Decagon AI: Complete Guide to Features, Pricing & Limitations (2026)

I know, I know, you're here because someone dropped Decagon into a Slack thread, a vendor shortlist, or a board deck - and now you need to figure out what it actually does, what it costs, and whether it's the right fit.

Most likely, one of these three things happened:

1. Your company is evaluating AI support tools and Decagon showed up on the enterprise shortlist alongside Sierra and Forethought.
2. You saw the $4.5 billion valuation headline and want to understand what's behind it.
3. You're already using another AI support tool and wondering if Decagon is worth the jump.

This guide covers every angle: features, setup, pricing, resolution rates, limitations, and who Decagon is actually built for.

Every claim links to a source. Where the evidence is thin, we say so.

What is Decagon AI?

Decagon is a conversational AI platform that builds and scales AI agents for enterprise customer support.

Their tagline is "The AI concierge for every customer" — the pitch being that every customer interaction gets a personalized, intelligent response without a human in the loop.

The core differentiator is ‘Agent Operating Procedures (AOPs)’ — a proprietary system where non-technical teams define complex support workflows in plain language rather than coded decision trees.

AOPs "combine the flexibility of natural language with the precision of coded logic," according to Decagon's product page.

The product follows three phases:

1. Build (define workflows)
2. Optimize (test and experiment), and
3. Scale (analytics and insights).

Co-founders Jesse Zhang (CEO) and Ashwin Sreenivas (President) started the company in 2023 after surveying businesses about AI pain points at an Andreessen Horowitz retreat.

Zhang previously founded Lowkey (acquired by Niantic in 2021); Sreenivas co-founded Helia (acquired by Scale AI in 2020).

The company emerged from stealth in June 2024 and has since raised ~$481M across five rounds, reaching a $4.5B valuation in January 2026. Named to the 2025 Forbes AI 50 list, it now has 300+ employees in San Francisco, New York, and London.

The customer list reads like a tech who's-who: Notion, Rippling, Duolingo, Chime, Hertz, Eventbrite, Substack, and Riot Games.

How easy is it to set up Decagon AI?

Decagon doesn't have a "sign up and go" flow. The onboarding is a white-glove enterprise engagement.

The company assigns dedicated Agent Product Managers (APMs) and Forward-Deployed Engineers (FDEs) who embed with your team to accelerate deployment.

The documented onboarding process follows a phased approach: pre-kickoff audit of your current workflows, a kickoff meeting to define success metrics, two parallel tracks (content: converting your SOPs into AOPs; technical: CRM access, authentication, API setup), configuration of routing rules and escalation paths, internal testing with simulated conversations, and gradual launch via A/B testing with increasing traffic.

Decagon's own blog describes a ~6-week timeline from initial discovery to full deployment.

Co-founder Ashwin Sreenivas told OpenAI in a case study that "core infrastructure can be up and running in days."

G2 reviewers partially confirm this — one wrote:

"Implementation was very quick (<1 week) and the team is constantly supportive."

The catch: there's no self-serve signup, no public documentation site, and no way to test the product without engaging sales.

For teams that want to get an AI agent running this week, that's a problem. Decagon also offers Decagon University — a self-paced education program with sandbox environments and certifications — but only for existing customers.

What channels does Decagon AI work in?

Decagon supports four primary channels: chat, email, voice, and SMS. These run under a single intelligence layer with cross-channel memory, meaning a conversation started in chat can continue in voice or email without losing context.

Helpdesk integrations include Zendesk (knowledge base, ticketing, escalation via Sunshine), Salesforce (Customer 360, case creation), Intercom (email routing, chat), and Kustomer (knowledge base sync).

Voice is a standout feature. Decagon Voice 2.0 supports inbound and outbound calls with sub-second latency, customisable tone and speed, interruption handling, and branded caller IDs. Voice integrations include Amazon Connect, RingCentral, and SIP trunking. The Spring 2026 release added outbound voice — proactive AI-initiated calls, campaigns, callbacks, and voicemail handling.

Knowledge base integrations cover Confluence, Contentful, Guru, Zendesk Help Center, Slack, and custom databases. E-commerce integrations include Shopify and Stripe for order lookup and payment processing.

