Decagon vs Lorikeet (2026)

Connectors span CRM, helpdesk and ticketing (Salesforce, Intercom, Zendesk), knowledge bases (Confluence, Contentful, Kustomer), and CPaaS and telephony (Amazon Connect, RingCentral, SIP trunking), plus custom endpoints and MCP connectivity for any other system. Agents both retrieve data and take action.

Sourcedecagon.ai/product/integrationsread 2026-09-05

Connectors cover ticketing and messaging (Zendesk, Zendesk Messaging, Front, Help Scout, Slack), CRM (Salesforce, including running workflows inside it) and revenue tooling. Tools are set up as APIs or webhooks, so the agent reads data and takes actions in any internal system, scoped least-privilege with short-lived single-operation tokens and HMAC-signed inbound webhooks.

Sourcelorikeetcx.ai/integrationsread 2026-09-05

Agent Operating Procedures turn natural language inbound and outbound workflows into executable agent logic that pulls live data and triggers actions across connected systems. Versions are tracked in Git, and rollout is staged across agent versions.

Sourcedecagon.ai/product/aopread 2026-09-05

The agent runs multi-step work by following the customer's own standard operating procedures without improvising, whether a three-step form or a forty-step refund flow. Workflow steps include workflow matching, ticket triage, gather, get status, choice and route based on problem, with tool calls such as retrieving recent transactions shown inline. Lorikeet describes specialized agents working together in real time across channels to deliver end-to-end resolutions, and the login surface is a workflows console.

Sourcelorikeetcx.airead 2026-09-05

Chat, voice and email run on one intelligence layer. Inbound calls route over SIP and CPaaS platforms, email comes in through Zendesk and Intercom, live chat escalates through Zendesk Sunshine and Salesforce, and outbound and proactive agents are supported as a core feature.

Sourcedecagon.ai/product/overviewread 2026-09-05

Five channels run on one engine, phone, SMS, chat, email and WhatsApp, with an SDK as a further delivery surface, and specialized agents work together in real time across chat, email, voice and SMS to deliver end-to-end resolutions. Voice runs at sub-second latency on the same engine, so a customer who starts in chat does not repeat themselves on a call, and every channel inherits the same guardrails and logging.

Sourcelorikeetcx.airead 2026-09-05

Decagon maintains the agent's own knowledge base, filled by sync from customer systems including Confluence, Contentful and Kustomer. Suggestions keeps it current by finding gaps from conversations, ranking their impact on coverage and issuing monthly content updates. Traces name the knowledge articles the agent referenced. The knowledge base persists between conversations.

Sourcedecagon.ai/product/suggestionsread 2026-09-05

The agent is grounded in the customer's own documented procedures, not retrieval alone. Lorikeet follows the customer's SOPs, taps into its systems and handles its highest-stakes workflows, and says its architecture goes beyond retrieval-augmented generation by following the procedures themselves, so the agent neither just retrieves nor invents process. Help centers, reference materials and internal tools supply the surrounding material, and the customer writes and owns the procedures.

Sourcelorikeetcx.airead 2026-09-05

Memory is a named part of the agent engine. It holds customer context across sessions and channels alongside the customer's systems of record, with stated data portability and control, so a conversation started in chat continues by voice without repetition. Decagon does not publish how long memory is kept or how to delete it for one customer.

Sourcedecagon.ai/product/overviewread 2026-09-05

Customer memory keeps context across every touchpoint, so a customer who starts in chat does not repeat themselves on a call, because the channels run on one engine and persistence spans channels. No scope, lifetime, retention window or deletion control for that memory is described.

Sourcelorikeetcx.airead 2026-09-05

Guardrails can be configured for brand voice, escalation and hallucination. A supervisor model revises ungrounded responses before they are sent, bad actor detection escalates to a human, and Watchtower flags conversations for triage against criteria the customer sets. Oversight is automated gating and quality review after the fact, with no customer facing approval queue that holds an agent action until someone signs off.

Sourcedecagon.ai/product/aopread 2026-09-05

Runtime guardrails detect sensitive topics and escalate to the customer's team before the agent proceeds, and outbound checks block the agent from disclosing withheld fields or acting outside policy, independently of the model. Tools are scoped least-privilege, dangerous actions need per-user elevation, and the integration is red-teamed against prompt injection and scope escalation before go-live. People write the procedures the agent follows, and handoffs arrive with the investigation complete.

Sourcelorikeetcx.ai/product/trustread 2026-09-05

Access controls include role based access control, SSO with Okta and Microsoft Entra, tamper protected audit logs and short lived scoped JWTs for agent access. Data is encrypted with AES-256 at rest and TLS 1.2+ in transit, DLP based redaction removes PII, and model providers are held to zero-day retention. Decagon shows SOC 2, ISO 27001, PCI DSS, HIPAA, CCPA and Data Privacy Framework marks and runs a live trust portal.

SourceDecagon security page and trust.decagon.ai; decagon.ai/securityread 2026-09-05

Lorikeet holds SOC 2 Type II, ISO 27001:2022, HIPAA and GDPR, all independently verified and published on a public Vanta trust center at trust.lorikeetcx.ai with reports downloadable under NDA and refreshed annually, and it signs BAAs for healthcare customers. Controls include automatic PII redaction, minimum-necessary PHI handling, TLS 1.3 in transit and AES-256 at rest, no public internet access to production, row-level-security tenant isolation, SSO with hardware-key MFA, and third-party penetration testing.

