Skip to content
NagentNagent
Log inSign upHire your AI team
Atlas/Decagon
Vendor dossier / Customer experience

Decagon

Conversational AI agent platform for complex customer-support workflows across channels.

REVIEWED 29 SEPTEMBER 2026CXNO HANDS-ON BENCHMARK

What it is

Official-source notesCustomer experiencescaling

Decagon uses Agent Operating Procedures to specify natural-language workflows. It describes testing, observability and experimentation across chat, voice and email. [V1] [V2]

Documented means the statement is present in the cited material. Capability effectiveness, customer outcomes and security assurances have not been independently audited here.

Promoted use cases

  • Resolve complex support workflows
  • Operate a customer concierge across channels
  • Refine procedures using conversation analytics

CX agents operate where answer quality and action correctness directly affect customers. A response that sounds useful can still be wrong, and a conversation marked resolved can reopen. The comparison should combine customer experience, transactional accuracy, human handoff, operational visibility and the exact commercial definition of an outcome.

Pricing and buying model

Published commercial evidence

Sales-led quote

No reliable public numeric rate verified.

Check current commercial source ↗

Budget for implementation, model and tool usage, data, human review and ongoing support. Credit units and outcome definitions differ between vendors. A missing numeric rate is marked unverified rather than replaced with an old third-party estimate.

Collaboration and governance

AOP editing is relevant to business ownership. Inspect who can change procedures, how versions are evaluated and how releases are rolled back.

The relevant unit of control is permission to change an order, account or service record. Inspect who can propose, approve, execute, interrupt and audit that action. Shared seats, shared content and a shared live agent session are different capabilities; require a demonstration of the one your process needs.

Evaluate memory correction, permission revocation and release control for updated instructions. An improvement loop should preserve the original evidence, proposed change, test results and named approver. A safety or security badge alone cannot establish those workflow properties.

Review and customer evidence

No independent review sample was verified in this research pass. Vendor-hosted testimonials are selection-biased customer evidence, not an aggregate rating.

Before relying on a testimonial, confirm the exact product, version, package, workload and baseline. Ask a relevant customer about setup effort, failed cases, ongoing manual work and support after launch. This dossier does not convert customer logos or vendor-hosted awards into independent proof.

Focused feature documentation

Focused documentation check · Sep 2026

AOPs and approval-gated self-improvement

Decagon’s Agent Operating Procedures let CX teams write inbound and outbound workflows in natural language; its product page says engineering teams can use Git-based version tracking and inspect reasoning, action timing and latency. The page also promotes guardrails across chat, voice and email. These are specific product claims rather than independent tests of how safely a real refund or account update runs.

Its June 2026 Duet Autopilot announcement describes a self-improvement loop: production conversations surface a problem, the system proposes an AOP change, tests it against the originating conversation and a curated golden set, and updates the test set as new issues appear. Every proposed change requires human approval before production, according to Decagon; reviewers see a versioned diff, issues and validation results. That is directly relevant to Nagent’s governance thesis. Verify whether the proposal, test set and approval reflect all channels and whether a reviewer can reject a regression without affecting live traffic.

The reviewed public material did not give a reliable universal numeric rate. Request a scoped quote by channel, conversation volume, integration and human support. Pilot the same difficult cases across the old and proposed AOP, including repeated messages, revoked permissions and a sensitive escalation. Measure independently accepted resolutions and correction effort, and distinguish Decagon’s own benchmark and customer-story results from a controlled buyer comparison.

Primary sources: Official AOP product page ↗ · Official Duet Autopilot announcement ↗

Published documentation and public illustrations; account behaviour was not independently tested.

Visual source

Public Decagon source page captured 23 September 2026
Decagon · Public page captured 23 Sept 2026. Vendor presentation, not an authenticated product test. Open source ↗

Fit, limitations and proof requests

Editorial assessment

Decagon should be evaluated on complex resolution and operational control, not just conversational fluency. A custom Nagent comparison needs equivalent channels and transaction permissions.

Priority question: Can we replay the same difficult cases against an old and new AOP before promoting the change?

The evaluation owner should be the customer-experience and support-operations leads. Use durable, accurate resolutions with appropriate human escalation as the business target. Judge the solution in its own stack role: a cloud runtime, a specialist production tool and a managed AI team can be complementary purchases.

Use-case evaluation plan

Knowledge answer with a verifiable source

Provide a current policy, an obsolete policy and an ambiguous customer question. Require the agent to use the right version, identify uncertainty and avoid inventing exceptions. Ask a human reviewer to verify the answer without reading a long hidden history.

Measure: answer accuracy, source freshness and appropriate escalation.

A bounded customer action

Use a test customer, a low-risk transaction and an explicit action limit. Try a request outside that limit and a case with conflicting identity information. Verify the backend state after execution; a conversational confirmation alone is not evidence that the transaction succeeded.

Measure: correct state changes, blocked unauthorized actions and recoverability.

Learning from a failed resolution

Introduce a repeat contact after an apparently successful interaction. Ask the system to identify the failure, propose a policy or knowledge correction and show the evidence. Require a human owner to approve the change before it affects other customers, then replay similar cases.

Measure: reopen rate, correction quality, regression results and human workload.

Detailed comparison

Read the detailed Nagent vs Decagon guide for operating models, shared context, governance, cost, evidence gaps and a staged pilot.

Nagent vs Decagon

Sources and evidence register

Nagent · Multiplayer AI for growth teamsResearch: 21–29 September 2026 · Public-source analysis, no hands-on benchmark · Sources & method