Decagon
80% of Decagon's own inference now runs on models it fine-tuned in-house, and case studies show real actions, not just deflection — one OpenAI-cited customer resolves 91% of tickets with zero human involvement. But it's a $4.5B valuation against a third-party estimate of $35M ARR, roughly 130x, and reviewers keep hitting the same wall: nobody can tell you why the agent did what it did.
| Platform fee (est.) | ~$50,000/yr | third-party estimate; not vendor-published |
| Usage (est.) | ~$0.99/conversation or ~$0.50/resolution | negotiated per contract; definition of 'resolution' is a documented source of billing disputes |
| Median enterprise ACV (est.) | ~$386,000/yr | range ~$95,000–$590,000+, per Vendr data via aggregator reporting |
No pricing page has ever existed (Wayback Machine CDX check returns zero snapshots). Every figure here is a third-party estimate (Vendr data, aggregator reporting), not vendor-published.
checked 2026-07-30 · vendor pricing page
Element scores
Strengths
Agent Operating Procedures are a real natural-language-plus-code workflow layer with Git versioning, not just a marketing label, and the evaluation setup behind it — offline LLM-judge scoring plus staged online A/B tests with real statistical significance — is more rigorous than most of this category ships. The small-model strategy is paying off: over 80% of Decagon's own inference now runs on models it fine-tuned in-house, and named case studies back the resolution claims (Duolingo 80% deflection, one OpenAI-cited customer at 91% with zero human involvement).
Honest dings
There has never been a public price. The Wayback Machine shows zero snapshots of a pricing page ever existing, and third-party deal data puts the median enterprise contract at roughly $386K/year, ranging from about $95K to $590K+. Despite the Trace View and Watchtower tooling on the tin, G2 reviewers and repeated independent write-ups converge on the same complaint: operators can't always tell why the agent decided something, and audit logs and user roles are described as primitive. Agent Assist, the human-copilot feature, only works with Zendesk. The $4.5B valuation sits against a third-party ARR estimate of roughly $35M, about 130x, a number Decagon itself has not confirmed or denied.
Sources
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- decagon.ai — Homepage — product positioning, customer-logo wall (accessed 2026-07-30)
- decagon.ai/about — Official metrics (10M+ served, 80% deflection, 65% cost decrease), founders, investor list (accessed 2026-07-30)
- decagon.ai/product/aop — Agent Operating Procedures — feature detail, Git versioning (accessed 2026-07-30)
- decagon.ai/product/watchtower — Watchtower always-on QA feature detail (accessed 2026-07-30)
- decagon.ai/product/duet — Duet AI co-pilot feature detail (accessed 2026-07-30)
- decagon.ai/product/testing-qa — Simulations/Testing & QA — synthetic test generation, CI/CD integration (accessed 2026-07-30)
- decagon.ai/product/integrations — CRM/helpdesk connectors, MCP open-connectivity confirmation (accessed 2026-07-30)
- decagon.ai/security — Security architecture — RBAC, SSO, bad-actor/supervisor models, compliance badges (accessed 2026-07-30)
- decagon.ai/product/experiments — A/B testing product — statistical significance, concurrent experiments (accessed 2026-07-30)
- decagon.ai/glossary/what-is-explainable-ai — Decagon's own XAI definition contrasted with its observability-focused approach (accessed 2026-07-30)
- decagon.ai/glossary/what-is-the-a2a-protocol — A2A protocol explainer — educational only, not confirmed shipped (accessed 2026-07-30)
- decagon.ai/blog/introducing-decagon-labs — 80%+ in-house model traffic disclosure (March 2026) (accessed 2026-07-30)
- decagon.ai/blog/fine-tuning-ai-agents — SFT/RL fine-tuning architecture, network-of-specialized-models description (accessed 2026-07-30)
- decagon.ai/blog/evaluation-engine-ai-agents — Two-phase evaluation architecture, multi-provider model use (accessed 2026-07-30)
- decagon.ai/blog/series-d-announcement — Official Series D announcement — 100+ new enterprise customers in FY2025 (accessed 2026-07-30)
- openai.com/index/decagon — OpenAI's own customer case study — model stack, fine-tuned RAG query rewriting, 91% no-human resolution (accessed 2026-07-30)
- bloomberg.com/news/articles/2026-01-28/ai-customer-… — Independent confirmation of $250M Series D at $4.5B valuation (accessed 2026-07-30)
- sacra.com/c/decagon — Third-party ARR estimate (~$35M as of October 2025) (accessed 2026-07-30)
- g2.com/products/decagon/reviews?qs=pros-and-cons — G2 pros/cons — 'primitive' user roles/audit logs, Zendesk-only Agent Assist restriction (accessed 2026-07-30)
- quiq.com/blog/decagon-reviews — Independent critical review — black-box transparency complaint, single-generalist-agent limitation (accessed 2026-07-30)
- yardstickresearch.app/tear-sheet/decagon-ai — Wayback Machine confirmation that decagon.ai/pricing has never existed (accessed 2026-07-30)
- komo.ai/directory/decagon — Vendr-sourced pricing data — platform fee, per-conversation/resolution rates, median ACV (accessed 2026-07-30)