Sourcelorikeetcx.ai/product/trustread 2026-09-05

Traces explain how and why the agent decided as it did at any point in a conversation, covering model calls, workflow triggers, the knowledge articles referenced, and runtime and latency figures. Watchtower reviews every conversation against criteria the customer defines and flags outliers for triage.

Sourcedecagon.ai/product/overviewread 2026-09-05

Every customer interaction, model choice and action is tracked and fully explainable, not a black box. Every tool call, elevation request and reasoning step is logged with timestamps and can be replayed for a regulator examination, and Coach scores the quality of every conversation, human or AI.

Sourcelorikeetcx.ai/product/trustread 2026-09-05

The platform runs on multi region infrastructure with autoscaling and automatic failover for resilience. There is no customer selectable region, VPC, single tenant or on premises option, so where data sits depends on Decagon's architecture, not the customer's choice. The product is delivered only as SaaS hosted by Decagon.

Sourcedecagon.ai/securityread 2026-09-05

Production runs on Google Cloud Platform in a private VPC with no public internet access and row-level-security tenant isolation, with US primary hosting and data storage residency available in Australia or the EU. Lorikeet's marketing articles name the US, Australia and the UK, while its trust page lists Australia and the EU.

Sourcelorikeetcx.ai/product/trustread 2026-09-05

Customers build agents by writing Agent Operating Procedures in natural language, or Duet generates them from the customer's own conversation transcripts, applying Decagon's deployment learnings behind the scenes. The seven industry pages are positioning and case studies, not solutions to adopt. There is no template gallery, named prebuilt agent set, blueprint library or industry pack.

Sourcedecagon.ai/product/aopread 2026-09-05

Each customer's configuration is written from its own standard operating procedures, by design. Lorikeet's test is that if you can explain to a human agent how to resolve a ticket, the agent can handle it, so time to value depends on the customer's procedure documentation. There is no template library, prebuilt agent catalog or marketplace, and the specialized agents in the Team of Agents architecture work together internally and are not assets a customer browses and adopts. Lorikeet also runs a separate site at lorikeet.tools, branded Toolshed, whose contents are not described.

Sourcelorikeetcx.airead 2026-09-05

The agent engine runs on a network of frontier and specialized models, with OpenAI and Anthropic named as providers held to zero-day retention, and model redundancy is a platform property. Routing happens inside Decagon, and customers and admins cannot select the model that powers the product.

Sourcedecagon.ai/product/overviewread 2026-09-05

OpenAI, Anthropic and Baseten are among the core subprocessors, so several model providers are in play, all under zero-data-retention agreements with no fine-tuning on customer data, and Lorikeet routes between providers by task. The routing is Lorikeet's own and model choice is tracked, not made by the customer, with no customer-facing selection or bring-your-own-model control.

Sourcelorikeetcx.ai/product/trustread 2026-09-05

A REST API for the platform sits at api.decagon.ai, authenticated with a customer issued key passed as a bearer token, and rate limited. An API key can be generated from the Developer page of the Decagon dashboard. The documentation at docs.decagon.ai sits behind a customer login, nothing on the public site links to it, and the public Integrations page describes only connections into the customer's own systems.

SourceDecagon developer documentation at docs.decagon.airead 2026-09-05

A developer documentation site at docs.lorikeetcx.ai covers a quickstart and API reference, and integrations use scoped Bearer-credential access and HMAC-signed inbound webhooks that the customer's systems call. Lorikeet also publishes public HTTP endpoints that outside AI assistants can call to query its support agent and book demos. The documentation sits behind a customer access code, so the platform API and its scope are open only to customers, and no SDK is advertised.

Sourcedocs.lorikeetcx.ai/guides/quickstartread 2026-09-05

Testing & QA and Experiments are customer facing products. Agents are validated before launch with end to end simulated conversations and unit tests. Live traffic is then routed across agent versions to A/B test changes and measure their effect on CSAT, deflection rate and other metrics at scale, with readable results the customer can compare.

Sourcedecagon.ai/product/overviewread 2026-09-05

Testing covers both sides of go-live. Before launch, a pre-go-live simulation suite lets a security or compliance team validate agent behavior before unsupervised resolution begins. After launch, Coach reviews every ticket, not a sample, whether handled by people, AI or both, shows why performance is trending up or down, proposes fixes and, with approval, implements them, with automated quality assurance, thematic analysis, segmentation and export. Coach is sold to Lorikeet customers and standalone, is metered per ticket and is generally available.

Sourcelorikeetcx.airead 2026-09-05

The agent acts only through API and MCP connectors, CRM and ticketing integrations, CPaaS platforms and SIP telephony. It does not control a browser or virtual desktop or operate software that lacks a programmatic interface.

Sourcedecagon.ai/product/integrationsread 2026-09-05

Work happens programmatically inside the customer's systems through controlled integrations, tapping into Zendesk, Stripe and internal APIs, with tool calls such as retrieving recent transactions shown as named workflow steps. No browser control, hosted session or computer use is described. Lorikeet's public HTTP endpoints let outside AI assistants call Lorikeet and do not operate any interface.

Sourcelorikeetcx.airead 2026-09-05