Cognigy

Cognigy takes authenticated action across the contact center estate, with the connectors named and the mechanism documented. It is technology agnostic by design, integrating with any CCaaS, CRM or business system, including Genesys, Avaya, NiCE CXone, Amazon Connect, SAP and Salesforce.

Handover Providers are configured per platform, with a NiCE CXone non integrated provider added in 2026.9 for handover without a platform level integration, a Salesforce MIAW provider maintained, and Agent Copilot embeddable directly in the Salesforce Agent Console. Beyond handover, Connections hold credentials, Extensions package custom Nodes, and Code Nodes and the Node Execution API reach complex legacy systems and internal tools that have no connector. The actions are transactional: rebooking, cancellation, order status, invoice requests, payment and refunds are named as outcomes the agents execute.

Sourcedocs.cognigy.com Handover Providers and release notes 2026.9, cognigy.com AI Ops and orchestration pageread 2026-09-06

A real step model for orchestration is documented. Agents are built as Flows of Nodes in a unified visual builder covering chat and voice together, and the documentation names the node classes that make execution work: Say and Question Nodes for turns, LLM Prompt and LLM Entity Extract for model calls, Go To for jumps within a Flow, Code Nodes and Transformers for logic the builder does not express, Search Extract Output for retrieval, and Handover to Human Agent as a terminal path.

Agentic AI reasons across channels and orchestrates actions to move from conversation to resolution, executing transactions such as rebooking, cancellation, order status, invoice requests, payment and refunds. Flows compose into Snapshots that carry all resources, logic, content, configuration, NLU models and Extensions as one deployable unit, so the orchestration comes with a release artifact.

Sourcedocs.cognigy.com Node Reference and Snapshots, cognigy.com AI Ops and orchestration pageread 2026-09-06

Knowledge AI is a named retrieval product with its own license, quota model and vector store. Customers upload knowledge as PDF, Cognigy Text or web pages, or connect external systems through Knowledge Connectors, to create Knowledge Sources, and Cognigy describes the purpose plainly as uploading documents and using RAG so agents give accurate, context aware responses.

Retrieval is wired into the Flow through the Search Extract Output Node for agents and the Copilot Knowledge Tile for human agents, which returns source links naming the files an answer came from. The corpus is licensed in Knowledge Chunk quotas allocated across organizations through the Management UI or the Cognigy.AI API, and the on premises release notes name Weaviate as the vector database, telling operators to upgrade it to v1.22.5. That is a maintained, queryable index over the customer's documents.

Sourcedocs.cognigy.com Knowledge AI overview and activation, release notes 4.68read 2026-09-06

Handover to people is engineered as a first class subsystem; approval before the agent acts is not.

Handover Providers are configured per contact center platform, a dedicated Handover Flow controls the interaction with the human agent and the transfer, a Handover to Human Agent Node sits in the Flow, and Agent Copilot workspaces suggest next actions and surface knowledge with source links to the person now holding the conversation.

Escalation is a designed path with context intact rather than a failure exit, and role based access control enforcing separation of duties limits who can change agent behavior.

No review and approve surface over agent output is documented: no pre send queue, draft review mode, confidence threshold or per action sign off appears in the material reached. The agent answers and acts on its own until a Handover Node or a configured condition moves the conversation to a person.

Sourcedocs.cognigy.com Agent Copilot getting started, Handover Providers and Handover Featuresread 2026-09-06

Cognigy holds a broad set of certifications and documents its access model beside them. Its public trust center lists ISO 27001, ISO 27701 and ISO 42001. Cognigy also holds SOC 2 Type II and PCI DSS 4.0, the latter with independent QSA attestation, along with the TISAX and BSI C5 attestations, GDPR and HIPAA alignment, and work under way mapping SOC 2 controls to HITRUST CSF. ISO 42001 certifies an AI management system specifically, which few vendors hold.

Access is governed by granular role based access control enforcing separation of duties, with the Management UI distinguishing admin rights for organization level tasks such as allocating Knowledge Chunk quotas, alongside end to end encryption in transit and at rest and detailed audit logs.

Sourcetrust.cognigy.com, docs.cognigy.com Administration and Knowledge AI activationread 2026-09-06

Both who changed the agent and what it did in a run can be inspected. Detailed audit logs capture every configuration change and user action, recording who altered the agent and when, and extend to platform administration rather than only conversations.

The Interaction Panel executes a Flow node by node in Debug Mode against a chosen Flow, Locale and Snapshot, shows the agent's memory during a session, and can follow a live session by user identifier, so a specific run can be reconstructed step by step rather than summarized.

Around them sit Insights transcripts, from which a Playbook can be generated, and analytics and reporting across the conversational estate.

Sourcedocs.cognigy.com Interaction Panel and Insights, cognigy.com AI Ops and orchestration pageread 2026-09-06

Conversation state exists and can be inspected, but no memory layer is described around it. Cognigy keeps conversation context across channels, so a handover carries history intact, and the Interaction Panel lets a developer inspect what sits in the agent's memory during a session, which confirms a runtime state store.

No memory layer with a stated scope and lifetime is documented in the material reached. Nothing says what an agent keeps about a contact between separate conversations, whether state is scoped per user, per session or per organization, how long it lasts, or how an operator inspects or deletes it for one person. Knowledge Sources are the knowledge layer, not memory. The developer reference is large, and a targeted read may well find a documented state model behind the memory inspection.

Sourcedocs.cognigy.com Interaction Panel and Developer Referenceread 2026-09-06

Cognigy runs as a genuine self hosted product, and the operational detail shows it.

