progress is ongoing, with most researchers2 expecting AGI to emerge by the end of this decade.
Open source
I want to explore the points raised by the host of the All-in Podcast last week, particularly the question: Why wouldn't Klarna make this AI agent they built internally open-source? The key argument presented by the host is that it would be a net positive for them because (1) they would have a community improving the AI assistant, allowing them to reduce engineering costs internally, and (2) it would advance the whole industry by putting pressure on high-priced support software. Usually, open-sourcing often occurs when markets saturate, or proprietary software loses its competitive advantage3.
I tend to agree with the podcast’s hosts; going open-source be a smart move for Klarna, likely yielding fast results and having a broad impact on the industry. After all, Intercom prices its Fin AI assistant at $0.99 per resolution, which would have been a significant cost for Klarna.
Now, there are two important questions to consider:
If AI is a commodity4, what prevents AI agents from being commoditized through open source?
Would this AI assistant be able to perform well across all support tiers and channels?
I suspect that the first question cannot be answered without addressing the second question regarding support tiers.
AI and support Tiers
As you know, Tier 1 handles basic inquiries and common problems, escalating more complex issues to Tier 2, where experienced personnel offer advanced technical support. Tier 3 involves specialists for highly specific or intricate issues.
Across industries, the majority of queries (70-80%5) are in the first tier, involving basic support questions. This tier is commonly automated, sometimes extensively, using basic chatbots. The main issue with current chatbots is not their ability to provide answers, but rather the challenge customers face in reaching those answers due to rigidity. AI assistants can address this issue, but as we have experienced at Rasayel, it still requires very good AI retrieval and knowledge from the businesses. As a customer issue moves through the different tiers, it is less likely to be automated as it often requires more interlinked or very deep-specific knowledge.
Where we have seen the most friction at Rasayel when building our support AI agent, specifically for WhatsApp, is the ambiguity between Tiers. Out of the box, AI agents aren't good at detecting if an issue is likely to be on Tier 1 or 3, mainly due to the lack of contextual knowledge. Furthermore, I believe that these open-source agents, for now, will be very efficient at working within Tier 1, as is probably the case for Klarna. After all, these customers in Tier 1 could probably just answer their questions in a well-designed FAQ. What the AI agent does in this case is that it brings the FAQ within a conversational format.
AI alone will have limitations in advancing through tiers, particularly in Tier 2 and 3 of certain industries that demand specific knowledge of how systems operate. For instance, a SaaS business may need a support engineer well-versed in their evolving backend systems for these tiers. To reach Tier 2 and 3, AI must be closely integrated into systems, which may still take years to be the case.
To make this more practical, I requested GPT-4 to create a table listing the responsibilities in each tier and their likelihood of automation by AI, specifically for support in the SaaS industry.
The good news for us working in the support industry is that after a few months of using our own support agent across tiers, I came to a major conclusion: it’s not only about the complexity of the query.
Context is key: We notice it on a daily basis; our customers are often multichannel. They create relationships with their customers, meaning that they have a video call with them, followed by an email and a WhatsApp message. Unless the AI can capture this important context, most customers will get frustrated as they need to repeat their query over and over again.
The human touch: This is somehow a more intangible observation, but some customers buying specific products or services just have a different set of expectations. Imagine you buy a subscription for a premium travel service; would you be okay to talk to an AI? Well, this is something we have seen even at Rasayel; many customers just ask for a human directly. They want to create and maintain a relationship. The human touch extends to relationship-building as well. Fostering trust and interpreting subtle cues play a big role in creating long-lasting customer service.
Market impact
It is obviously difficult to predict the impact on the market of first open-sourcing software. But following the simple logic outlined above, here is my prediction:
Support software operating and selling to B2C companies that mainly operate in Tier 1 have probably reached their peak. In this case, I am mainly referring to market leaders such as Intercom. On the other hand, if your support software sells to complex B2B, you definitely have more time to prepare for this shift and find where your software's value really lies.
I wouldn’t underestimate the advancement of AI and deep integrations in every software that will happen in the coming months/years. I see, at least for us at Rasayel, many great opportunities to distinguish ourselves by creating clear value on top of these AI agents.
Conclusion
My conclusion is threefold. Firstly, it is evident that the major players in the support industry are at risk. Their key advantage lies in strong integrations with specific channels and the lock-in they maintain with large enterprise customers. When you have a high market share, the more likely way is down.
Secondly, if you are a support agent working on the front lines (Tier 1), I strongly advise you to read this article from Intercom about the new jobs it may create, such as Knowledge Managers, Conversation Designers, Conversation Analysts, Support Design Strategists, etc. As they outline, new jobs will be created, and it is time to jump on board.
Lastly and most importantly, it is crucial for every customer service SaaS to consider how they will add value beyond AI agents (Tier 1), as open-source alternatives could significantly affect them. Their strength may lie in their refined data models or seamless integrations. Whatever it is, they have a lot of work ahead. Best of luck to you all!
At Rasayel6, we are aware of all these challenges and have an experienced team in place and growing fast. If you want to join our design team and get to work, my DMs are open.
Data referring to Ark Invest Big Ideas 2024 page 10 “AI Is Accelerating Faster Than Forecasters Anticipated” — These forecasts are limited and unreliable, like any forecasts.
While there is no direct hard evidence for this statement, there are many indirect pieces of evidence. Here are some:
Assumptions of AI being commoditized are made based on the following:
The data heavily depends on tier segmentation for companies or industries. In other words, these numbers may be inaccurate depending on your company or industry. Support complexity typically follows a Pareto distribution, with most complexity concentrated in the distribution's tail.