Decagon

Enterprise AI customer-service agent platform — $4.5B valuation, ~$35M ARR by early 2026, AI-only contrast to Crescendo's HITL and Sierra's outcome-pricing thesis.

1. Core Product / Service

Decagon sells an enterprise-grade AI customer support agent that companies configure with an "Agent Operating Procedure" (AOP) — a plain-language workflow definition rather than a dialog tree [1][5]. Three architectural layers:

Agent runtime — frontier-model-backed conversational agent that resolves customer inquiries autonomously across channels (chat, email, voice as of 2025-26).
AOPs (Agent Operating Procedures) — plain-language workflow specs (think "company SOP for support") that non-technical CX leaders can edit. The AOP is the durable artifact: it's tunable, auditable, and replaces traditional dialog trees and intent classifiers.
Concierge layer (2026 pivot) — Decagon's Series D announcement reframed the product from "support deflection" to "concierge experience" — proactive, personalized, multi-modal CX [2].

Materially differentiates from crescendo (which keeps humans in the loop by design) and from sierra (which leans into outcome pricing as the wedge).

2. Target Users & Pain Points

Buyer: VP/Director of Customer Support / CX at high-growth tech and large enterprise (F100). Notion, Duolingo, Rippling, Bilt, Eventbrite, Substack, Oura Health, Affirm, Chime are reference customers [5][6].
Pain points: Pure-AI deflection bots (Ada, Forethought-class) plateau at ~70-80% resolution and frustrate customers on edge cases. Decagon's AOP architecture lets CX teams iteratively tune behavior in plain language, pushing resolution rates higher without engineering tickets.
Wedge: Non-technical configurability + enterprise-grade observability + frontier-model quality.

3. Competitive Landscape

Vendor	Model	Pricing	Wedge	Vs Decagon
Decagon	AI agent + AOP plain-language config	Per-conversation OR per-resolution	Configurable AOPs, $35M ARR / $4.5B val	High-growth tech + F100
crescendo	AI + human experts (HITL), fully managed	Per-resolution outcome-based	Owns the human ops layer	Different bet: HITL by design
sierra	AI agents, outcome-priced (rewarded for resolution)	Outcome-based	Bret Taylor brand, $15.8B val, $150M ARR	Outcome-pricing wedge, blue-chip enterprise
Ada	Self-serve automation, no-code	Performance-based, low-6-figure start	Mid-large digital brands	Older platform, less LLM-native
Forethought	Generative agent assist	Custom enterprise	Triage + agent assist	Augments humans, doesn't replace
Zendesk AI / Salesforce Einstein	Bolted onto helpdesk incumbents	Bundled with seats	Distribution moat	Distribution > AI quality
Cresta	AI for human agents (assist)	Custom	Real-time agent coaching	Different problem space (assist)

4. Unique Observations

Pricing model: Hybrid usage-based with two metering options [6][7]:
- Per-conversation: ~$0.99/conversation list reference, volume discounts kick in.
- Per-resolution: higher per-unit rate, only billed on autonomous AI completion.
- Most buyers choose per-conversation (predictable forecasting; avoids "what counts as resolution" disputes).
- List pricing is private — sales-led. Marketplace data: median $400K/yr ACV, range $100K-$580K, redline floor ~$50K [7].
Implied $/1M tokens consumed: A typical resolved conversation has 5-10 turns × 200-500 tokens with retrieval = ~5-15K tokens. At ~$0.99/conversation, retail rate ~$66-$200 per 1M tokens. Frontier API cost (2026): ~$2-$5/1M for the chat side. Markup: ~15-100×. Most heavily-used Decagon customers process 100K+ conversations/month, so customer math is "0.99 × 100K = $99K/month" vs. legacy BPO's $30-$50/hr-per-agent — Decagon wins decisively.
Moat type: Workflow (AOP framework + audit/observability tooling) + early-mover brand among AI-forward enterprise customers. NOT a model moat (frontier-model abstracted). Modest data moat from cumulative AOP refinements per customer.
Customer profile: Enterprise / F100 + high-growth tech (especially fintech, marketplaces, consumer subscription). Explicitly not SMB. Median deal $400K means even a 100-customer book = $40M ARR (roughly aligned with reported $35M).
Markup multiple over raw API: ~15-100× depending on conversation length and underlying model used. The buyer-side comparable is BPO labor cost, not API cost — so the markup is structurally defensible while AI-vs-human-agent competition lasts.
Series D pivot: The Jan 2026 round (3× valuation jump) was framed as "concierge experience" — proactive, multi-modal CX that goes beyond reactive deflection. Strategic insight: pure deflection is commoditizing fast; the next wedge is outbound AI engagement (proactive notifications, personalized journeys), where the analog is martech automation, not legacy BPO.
Vs Sierra: Sierra's market cap is 3.5× Decagon's despite ~4× the ARR — Bret Taylor brand premium plus deeper F50 penetration. Decagon's bull case is faster product velocity in mid-enterprise + AOP framework as a moat against Sierra's scale.

5. Financials / Funding

Total raised: $481M cumulative [5][6].
Series D (Jan 2026): $250M led by Coatue + Index Ventures at $4.5B valuation.
Series C (Jun 2025): $131M at $1.5B valuation, led by a16z + Accel — 3× valuation step in 7 months.
Tender offer (Mar 2026): completed first tender at $4.5B valuation, providing employee/early-investor liquidity [4].
ARR: ~$35M (early 2026 estimate per Sacra analysis); >100 new enterprise logos signed in 2025.
Headcount: not disclosed precisely; lean for ARR — suggesting strong unit economics.

6. People & Relationships

Co-founders:
- Jesse Zhang (CEO) — ex-quant trader; the GTM / product anchor. Stanford CS.
- Ashwin Sreenivas — ex-Helia; the engineering anchor.
Investors: a16z (early lead), Accel, Bain Capital Ventures, ChemistryVC, Coatue (Series D co-lead), Definition Capital, Elad Gil, Forerunner, Index Ventures (Series D co-lead), Ribbit Capital, Starwood Capital, T.Capital, Avra, A* Capital.
Reference customers: Notion, Duolingo, Rippling, Bilt, Eventbrite, Substack, Oura Health, Affirm, Chime.
Related wiki: crescendo (HITL CX, $500M val), sierra (outcome-priced CX, $15.8B val), magic (HITL pattern).