Intercom’s patented AI Engine™ allows Fin AI Agent to refine every query, optimize every response and validate the quality of each answer. As a result, it's the only AI agent that can balance industry-high resolutions with industry-low hallucinations.

Many AI agents can optimize toward either high resolutions or low hallucinations, but struggle with doing both. This is because they often simply apply a ‘wrapper’ to a generative Large Language Model (LLM), instead of building on top of it, because it’s cheaper and easier.

However, this approach ignores the tendency of LLMs to hallucinate and expose your customers to incorrect or irrelevant information. Without a system like Intercom's that refines the LMM’s inputs and outputs, an AI agent cannot be effectively optimized for accuracy and reliability.

In order to optimize the accuracy of an answer that an LLM generates, the inputs the LLM receives must be refined for comprehension. The clearer and more understandable the query, the better the output.

Often in customer service, customers can write in to support without fully explaining or contextualizing their query. To solve this problem, the Intercom AI Engine™ has been designed to refine the inputs that are sent to the LLM. This ensures that each customer message is optimized in terms of its meaning and context so that the LLM has the best possible chance of producing an accurate answer.

In addition, the Intercom AI Engine™ performs checks to see whether a Workflows automation or Custom Answer should be triggered based on the topic and context of a customer's query, as well as performing safety checks to filter out anything that Fin shouldn’t be answering.

Once a query has been checked and optimized the next stage is to generate a response using the LLM. For this task, the Intercom AI Engine™ has been designed to use a bespoke and enhanced retrieval augmented generation architecture (RAG for short).

RAG is a process that involves retrieving relevant information from a data source and combining it with a user's prompt before passing it to an LLM. This additional context helps improve the model's output by enhancing its base knowledge, and in doing so reducing the risk of inaccuracies like hallucinations.

Intercom’s application of RAG is highly unique. The AI group at Intercom has invested heavily in optimizing our unique application of RAG and continuously tests both the accuracy of the LLM and the individual steps of RAG to improve overall performance.

In the final step of the process, Intercom’s AI Engine™ performs checks to understand whether the output from the LLM meets the necessary response accuracy and safety standards. Many checks are performed, covering whether there is enough confidence in the response, the response is accurate enough, and whether the response is grounded enough in reality to address the question adequately.

To calibrate and enhance engine performance, the Intercom AI Engine™ has advanced integrated tools that help optimize answer generation, efficiency, precision, and coverage.

Fin customization and control - Intercom has incorporated features and tools designed to help users customize and control how Fin responds, what it can do, what information it can use and much more. Each of these pieces plays a part in how well Fin performs. The more it knows how to do, the more it can do, and the more you will be able to automate your support with human-quality customer experiences.

AI analytics and reporting - The Intercom AI Engine™ has been designed to facilitate analyses on the effectiveness of each stage of the answer generation process. This gives the AI group at Intercom the tools they need to improve each stage of the process and overall performance. A rigorous amount of testing happens before any changes are made to the AI engine architecture, which takes into account how each small change impacts the engine as a whole. In addition, the AI engine offers Intercom users access to pre-built and customizable reports that help to understand where Fin is working well, and what can be improved.

AI recommendations - the AI engine offers recommendations to continuously improve performance over time. This ranges from identifying which content could help fill gaps in Fin’s knowledge, to highlighting underperforming content that should be further optimized, or suggesting actions should be set up to help Fin can resolve more queries for customers.

Relying exclusively on the generative capabilities of an LLM to answer or solve a customer problem is not a reliable way to serve customers. Without the proper safeguards in place, LLMs can be open to manipulation or hallucinations, which could then impact your customers. To ensure safety and reliability, Intercom’s AI engine™ has been designed with strict safety controls at each stage. If the necessary parameters for safety have not been met at any of the steps or stages, Fin will let the customer know that it cannot answer the query and escalate to human support.

Intercom has implemented state-of-the-art security measures to protect Fin against a wide range of LLM threats, including those identified by the OWASP LLM Top 10. By consistently testing a variety of high-end LLMs, and deploying rigorous internal controls, security protocols, and safeguards, Fin is able to achieve the highest level of security and reliability while avoiding potential limitations and threats.

That means you and your customers can always trust Fin’s answers as the safest, most accurate, and reliable of any AI agent.

Comprehensive overview of the security measures and testing protocols implemented for Intercom’s Fin AI features available here.

Fin AI Agent is available on US, EU and AU hosted workspaces.

Intercom has international accreditations and controls in place in order to ensure the highest standard of safety and security, including:

Customer data is not used for model training by LLM providers. Any data submitted to an AI Product becomes an Input, used to generate an Output (as those terms are defined in our Additional Product Terms). OpenAI is contractually required to use Inputs and Outputs: