‘It Delivers’: Intercom’s Next-Gen Fin AI Agent Switches from OpenAI to Anthropic

Wrapping up a week packed with AI agent announcements, Intercom has unveiled Fin 2, the newest version of its customer service bot. The company claims its agent will be able to “answer more questions, in more ways, in more places.” To accomplish this, Intercom has added features designed to improve Fin’s intelligence and help it make better decisions.

However, the most significant change to the Fin agent is that it’s now powered by Claude, Anthropic’s sophisticated large language model (LLMs). It’s a marked change from when Fin launched in 2023 and used OpenAI’s ChatGPT.

Disclosure: I used to work at Intercom as a senior editor. This post is based solely on public information and my independent analysis. The company did not influence or compensate me to write about this news.

Dropping OpenAI For Anthropic

For those who may not be well-versed in large language models (LLMs), the significance of this decision might raise questions. On the one hand, it’s striking, given the extensive media coverage surrounding OpenAI and ChatGPT over the past year, mainly due to their prominent partnership with Microsoft. From a public perspective, this move could be seen as a notable win for Anthropic.

However, it’s not uncommon for software providers to adopt a model-agnostic approach. Several companies have shared that while they utilize GPT to train their agents, they also incorporate other models into their systems. Although Intercom’s switch to Anthropic is significant, it may also reflect a strategic decision based on development timelines—the company likely aimed to align this transition with other updates for a major release.

Much has changed since Fin was first released. In 2023, OpenAI’s ChatGPT was the leading model everyone was discussing. However, this year has seen a surge in new LLMs, each boasting improved benchmarks and capabilities, creating a more competitive landscape.

So why did Intercom decide to switch to Anthropic? Before examining the company’s official reasoning, it’s important to highlight that there is a reciprocal exchange of services—Intercom is utilizing Claude while Anthropic is employing Intercom’s Fin. But in any event, according to Intercom co-founder and Chief Strategy Officer Des Traynor:

We landed on Claude for one simple reason: it delivers. And it doesn’t just deliver faster speed or scaled operations, but also high-quality service, performance, and reliability.

In an interview with Fortune, Trayor denied that a specific problem caused the shift from ChatGPT but that it was the result of an evaluation in which Claude performed better. Interestingly, the release of Fin 2 came more than a month after Anthropic introduced an enterprise plan for Claude.

Intercom claims the Claude-powered Fin has delivered an average resolution rate of 51 percent “across thousands of Intercom customers and millions of conversations.” It’s a marked improvement from the 23 percent that the first-generation bot achieved.

Keeping Up With the Customer Service Chatbots

Intercom was one of the first to capitalize on using AI in customer service. Initially, it intended to create a bot that could converse naturally with customers, answer questions about the business, minimize the risk of hallucinations, and be easy to set up. Over a year later, Intercom is no longer alone in developing these agents as it faces competition from Salesforce, Zendesk, Sierra, Microsoft, Freshdesk, LivePerson, HubSpot, Decagon, and others. The goal is to be the first to prove their AI can deliver human-level-like support that consumers can’t tell the difference.

The updates to Fin aim to keep Intercom at the forefront for companies by enhancing the agent’s performance and accuracy.

So, what’s new with Fin 2?

How Fin Learns

The first change to the bot is the addition of a Knowledge Hub, a repository for teams to control, update, and manage all the content Fin will be trained on. This is similar to Atlassian’s Rovo, an AI-powered search engine with a central storage area for the internal content, external websites, PDFs, and/or databases companies want the bot to learn from.

It’s not bad to have this container, especially if you want Fin not to hallucinate and only deliver responses specific to your business. Intercom claims the next-generation Fin agent can combine knowledge from multiple content sources to generate tailored customer responses.

How Fin Behaves

The second Fin update involves new settings that give administrators more control over how the AI agent communicates with customers. These include changing Fin’s tone of voice using five available presets—should it be neutral-sounding or matter-of-fact, professional, friendly, humorous, or custom? There’s also an option to choose the length of Fin’s responses—should they be concise or more thorough?

Moreover, with its new multilingual support, Intercom’s Fin can detect and resolve issues in over 45 languages. With this talent, the agent can provide real-time translation, meaning responses can be created to match the customer’s language preference. However, this capability won’t be available until later in 2024.

AI Category Detection has also been added, enabling companies to determine how specific topics are detected and addressed.

If, for example, a customer asks about refunds, cancellations, reports a bug, or event seems frustrated based on the conversation, Fin automatically categorizes those conversations and will route them according to your settings.

Later this year, Intercom will make it easier for administrators to use natural language to instruct Fin on the correct support policies and procedures to follow.

The final behavioral modification is the expansion of Fin to more digital channels, including email and WhatsApp.

What Fin Can Do

Now that we’ve explored Fin’s training and behavioral framework, let’s delve into the new outputs it can generate. To start, Intercom has enabled the AI agent to tap into external data sources such as Stripe and Shopify to provide more personalized answers and perform complex tasks based on customer inquiries.

Customers can also prompt Fin to change their information, such as shipping address or adjust their subscription.

Intercom provides a template for companies to create actions using natural language.

Learn How Fin Performs

The final component of Fin 2 focuses on analytics and rethinking a widely used metric called CSAT. Intercom describes it as being both “incredibly valuable and incredibly flawed” because it’s merely a “fraction of the total customer conversations” a team handles.

Fin 2 will generate an AI-created CSAT to correct this, providing companies with an analysis they can use to evaluate their teams’ conversations and better understand how their customers are receiving their service.

A new AI-powered Conversation Quality report also displays high- and low-performing topics. This could help inform companies about customer service content gaps.

Lastly, Intercom has released a Holistic Overview Report, consolidating insights into a single dashboard to show the entire support operation serviced by human and AI agents.

Intercom isn’t raising its prices for Fin 2, saying it charges $0.99 per resolution, but only if one is delivered. Should the AI agent be unable to generate a response, “Fin is free to use.”