What's missing is worth noting. Freshdesk is not listed on Decagon's integrations page — a significant gap for the thousands of companies running on it. And Decagon has no marketplace listings on the Zendesk Marketplace, Intercom App Store, or Salesforce AppExchange — all integrations are direct API connections.

What are the limitations of Decagon AI?

If you're evaluating Decagon against alternatives, start with the limitations below.

Pricing opacity and high cost barrier

Decagon has no public pricing page.

All pricing requires a sales call.

Third-party contract data puts the median annual contract at ~$432,750 with a range of $105,000 to $923,183.

Usage-based pricing means unpredictable costs during seasonal spikes.

Agent Assist was limited to Zendesk

Agent Assist's Zendesk-only access is a recurring complaint in G2 reviews:

"Users are frustrated by access restrictions limiting 'Agent Assist' to Zendesk, affecting broader usability across departments."

Decagon relaunched its copilot as Decagon Assist in August 2026 and now names Salesforce, Zendesk and Front as host systems, so if your team runs on Intercom or another helpdesk, confirm coverage before you buy.

No standalone helpdesk

Decagon requires a separate helpdesk (Zendesk, Salesforce) for human agent handoffs.

It doesn't include a native human inbox. You're maintaining two platforms — and paying for both.

Product immaturity

Multiple G2 reviewers flag this directly:

"Decagon is still a new product, and lacks maturity in some of its features. For instance, regression testing only recently became available, and they are still building out guardrails that are necessary for the long-term quality of our chatbots."

Basic user roles and shallow audit logs

G2 reviewers report rudimentary user roles that make granular permissions difficult. Audit logs "lack depth, which can cause issues when tracing activity or ensuring compliance."

"Black box" AI decisions

Despite marketing transparency, third-party reviewers note:

"You can't always see why the AI did what it did. That makes it tough to review conversations, tweak the agent's behavior, or figure out what went wrong when it makes a mistake."

Engineering resources still needed

While AOPs promise natural-language workflow creation, advanced customisation still requires developers. "Both [Decagon and Sierra] still require code execution for AOPs/SDK, prolonging onboarding."

Limited analytics scope

Decagon's analytics cover the AI layer only — they don't show how human agents and AI work together.

No native workforce management integration exists for connecting AI activity with scheduling or staffing.

G2's own numbers tell a story

Decagon scores 9.7/10 for Quality of Support and 9.4/10 for Customer Interaction Automation. But its Ticket Resolution score is 7.9/10 — notably lower than Forethought's 8.7/10 on the same metric.

High deflection doesn't always mean high resolution.

What knowledge sources can I train Decagon AI on?

Decagon pulls knowledge from multiple structured and unstructured sources through what it calls a Unified Knowledge Graph — a system that processes all sources, infers relationships, and identifies key entities.

Supported sources include:

• help centre articles (synced from Zendesk, Kustomer, Guru, Confluence, Contentful), product documentation
• past customer conversations and historical transcripts
• CRM data (Salesforce Customer 360)
• internal business APIs (real-time order data, user accounts)
• internal knowledge bases (Slack, custom databases), and
• Shopify/Stripe product and payment data.

Custom API integrations are also supported for proprietary data.

AOPs allow teams to define how knowledge is retrieved and applied at the workflow level.

The Knowledge Suggestions feature analyses conversations where customers didn't get complete answers and automatically drafts new knowledge base articles to fill gaps.

A single knowledge source serves all languages — policy updates written in English become effective globally.

The notable gap: specific supported file types (PDF, CSV, etc.) are not publicly documented.

There's no public documentation site, and knowledge source configuration appears to happen during white-glove onboarding rather than through self-serve upload.

You can't just drag and drop a PDF into a dashboard.

What features does Decagon AI have?

Decagon's feature set is extensive and oriented toward enterprise-scale automation. Here's what it breaks down to in practice.

Agent Operating Procedures (AOPs)

AOPs are the core workflow engine. Teams define support workflows in natural language that "compile into code."

