Impact of AI on Customer Support

From: redpointai

The advent of AI, particularly large language models (LLMs) like ChatGPT, has profoundly impacted the customer support industry, fundamentally changing how companies approach service delivery and product development [00:00:16]. Intercom, a customer support platform, swiftly pivoted its strategy to integrate AI, recognizing that customer support is highly susceptible to disruption and transformation by these new technologies [00:02:14].

Intercom’s Rapid AI Integration

The release of ChatGPT in late 2022 prompted an immediate and significant shift at Intercom. Dez Trainer, co-founder and Chief Strategy Officer, recalled a rapid “all hands on deck” response, where the team deliberated ripping up their entire AI/ML roadmap to go “all in” on the new capabilities presented by LLMs [00:01:26]. They quickly shipped initial AI-powered features before Christmas, had a “reasonable release” in January, and launched their flagship AI product, Finn, in March 2023, followed by a broader release in July [00:01:58]. This rapid deployment aimed to establish Intercom as a leader in AI adoption within the customer support industry [00:02:55].

AI in Intercom’s Customer Support Products

Intercom’s AI applications in customer service and sales began with “zero downside” features integrated into their inbox [00:03:52].

• Initial Features (Inbox AI): These included functionalities like summarizing conversations, translating messages for multilingual support, expanding on text, and collapsing content [00:03:57]. Such tools addressed real customer support tasks like summarizing issues to create tickets, offering immediate utility with minimal risk [00:04:20].
• Finn: The User-Facing Chatbot: The development of Finn, a user-facing chatbot, was enabled by access to GPT-4 beta, which significantly improved the ability to “contain the bot” and reduce “hallucinations” [00:05:16]. Finn was designed to answer questions based on a high confidence threshold, ensuring trustworthiness and reliability [00:05:39]. Customers emphasized the need for the bot to stay on topic, avoid political opinions, and not recommend competitors [00:05:51].
• Evolution of Features: Inbox AI later expanded to include features allowing users to match or adopt the standard Intercom tone of voice [00:06:09]. Future developments for Finn include enhancing its capabilities within the inbox, such as answering support emails [00:31:58].

Challenges in AI Implementation

Implementing AI in customer support involves navigating several complex challenges:

• Hallucinations and Guardrails: A core aspect of building reliable AI is creating robust “torture tests” to evaluate the model’s behavior across various scenarios, identifying misbehaviors and desired outcomes [00:06:57]. This involves striking a balance: overly constrained models might miss correct answers, while less constrained ones might generate undesirable content [00:07:23]. Effective “prompting” is crucial to ensure the LLM prioritizes specific contexts over its general knowledge [00:08:33].
• Cost Optimization vs. Exploration: While early on, automatic summarization of 500 million conversations monthly would have been prohibitively expensive, Intercom remains in “deep exploration mode” rather than primarily focusing on cost optimization [00:10:01]. The focus is on finding all potential opportunities for AI augmentation across the platform, from reporting to human agent assistance [00:11:32]. The belief is that model costs and speeds will naturally improve over time [00:14:43].
• Latency: Speed is a critical factor, as current AI responses can feel like “modem internet days” [00:12:25]. Faster AI models, potentially running on devices like phones, are anticipated to normalize instant AI experiences [00:12:47].
• Missing Tooling: There is a need for better tooling around prompt management, including versioning, AB testing, and conflict resolution for prompts across different models [00:16:16]. Robust infrastructure, especially for compliance and data sovereignty (e.g., EU server locations), is also crucial [00:16:55].

Deployment and Adoption Strategy

Intercom adopted a “crawl, walk, run” approach to deploying AI, making it easier for customers to “dip their toe in” rather than committing to a full “trustfall” [00:02:22].

