What is Decagon?

Decagon is a customer-experience AI company that builds, deploys and operates autonomous AI agents for enterprise support. It calls its category the “AI concierge”: agents that handle the full interaction across chat, email, voice and SMS, take actions, escalate when needed and improve by learning from past conversations. It sits on top of your existing stack rather than replacing your helpdesk, connecting via direct API. The pitch, in CEO Jesse Zhang’s framing, is human-like agents that understand and anticipate needs rather than acting as a glorified rules engine.

The company was founded in August 2023 by Jesse Zhang (CEO; Harvard CS, previously founded Lowkey, acquired by Niantic) and Ashwin Sreenivas (President; Stanford CS, previously co-founded Helia, acquired by Scale AI), who met at an Andreessen Horowitz retreat. It emerged from stealth in June 2024 and made the 2025 Forbes AI 50.

The founders and the funding

Decagon has scaled fast on the back of feverish investor demand. It has raised roughly $481M across five rounds, most recently a $250M Series D in January 2026 led by Coatue and Index Ventures at a $4.5 billion valuation — a mark a March 2026 employee tender offer then matched. ARR reached around $35M by late 2025 (up from $10M a year earlier), with 100+ enterprise customers added in 2025 across airlines, banking, telecom and retail. Named logos include Notion, Duolingo, Chime, Rippling, Oura, Substack, Bilt, Eventbrite, Affirm, Avis Budget Group, Mercado Libre and Deutsche Telekom (a commercial pilot backed by a strategic investment from Deutsche Telekom’s venture arm). It now runs 300+ employees across San Francisco, New York and London. That trajectory has turned Decagon from interesting startup into a genuine category contender.

How Decagon works — Agent Operating Procedures

The heart of Decagon is Agent Operating Procedures (AOPs): a system for defining support workflows in natural language that, in Decagon’s words, “combine the flexibility of natural language with the precision of coded logic.” A CX team writes the procedure in plain English; it compiles into structured logic the agent executes reliably; and technical teams retain control over integrations, guardrails and versioning through Git. The AOP Copilot (and the newer Duet builder) turn rough notes or existing SOPs into production-ready AOPs in seconds, with templates for common cases like refunds or account verification. This is the cleanest answer in the category to the perennial tension between “easy for business teams to change” and “safe and controllable for engineering,” and it’s Decagon’s strongest differentiator.

Increasingly, the intelligence is Decagon’s own: since March 2026 it reports that around 80% of traffic runs on in-house models trained specifically on customer-support conversations rather than general-purpose third-party LLMs, which it says yields better support performance (and reduces, though doesn’t eliminate, model-provider dependency).

Key features

AI Actions — agents that do, not just say

Through integrations with Stripe, Shopify and Salesforce, Decagon’s agents take real actions: processing refunds, updating orders, verifying identity, disputing transactions, cancelling subscriptions, replacing credit cards and creating tickets — without escalating to a human. This action-taking, governed by AOPs, is what pushes its reported average deflection above 80%.

Voice 2.0 and proactive outbound

Decagon Voice, built in partnership with ElevenLabs, runs the same AOP logic across phone as chat and email. Voice 2.0 (September 2025) cut latency by around 65% to sub-second, with interruption handling, customisable tone, branded caller IDs and cross-channel memory. The Spring 2026 release added outbound voice — proactive AI-initiated calls, campaigns, callbacks and voicemail handling — and Proactive Agents that combine user memory with outbound calling to turn support into concierge-style engagement. Voice integrates with Amazon Connect, RingCentral and SIP trunking.

QA, debugging and testing tooling

Operators get a deep toolset: Watchtower for QA monitoring (resolution and fallback rates, retraining needs), Trace View to see how an AOP executed for a given interaction, Agent Workbench (Spring 2026) for autonomous debugging and root-cause analysis, Simulations to test agent behaviour before going live, plus A/B testing and Decagon University for training. It’s a mature operating layer for running agents at scale.

Scorecard

Overall score: 7.5 / 10 — the average of the eight categories above.

Pricing

Decagon publishes no pricing and offers no self-serve signup, trial or public docs — you evaluate through a sales process. It uses two models, and the figures below come from third-party procurement data; confirm directly with Decagon.

The honest read: Decagon’s pricing is flexible in model but opaque in practice. Offering both per-conversation (predictable, volume-discounted) and per-resolution (pay-for-success) is sensible, and lets you match billing to your support profile. But with no published rates, no trial and no self-serve, you commit to an enterprise contract and a multi-week implementation before you can independently model ROI — the same accessibility wall as Sierra and Ada. As with any per-resolution element, pin down exactly how “resolution” is defined and measured, whether routing interactions are billed, and what happens on repeat contacts, because those definitions move the invoice more than the headline rate does.

Pros & cons

Honest weaknesses

The most practical issues are integrations and accessibility. Decagon connects by direct API and has no listings on the Zendesk Marketplace, Intercom App Store or Salesforce AppExchange, and — notably — Freshdesk isn’t on its integrations page at all, a real gap for the many teams running on it. Its agent-assist copilot for human reps is restricted to Zendesk, so everyone else loses that half of the value. And like its peers, Decagon offers no trial, no self-serve and no public pricing or docs, so evaluation means entering a sales process and committing to an enterprise contract plus a multi-week-to-multi-month rollout before you can prove it on your own data.

The subtler weakness is transparency. Decagon built Trace View, Watchtower and the newer Agent Workbench precisely because early customers struggled to understand why an agent made a given decision — and while the company says decision-making is fully traceable with audit logs, user feedback suggests the day-to-day auditing experience is still inconsistent and depends heavily on how deeply your team adopts the tooling. Its analytics also stay largely within the AI layer, offering limited visibility into how AI and human agents coordinate. None of this undermines what is a genuinely strong, well-differentiated product — but the right test, as ever, is how it handles your hardest 20% of tickets in a pilot, which the no-trial model makes harder to run cheaply.

Who is Decagon for?

Decagon is the right pick for fast-scaling, data-rich enterprises that want a buildable, action-taking AI agent and the control to iterate it in-house — high-volume tech, consumer, fintech, travel and telecom teams that value AOPs’ blend of plain-language editing and engineering governance, and that have the volume and budget for an enterprise deployment. If your CX team wants to own how agents behave without waiting on a vendor for every change, AOPs are a standout. It’s the wrong tool for smaller teams, anyone who needs transparent pricing or a free trial, teams on Freshdesk or relying on marketplace-style plug-and-play integrations, or those who need agent-assist on a non-Zendesk helpdesk. For the category’s momentum leader, especially on voice, see Sierra; for the strongest agent inside a modern helpdesk with published pricing and a trial, Intercom Fin; and for a comparable managed enterprise agent, Ada.

FAQ

AI customer supportDecagonconversational AIAI agententerprise AIAI tool review