The AOP Copilot (launched September 2025) converts rough ideas or existing SOPs into production-ready AOPs in seconds. AOP Templates provide pre-built blueprints for common use cases like refund processing or account verification.

AI Actions

AI Actions let the agent do things, not just say things. Through integrations with Stripe, Shopify, and Salesforce, Decagon can process refunds, update orders, verify identity, and create tickets — without escalating to a human.

Testing and QA

Testing and QA goes deeper than most competitors. End-to-end simulated conversations using AI-generated mock customer personas. Unit testing for individual workflow components. Regression testing with historical transcripts. And Experiments — A/B testing that routes live traffic across agent versions and measures CSAT, deflection, and other KPIs.

Watchtower

Watchtower is an always-on QA system that reviews every conversation (AI and human) against custom scoring rubrics. It monitors sentiment, detects fraud mentions and regulated complaints, and alerts for compliance risks. Trace View provides transparent observability — trace every AI decision, see which model was called, which workflow triggered, and which knowledge article was referenced.

Reporting

Reporting includes built-in analytics with theme identification and anomaly detection. Ask AI allows natural-language querying of conversation data ("Why are refunds increasing this week?"). Knowledge Suggestions identify gaps and auto-draft new articles.

Agent Assist

Agent Assist automates mundane tasks for human agents, drafts response suggestions, and learns from expert interactions. Since August 2026 it runs as Decagon Assist inside Salesforce, Zendesk and Front, though older G2 reviews describe it as Zendesk-only.

Newest features (Spring 2026)

The newest features include Proactive Agents combining user memory and outbound voice, an Agent Workbench for autonomous debugging, and Duet — an AI partner that manages the full agent lifecycle through conversation.

How do I improve Decagon AI responses?

Decagon's optimisation loop centres on iterative AOP refinement, testing, and data-driven feedback.

Teams continuously update Agent Operating Procedures based on conversation analytics. The AOP Copilot suggests improvements and converts rough notes into production-ready instructions.

Agent Versioning applies CI/CD discipline — roll back to previous versions if new changes degrade performance.

Before deploying changes, the Simulations feature lets teams test core workflows with AI-generated mock personas. Unit testing validates individual workflow components. Regression testing replays historical transcripts against new agent versions to catch regressions before they hit customers.

The Experiments feature enables live A/B testing — route a percentage of traffic to a new agent version, measure CSAT and resolution rates, then scale up once the numbers confirm improvement.

Decagon recommends a gradual rollout approach: start with a small traffic percentage, use Watchtower to monitor every conversation against custom quality criteria, and leverage Knowledge Suggestions to identify where customers aren't getting complete answers.

The system analyses these gaps and auto-drafts new articles based on how expert human agents resolved similar issues.

The data flywheel means every conversation improves future performance. Ask AI lets teams query conversation data to surface emerging trends or recurring failure patterns ("Show me all conversations about refund policy confusion this week").

What resolution rate can I expect from Decagon AI?

Decagon's homepage headlines an 80% deflection rate and a 93% agent quality score. Individual results vary significantly.

Published case study metrics (all from Decagon's own marketing):

Deflection vs. resolution matters here.

Decagon's own blog acknowledges the tension: "High deflection rates can hide customer rage and brand damage."

Deflection means the customer didn't reach a human. Resolution means the issue was actually solved. These are different things.

Decagon has moved toward per-resolution pricing where customers only pay for fully resolved conversations.

But the definition of "resolved" remains somewhat ambiguous — Decagon determines resolution algorithmically, this can lead to billing disputes.

Independent corroboration is limited. G2 reviewers mention 75-80% ticket deflection at launch.

One reviewer reported cost-per-ticket dropping from $78.43 to $73.92 in a partial first quarter. The most impressive metrics (90%+ resolution, 95% cost reduction) come exclusively from Decagon's own case studies.

Decagon's G2 Ticket Resolution score of 7.9/10 — the lowest of its category scores — suggests real-world resolution quality may not match marketing claims.

What AI model does Decagon AI use?

