Our Guarantee
Performance backed by our $1 million guarantee
04Fin is the highest-performing Agent for customer service. We know others make the same claim, which is why we back ours with the Fin Million Dollar Guarantee.
Learn more
Fin's CX Model Suite: Purpose-built for customer serviceMeet Fin Apex 1.0The best-performing model
Meet Fin Apex 1.0The best-performing model
for customer service
THE WORLD'S BEST SUITE OF MODELS FOR CXMost models are built to be good at everything. Fin Apex 1.0 is built to be exceptional at one thing: customer service.
Most models are built to be good at everything. Fin Apex 1.0 is built to be exceptional at one thing: customer service.
Fin Apex 1.0 is the flagship of the Fin model suite - our first model that generates the final answer Fin gives to every customer.
Built by our world-class AI team and trained on years of support interactions, Fin Apex 1.0 is optimized for customer service - always providing accurate answers, following your policies reliably, and knowing when not to answer.
In production, Fin Apex 1.0 outperforms frontier models with higher resolution rates, fewer hallucinations, and faster responses.
2.8% higher resolution rate
0.6 seconds faster time to first token
-65% reduction in hallucinations*
*Compared to Sonnet 4.6
Fin model suite
Introducing the
Fin model suite
- [ 1 ]
Fin Apex 1.0 [ new ]
Fin Apex 1.0 is the model that produces the final answer Fin gives to every customer. It takes the most relevant content surfaced by the Fin model suite, applies your policies, and produces a direct answer or decides the question needs a human. Every output is grounded in your knowledge base, not inferred from general training data.
Specification- 1.1Grounded in your knowledge base, not general data
- 1.2Answers directly, escalates honestly when it can't
- 1.3Follows your configured guidelines every time
- 1.4Trained to minimize hallucination at the source
- 1.5Post-trained on Fin's own production data
- 1.6Compounds in accuracy as the suite improves
- [ 2 ]
Fin Retrieval
The Fin Retrieval model scans all available knowledge sources and pulls out a small set of useful information.
Specification- 2.1Understand intent behind the question
- 2.2Search across all available content sources
- 2.3Match on semantics, not just keywords
- 2.4Select top N most likely candidates
- [ 3 ]
Fin Reranker
The Fin Reranker model takes the retrieved content and scores each piece for relevance, accuracy, and usefulness in context. It then selects the final piece(s) for the LLM to use.
Specification- 3.1Score relevance with respect to query
- 3.2Evaluate context match and resolution fit
- 3.3Downrank outdated or low-confidence sources
- 3.4Output: final selected content for generation
- [ 4 ]
Fin Issue Summarizer
The Fin Issue Summarizer model detects and extracts the user's issue from conversation history. Built on a fine-tuned 14B model, it transforms unstructured chat exchanges into clear, actionable issue summaries for downstream retrieval.
Specification- 4.1Detect whether conversation contains an addressable issue
- 4.2Extract summarized issue from multi-turn exchanges
- 4.3Handle edge cases: greetings, feedback, and noise
- 4.412.5% cost reduction vs frontier LLMs
- [ 5 ]
Fin Feedback Parser
The Fin Feedback Parser model uses a multi-task ModernBERT architecture to interpret user responses. It classifies feedback sentiment, detects follow-up questions, and identifies conversation endings with state-of-the-art accuracy.
Specification- 5.1Classify feedback type: positive, negative, or none
- 5.2Detect follow-up questions in user messages
- 5.3Identify when conversations have ended
- 5.4Match LLM accuracy at a fraction of the cost
- [ 6 ]
Fin Language Detector
The Fin Language Detector model uses XLM RoBERTa to accurately identify the user's language across 45 supported languages. It handles real-world challenges like typos, short messages, and script mismatches.
Specification- 6.1Support 45 languages with high accuracy
- 6.2Handle spelling mistakes and short context
- 6.3Process script mismatches (e.g., Romanized Hindi)
- 6.4Enable intelligent language fallback logic
- [ 7 ]
Fin Escalation Router
The Fin Escalation Router model uses a multi-task ModernBERT architecture to decide when conversations should be handed off to human agents. It provides reasoning and cites matching business guidelines with over 98% accuracy.
Specification- 7.1Decide: escalate, offer to escalate, or continue
- 7.2Cite matching business escalation guidelines
- 7.3Provide reasoning for escalation decisions
- 7.40.5s faster than LLM-based routing
AI TeamBuilt by a world-class
Fergal ReidChief AI Officer
Brian McDonnellSenior Director, Engineering
Mario KostelacPrincipal Machine Learning Engineer
Alexey TarasovSenior Manager, ML Engineering
Fedor ParfenovPrincipal Machine Learning Scientist
Pedro TabacofPrincipal Machine Learning Scientist
Molly MaharPrincipal AI Product Designer
Pratik BothraPrincipal Machine Learning Engineer
Rati ZvirawaSenior Director, Product Management
Rob ClancyPrincipal Machine Learning Engineer
Built by a world-class
team of AI experts
02Building models of this quality is only possible thanks to our 50-person, world-class AI team, led by Fergal Reid, and our decade-long experience building customer service software.
AI Group Leadership
Security
Customer-first approach to data privacy and security
05We have invested heavily in data privacy and security. These models are trained on deidentified data from past Fin interactions.
The following existing customers are automatically excluded from our training process:
- Customers that have a BAA in place with us
- Customer using an EU/AU regionally hosted workspace
- Any customer that was previously granted an opt-out
FAQs
FAQs
What is the Fin CX Model Suite?
The Fin CX Model Suite is a system of seven purpose-built AI models, each handling a discrete stage of resolving a customer support query: detecting language, summarizing the issue, retrieving knowledge, reranking results, generating the final answer, parsing feedback, and routing escalations. Unlike general-purpose LLMs, every model in the suite is trained specifically for customer service rather than broad capability.
How does Fin Apex 1.0 compare to general-purpose AI models?
Fin Apex 1.0 is post-trained on Fin's own production data from real support interactions, not general internet data. In production, it delivers a 2.8% higher resolution rate, responses 0.6 seconds faster, and 65% fewer hallucinations compared to Sonnet 4.6, because it is optimized for one task rather than trying to be broadly capable.
How does Fin decide when to answer versus escalate to a human?
The Fin Escalation Router model uses a multi-task ModernBERT architecture to choose whether to continue, offer escalation, or escalate the conversation. It cites matching business guidelines, provides reasoning for each decision, and is 0.5 seconds faster than LLM-based routing, with over 98% accuracy.
Is my customer data used to train Fin's models?
Fin's models are trained on de-identified data from past Fin interactions. Customers with a BAA in place, those using an EU or AU regionally hosted workspace, or any customer previously granted an opt-out are automatically excluded from the training process.
Who built the Fin CX Model Suite?
The suite was built by a 50-person AI team at Intercom, led by Chief AI Officer Fergal Reid, combining over a decade of customer service software experience with specialized ML expertise. The team trains purpose-built models at each stage of the pipeline, including fine-tuned smaller models like the 14B Issue Summarizer and transformer architectures like ModernBERT for feedback parsing and escalation routing.