Sources & Research

“40-60% of support tickets are repetitive”

This is a well-established industry finding, supported by multiple sources:

• Zendesk CX Trends Report (2023): Reports that common, repetitive questions make up the majority of support ticket volume across their customer base of 100,000+ companies.
• Freshworks / Freshdesk research: Estimates 40-50% of incoming support tickets are repetitive, routine questions that could be automated.
• IBM Watson contact center research: Found similar patterns in enterprise contact centers, with simple, predictable inquiries making up the bulk of volume.
• Harvard Business Review (2017): “Kick-Ass Customer Service” research found that the vast majority of support interactions are for routine issues that don't require human judgment.

“LLM response time: 1-5 seconds”

This refers to the full round-trip time for an LLM-powered support resolution, not just the raw API call:

• Raw API latency: Modern LLMs (GPT-4o, Claude Sonnet) return first tokens in 200-500ms, but full responses for support-length answers take 1-3 seconds.
• Smaller models: GPT-4o-mini and Claude Haiku are faster (500ms-1.5s for classification-length tasks), but still 3-10x slower than a purpose-built classifier.
• Full pipeline: A production LLM support system typically adds prompt construction, retrieval (RAG), guardrails, and response parsing on top of the raw API call, pushing total latency to 2-5 seconds.

We use the range “1-5 seconds” to reflect the full spectrum from optimized lightweight setups to production RAG pipelines. The key comparison point is that a dedicated classifier avoids the token generation step entirely.

“LLM cost per resolution: $0.50 – $2.00+”

Important context: This is the total cost of ownership per resolved ticket, not the raw API cost of a single LLM call.

• Raw API cost per call: A single GPT-4o-mini or Claude Haiku classification call costs roughly $0.001-$0.05 depending on prompt length. This is inexpensive.
• Multi-turn resolution: Most support tickets require multiple LLM calls (classify, retrieve context, generate response, handle follow-up). A multi-turn conversation can cost $0.10-$0.50+ in API costs alone.
• Engineering and infrastructure: Building and maintaining a production LLM support pipeline requires prompt engineering, RAG infrastructure, guardrails, hosting, monitoring, and ongoing tuning. These costs are amortized per ticket.
• Industry benchmarks: Enterprise AI support platforms (Intercom Fin, Zendesk AI) typically charge $0.99-$2.00 per AI-resolved conversation, reflecting the true total cost of LLM-powered support.

The $0.50-$2.00+ range reflects what teams actually spend per resolution when they build or buy LLM-powered support. If you're only making a single API call for classification (comparable to what Supp does), the raw cost is much lower — but you won't get confidence scoring, automatic escalation, or the “wrong answers are free” guarantee.

“LLMs hallucinate — no built-in confidence scoring”

LLM hallucination is a well-documented phenomenon in AI research. Our specific claim is:

• Hallucination: LLMs can generate plausible but incorrect answers with high apparent confidence. This is well-established in research from OpenAI, Anthropic, Google, and academic institutions.
• Confidence scoring: Some LLMs provide log probabilities, but these don't directly map to “is this classification correct?” in the way a purpose-built classifier's softmax output does. Building reliable confidence thresholds on LLM logprobs requires significant calibration.
• Our approach: Supp's classifier produces a calibrated confidence score per intent. When confidence is below threshold, we don't guess — we escalate to a human. No charge.

To be fair, LLMs are improving rapidly at factual accuracy, and some providers are adding confidence features. Our comparison reflects the current state of general-purpose LLMs applied to support classification without custom calibration.

“Setup: days to weeks” for DIY LLM support

Building a production-ready LLM support system requires prompt engineering, testing across edge cases, building guardrails to prevent harmful outputs, setting up hosting infrastructure, implementing monitoring, and creating fallback/escalation paths. Based on developer experience reports and industry timelines, this typically takes days for a minimal viable setup and weeks for a production-grade system. Supp's setup is a single script tag because the model, infrastructure, and escalation logic are pre-built.