Decagon doesn't rely on a single LLM. The platform uses a model-agnostic, multi-model architecture with foundation models from OpenAI (GPT-3.5 fine-tuned, GPT-4), Anthropic (Claude), and Cohere, layered with company-specific training data.

Rather than routing all queries through one model, Decagon uses an "ecosystem of agents" that work together and review each other's work.

A supervisor model detects hallucinations before responses go out, automatically revising anything that strays from factual grounding. Different models handle different tasks — fine-tuned GPT-3.5 for RAG query rewriting, GPT-4 for complex decision-making.

Decagon invests heavily in fine-tuning open-source models, especially for voice. Their approach uses Supervised Fine-Tuning with curriculum learning, Reinforcement Learning for behaviour refinement, and mixed-objective fine-tuning.

These models reportedly surpass larger frontier models on targeted tasks while being smaller and faster. For voice, Decagon achieves p95 model latency under 400ms and a ~6x cost reduction per turn versus closed models. They ship model updates weekly, sometimes daily.

Production inference runs on Together AI (NVIDIA HGX B200 GPUs) with speculative decoding and prompt caching. Training uses Modal for SFT and RL workloads. Voice synthesis uses ElevenLabs. Cloud hosting is on Google Cloud.

On data handling: Decagon enforces zero-day data retention with all AI providers — no conversation data is stored by OpenAI, Anthropic, or others. PII is automatically redacted using Google's DLP service. No customer data is used for model training.

What languages does Decagon AI work in?

Decagon claims to support "any language" natively. The platform automatically detects the customer's language and responds accordingly.

Cross-language knowledge retrieval is a strength: a single English-language knowledge base serves all languages.

Policy updates written in English become effective globally without maintaining separate knowledge bases per language. The Rituals Cosmetics case study specifically confirms support across 15 languages in a single deployment.

Performance quality will vary by language given the underlying LLMs' training data distribution. Decagon's primary markets appear to be English-speaking countries, with notable deployments in Brazil (NG.CASH), Latin America (Mercado Libre), and Europe (Rituals Cosmetics, Deutsche Telekom). No specific language quality benchmarks are published.

How secure is Decagon AI?

Decagon positions security as foundational, with enterprise-grade protections across authentication, encryption, AI safety, and compliance.

Compliance: SOC 2 Type II certification (confirmed in blog content and the trust centre at trust.decagon.ai). GDPR compliant. HIPAA options for healthcare clients, including Business Associate Agreements.

Encryption: AES-256 at rest, TLS 1.2+ in transit. Centrally managed keys with strict access controls, expiration policies, and rotation schedules.

Access controls: Role-Based Access Control (RBAC), SSO via Okta and Microsoft Entra (SAML/SCIM), just-in-time JWT API tokens scoped for minimal privilege. Voice authentication for end-user verification. That said, G2 reviewers note that user roles are still basic and lack granularity.

AI-specific safeguards: Bad actor detection for adversarial prompts. Supervisor model for hallucination detection. PII auto-redaction using Google's DLP service. Zero-day retention with all LLM providers.

Infrastructure: Google Cloud hosting with Cloudflare + Google VPC, web application firewalls, regular vulnerability scanning, multi-region infrastructure, auto-failover, and uptime SLAs.

Who is using Decagon AI?

Decagon serves 100+ enterprise customers added in 2025 alone, claiming to have served 10M+ end customers.

Three case studies in particular standout:

Chime (fintech, 7M+ customers) achieved 70% AI resolution across chat and voice, cut support costs by 60%, reported a 30 sNPS, and processes AI-verified identity checks. The financial services vertical is a sweet spot — Decagon also serves Affirm, Block, Bilt Rewards, Varo Bank, and Wealthsimple.

Substack reached 90% resolution without human intervention. As a content platform with largely repetitive support queries (account access, billing, subscription management), this is the kind of use case where AI support excels.

Bilt Rewards reported a $1.75M cost reduction and 75% resolution rate. The company handles credit card rewards and rental payments — a mix of high-volume, moderate-complexity queries.