• Phased Rollout: This involves allowing customers to initially apply AI only to specific user segments (e.g., free users) or during off-peak times (e.g., weekend support) [00:27:21]. This strategy helps customers gain confidence in the AI’s value.
• Shifting Mindset: As customers observe the AI’s effectiveness (e.g., free users receiving better, instant support), their mindset shifts from cautious interest to a desire for broader implementation [00:28:03].
• Normalizing AI: Broader industry adoption, particularly by major tech companies like Apple and Google integrating LLMs into their consumer products (e.g., Siri, Bard), is expected to normalize the idea of “talking to software” and make customers more accepting of AI as a standard feature in applications [00:30:29]. This will transform AI’s impact on software engineering into a competitive battleground [00:30:45].

The Future of AI in Customer Support

The future of customer support will see varying degrees of automation, with some areas reaching nearly 100% AI handling [00:33:18].

• Automation Levels: The potential for automation depends heavily on the complexity and diversity of queries within a given vertical [00:34:13]. E-commerce, with its limited types of queries (e.g., order status, returns), might achieve full automation [00:33:31]. More complex products, like Google Docs, will likely see high automation (e.g., 80-90%) but may never reach 100% due to the vast array of unique questions [00:34:02].
• AI Taking Actions: Beyond providing text answers, AI will increasingly take actions, such as initiating refunds or canceling orders within other systems (e.g., Stripe) [00:35:07]. While challenging to implement due to the need for robust authentication, monitoring, and error handling, this is a significant area of development [00:37:10]. Some customers may prefer full automation, while others might opt for a “human in the loop” model, where AI proposes actions for a human manager to approve [00:36:00]. This highlights the changing role of AI in transforming job functions within customer support.

Broader Impact on Product Landscape

AI is not just augmenting existing products but is poised to completely reimagine entire product categories, leading to a “product stock” being “thrown out” [00:38:58].

• Reimagining Product Stacks: In areas like advertising optimization, AI could automate entire workflows from ad generation and deployment to performance tracking and optimization, eliminating the need for manual user interaction [00:39:40].
• Advice for Startups and Incumbents:
  • Startups: Should target areas where the incumbent technology stack is “irrelevant” – meaning, if built today, it would be designed entirely differently with AI at its core, rendering existing UI and features obsolete [00:41:39]. This means avoiding areas where core infrastructure (e.g., sending billions of emails) is complex and provides an incumbent a significant advantage [00:40:43].
  • Incumbents: Should first identify “asymmetric upside” areas for simple AI integration (e.g., automatic project titles) to gain experience [00:42:19]. Then, systematically evaluate each workflow:
    1. Remove: If AI can reliably perform a workflow, it should be entirely removed from human intervention [00:43:04].
    2. Optimize: If AI cannot remove a workflow, it should be used to augment it or reduce it to a simple decision [00:43:32].
    3. Enhance: Finally, AI can be sprinkled in for minor enhancements to complete the offering [00:43:51].

AI Trends and Predictions

• Overhyped: Productivity tools focused on generating basic content like emails or sales pitches are overhyped [00:44:35]. Humans will learn to detect AI-generated content, and people might forget what “good writing” truly entails [00:44:42].
• Underhyped: The profound changes AI will bring to creativity (e.g., in art, music, video) are underhyped [00:44:59]. Tools like Kaiber or Refusion offer new forms of creativity, much like Instagram filters democratized photography [00:45:11].
• Industry Adoption: Companies like Adobe, Figma, and Miro have impressed with their quick and useful AI integrations [00:46:11]. However, Apple and Amazon (with Siri and Alexa, respectively) have been surprisingly slow to integrate advanced LLM capabilities, leading to a perceived gap between their “primitive” voice assistants and advanced AI like ChatGPT [00:46:48]. This gap is expected to close, leading to widespread AI adoption and a shift in user expectations [00:47:41].

Intercom continues to explore and develop new AI features, with a dedicated AI lab working on cutting-edge models and potential proprietary solutions, while also prioritizing customer feedback to unlock broader AI usage [00:32:11].