Its release notes document on premises installations at a level of specificity that cannot be marketing: a YAML based data redaction configuration introduced in 2026.9 replacing environment variables, with rollout percentage control and per organization exclusions via values.yaml; an ingress controller feature flag governing Traefik middleware resources; parameterized certificate secret names in the sbc-sip DaemonSet and StatefulSet; a Weaviate version upgrade instruction addressed to on premises installations running Knowledge AI; and a concurrency limit for Playbook execution set through values.yaml.

Licensing distinguishes shared SaaS, dedicated SaaS and on premises routes, with Knowledge AI quotas administered differently for each, so a customer can run Cognigy inside its own infrastructure with Kubernetes level configuration exposed to the operator.

The on premises route needs an operator able to run the underlying Kubernetes cluster, as the configuration detail makes plain.

Sourcedocs.cognigy.com release notes 2026.9, 4.68 and 4.72, Knowledge AI activationread 2026-09-06

The customer gets reusable parts to assemble, not working agents to adopt. The building blocks are real: a Node library covering conversation, logic, LLM prompting, entity extraction and handover; Extensions for packaging custom Nodes; Packages for exporting and importing resources, including Knowledge AI resources and Playbooks, between installations; Snapshots capturing a complete agent as a portable archive; and prebuilt handover providers and connectors for the contact center estate. Those move work between projects and environments.

Nothing is published as a working agent a buyer adopts. No template gallery, agent library, industry starter pack or marketplace of prebuilt agents is documented; the route to a live agent is building a Flow from Nodes, configuring Endpoints and handover, and testing it with Playbooks.

Sourcedocs.cognigy.com Extensions, Packages, Node Reference and Snapshotsread 2026-09-06

Voice and digital channels are equal, with the telephony stack built in house.

Cognigy serves voice, chat and digital channels with intelligent IVR and self service in more than a hundred languages, delivered through Endpoints that separate the agent from the channel, with Webchat v3 as the web surface and a Voice Gateway with its own self service portal and SIP infrastructure; the release notes expose SBC configuration and TLS handling at the level a real telephony product needs.

Endpoints are deliberately kept out of Snapshots so channel configuration and agent logic version independently. Production scale is evidenced: tens of thousands of concurrent sessions and peaks around 375,000 interactions a day, across several independent inbound queues including telephony.

No catalog of scheduled or event driven triggers is documented in the material reached, so work arrives by channel rather than by event.

Sourcedocs.cognigy.com Endpoints, Voice Gateway and release notes, cognigy.com platform pageread 2026-09-06

The customer configures the provider and picks the model as an ordinary setup step. Cognigy's docs tell users to get an account with one of the supported LLM providers, configure the provider in Settings, and select the model from the supported list, with individual Nodes stating which model capabilities they need. The provider set is maintained as a shipped feature: the 2026.9 release added Google GenAI as a new LLM provider with the Gemini 3 family, naming gemini-3.1-pro-preview, gemini-3-flash-preview and gemini-3.1-flash-lite-preview.

Cognigy states that customers are subject to the terms of the generative AI providers they use and disclaims responsibility for those services, which only makes sense where the customer holds the provider relationship. The same choice extends beyond the language model to natural language understanding, speech to text and text to speech, so a buyer assembles the stack, brings the account, selects the model and can swap it later.

Sourcedocs.cognigy.com Generative AI and LLMs section, 2026.9 release notes, Knowledge AI overviewread 2026-09-06

Cognigy publishes several APIs onto its own platform, plus a code level extension model. The Developer Reference covers the Cognigy.AI API, the Playbooks API with its own OpenAPI specification for creating and scheduling test runs remotely, the Node Execution API function reference and the Knowledge Connector API function reference.

Alongside them are Extensions for building custom Nodes, Packages for moving resources between installations, and Code Nodes, Transformers and CognigyScript for dropping to code where the low code builder stops. Administration APIs reach the Management UI, and Knowledge Chunk quotas can be set through the API as well as the interface. An outside caller can drive Cognigy to run a test, execute a node, feed a knowledge connector or administer an organization.

No MCP server for the platform is documented in the material reached.

Sourcedocs.cognigy.com Developer Reference, Playbooks, Knowledge Connector API and Node Execution APIread 2026-09-06

A scored regression harness ships with its own documentation section, plus a release path. Playbooks sit under Test in Cognigy's docs: a Playbook holds up to 50 Steps, each carrying input text and data sent to the agent and Assertions that check the response and aspects of the Flow against a defined value, and assertions can be inverted to check that a value is absent.

Runs execute as background Tasks with results in a Runs tab filterable by status and creator; up to ten Playbooks run at once, with the limit configurable on premises through MAX_CONCURRENT_PLAYBOOK_EXECUTIONS in values.yaml; and a Playbooks API allows remote creation and scheduled runs against the OpenAPI specification. Playbooks can be generated from a live conversation or an Insights transcript and exported as a Package. Cognigy's own guidance calls Playbooks automated QA testing and uses the word regression.

Around them sit the Interaction Panel for manual runs against a chosen Flow, Locale and Snapshot with node by node debugging, and Snapshots, which package a complete agent for promotion between development, QA and production endpoints.

Sourcedocs.cognigy.com Test section, Playbooks and Interaction Panel, Cognigy support center Playbooks guidanceread 2026-09-06

No browser, desktop or remote computer control is documented. Cognigy's agents reach outside systems through Handover Providers, Connections, Extensions, Code Nodes and the Node Execution API, all credentialed programmatic paths. Its own surfaces are Endpoints, a Webchat widget, a Voice Gateway and the Agent Copilot workspace embedded in a contact center console: channels the agent speaks through, or the vendor's own interfaces, rather than software it operates. Cognigy describes its agents as acting on backend systems through APIs and connectors.

Sourcecognigy.com AI Ops and orchestration page, docs.cognigy.com Node Reference and Extensionsread 2026-09-06