The customer base spans financial services (Chime, Affirm, Block), technology (Notion, Rippling, Eventbrite, Figma), travel (Hertz, Avis Budget Group), health (Oura, Noom, ClassPass), retail (Mercado Libre, Quince), media (Riot Games, Duolingo), and food delivery (Gopuff, Grubhub). Strategic partnerships include TaskUs (which plans to use Decagon to cut support costs 25-50%) and Deutsche Telekom (commercial pilot + strategic investment).

The common thread: these are all mid-market to enterprise companies with substantial support volumes. G2 reviewers confirm features are "more tailored to mid-market businesses."

How much does Decagon AI cost?

Decagon does not publish any pricing. Everything requires a sales call.

Pricing models

According to Decagon's own pricing guide, two usage-based models are available:

Per-conversation pricing charges a fixed rate for every incoming conversation, regardless of whether it's resolved or escalated. Volume discounts apply. Decagon says this is what "the majority of our customers gravitate towards."

Per-resolution pricing charges a higher fixed rate per fully resolved conversation, with no charge for escalations. Larger commitment volumes lower the rate.

Play video

There are no per-agent or per-seat fees. The model is entirely usage-based.

What do contracts actually look like?

Third-party data from Vendr negotiation data indicates:

Worked example

Let's say you handle 10,000 support conversations per month.

Decagon publishes no per-conversation rate, so the closest real number is Vendr's lowest reported contract of $105,000 a year. The same volume on My AskAI would cost about $15,600 a year on the Scale plan ($1,299 a month), less than a sixth of that floor.

For context: Intercom Fin charges a transparent $0.99 per outcome.

Decagon's pricing is significantly less transparent and reportedly 50-100x more expensive on an annual contract basis than SMB-focused alternatives.

The platform fee covers all channels, integrations, AOPs, Watchtower, testing/QA tools, and analytics. Dedicated APMs and FDEs appear to be included. Whether voice calls carry additional per-minute charges beyond conversation pricing is unclear.

Does Decagon AI have a free trial?

No.

Decagon does not offer a free trial or a free plan.

The only path in is to request a demo through their sales team.

The entire go-to-market motion is enterprise-sales-driven: schedule a demo, go through discovery, receive a custom quote, sign an annual contract.

There is no self-serve signup, no credit card trial, and no sandbox access without engaging sales.

Decagon University provides educational content and sandbox environments, but only for existing paying customers.

Is Decagon AI worth it?

Decagon delivers genuine enterprise value for large organisations with high support volumes. The case study metrics are impressive — 80%+ deflection, 65-95% cost reductions, and the enterprise customer list speaks for itself. The voice capabilities are among the best in the market.

When Decagon IS worth it:

• You handle 50,000+ support conversations a year
• You run on Zendesk or Salesforce
• Your annual budget for AI support tooling exceeds $300K
• You need multi-channel support including voice with sub-second latency
• You want a white-glove onboarding experience with dedicated engineering support
• You operate in a regulated industry that benefits from compliance-focused features

When Decagon is NOT worth it:

• You're an SMB or mid-market company (you're priced out)
• You use Freshdesk, HubSpot, or another non-Zendesk/Salesforce helpdesk
• You want self-serve setup and transparent pricing
• You need to be live this week, not in 6 weeks

What are the Pros and Cons of Decagon AI?

Pros

• Enterprise-grade AI architecture: Multi-model, multi-agent system with hallucination detection, fine-tuned voice models, and a supervisor layer that catches errors before customers see them. The technical depth here is real.
• Comprehensive voice support: Sub-second latency, inbound and outbound calls, branded caller IDs, and SIP trunking integration. Few competitors match this for AI-powered phone support.
• White-glove onboarding: Dedicated APMs and FDEs embed with your team. If you want hands-on help deploying AI support, Decagon treats it like a consulting engagement.

Cons

• Opaque, enterprise-only pricing: No pricing page, no self-serve plans, median contracts of ~$433K/year. Inaccessible to most businesses.
• Narrow agent copilot: Decagon Assist names only Salesforce, Zendesk and Front as host systems, so human agents on other helpdesks may be left out.
• No self-serve anything: No trial, no signup, no public docs. You can't evaluate the product without going through sales.

FAQs