Talkdesk

Integration skills call external systems through endpoints, authentication methods, input parameters and response mappings configured in Talkdesk Integrations. Agents select a connection and action, fill its inputs from conversation variables, and receive the result in named outputs. MCP skills also invoke individually selected tools on customer hosted servers, using Streamable HTTP with bearer or OAuth authentication; the tools execute outside Talkdesk.

Sourcesupport.talkdesk.com/hc/en-us/articles/41553841564187-ai-agent-platform-build-skill-integrationread 2026-10-20

A supervisor agent receives the customer's message and delegates through optional routing agents to action agents that plan tasks and invoke skills. Conditions determine which child receives each request, and agents communicate only with their parent or direct children. Workflow skills combine actions, functions and branches in a separate runtime. Default limits are three hierarchy levels, ten agents per level and five skills per agent.

Sourcesupport.talkdesk.com/hc/en-us/articles/39100265145371-ai-agent-platform-core-conceptsread 2026-10-20

Knowledge Management connects a shared business knowledge base to individual AI agents, which search it while answering customer questions. Builders restrict retrieval with static Knowledge Segments or Queues, or pass a variable that changes the content scope for the interaction. The knowledge tool sits inside the agent configuration, allowing different specialists in one orchestration to use different subsets of the same central information store.

Sourcesupport.talkdesk.com/hc/en-us/articles/42992846528923-ai-agent-platform-build-ai-agent-noderead 2026-10-20

System guardrails remain active for every agentic interaction, and administrators add a custom boundary prompt of up to 300 characters. Rules can vary by channel and language, with one to five violations ending a conversation. Authentication Gate blocks sensitive skills until a deterministic verification variable is set, and unresolved interactions escalate to staff with context. Enabled actions otherwise execute without a staff approval queue.

Sourcesupport.talkdesk.com/hc/en-us/articles/49500566907291-ai-agent-platform-settings-guardrail-customizationread 2026-10-20

SAML single sign on connects the account to customer identity providers including Microsoft Azure, Google Cloud Identity and Auth0, and administrators load provider metadata and manage multiple certificates to support rotation. Agent configuration and guardrail settings require administrator access. The certification set covers SOC 2 and SOC 3 reports, ISO 27001, 27017, 27018 and 27701, ISO 22301 for business continuity and ISO 42001 for AI management, alongside PCI DSS, the Cloud Security Alliance STAR registry, Germany's BSI C5 and Spain's ENS, with HIPAA and GDPR programs on the same list.

Sourcetalkdesk.com/contact-center-platform/security-compliance/certifications-standardsread 2026-10-20

Supervisors follow live Autopilot sessions, read the exchanged messages and review recordings afterward from the History tab. The deeper AI Agent Observability view is a preview for select customers.

Its Session History lists each session's channel, orchestrator, duration, outcome (completed, escalated, timeout, error or abandoned) and error count, filters on any of them and exports session data.

Opening a session shows per turn latency, the workflows run with their inputs and outputs, warnings and errors, and a Thoughts panel tracing agent transitions and skill calls. At general availability the view moves into the CXA Operations Center.

Sourcesupport.talkdesk.com/hc/en-us/articles/50075303588507--preview-ai-agent-observability-overviewread 2026-10-20

Conversation variables and application inputs carry customer context through an active orchestration. A skill's output initially belongs only to the action agent that invoked it; instructions must pass relevant results back to a parent agent for the rest of the hierarchy to use. Escalation carries collected context to the receiving contact center. The agent has no separate customer memory with its own retention and deletion controls across conversations.

Sourcesupport.talkdesk.com/hc/en-us/articles/39100265145371-ai-agent-platform-core-conceptsread 2026-10-20

Hosted processing defaults to the United States, and customers can choose Talkdesk processing facilities in Europe, Canada, Australia or the United Kingdom to meet regional privacy requirements. Transfers out of the EEA, the UK and Switzerland run on standard contractual clauses and the Data Privacy Framework. CXA can sit above an existing cloud or on premises contact center through SIP trunks or telephone forwarding: calls travel to Talkdesk for automation and return to the original system on escalation, so the agent runtime itself always runs in Talkdesk's cloud.

Sourcetalkdesk.com/terms-of-service/privacy-noticeread 2026-10-20

A library of prebuilt AI agents is organized by industry across banking, insurance, retail, healthcare, travel and hospitality. Banking alone carries agents for phone banking, loan management, collections, card requests, payments and transfers and outbound loan prequalification, and insurance covers policy administration, claims management, billing and proof of coverage.

A customer picks an agent and pilots it in its own environment, a program pitched at proving ROI in 30 to 45 days. The commercial and residential services offering adds prepared scheduling, billing, payment and account management agents with workflows and connectors for dispatch, CRM and billing systems.

Sourcetalkdesk.com/use-casesread 2026-10-20

Incoming customer sessions enter a Studio flow that invokes the Autopilot component and starts the configured agent orchestration. Voice and digital have separate connection components, with chat and messaging routed through the same agent logic. Existing contact centers can forward calls over SIP trunks or telephone numbers, then receive an escalation back with intent and customer context carried in SIP headers or API data.

Sourcesupport.talkdesk.com/hc/en-us/articles/39696186925595-autopilot-agentic-overviewread 2026-10-20

Agentic Autopilot creation selects Talkdesk as the service provider and attaches a published AI Agent Orchestration. Builders control instructions, routing conditions, knowledge scopes and response guardrails, while voice settings govern language and speech delivery.

No foundation model is named or offered for selection anywhere in this setup, and Talkdesk does not say which provider runs the orchestration. External NLU providers belong to the separate intent based Autopilot, and connecting a third party bot swaps out the whole agent, so neither path gives a customer a model choice inside the agentic runtime.

Sourcesupport.talkdesk.com/hc/en-us/articles/40273031755675-creating-an-autopilot-agentic-in-ai-launchpadread 2026-10-20

The public platform API uses OAuth 2.0 client credentials for account access and an authorization code flow for individual user access. Identity endpoints expose caller authentication and fraud insights, validate phone numbers, and enroll or remove contacts from voice biometrics. These interfaces let outside systems operate Talkdesk services; the agent builder's MCP client instead calls tools hosted by the customer.

Sourcedocs.talkdesk.com/docs/authenticationread 2026-10-20

The built in Simulator runs chat and voice conversations against an orchestration from the canvas before a new version is published.

Builders inject application inputs of up to 5,000 characters to reproduce context a Studio flow would pass, then type messages or run Test Scripts: multiturn conversations entered in sequence or uploaded as JSON, or a script generated from a described scenario and goal, which teams keep for regression testing. Each reply opens an Insight view of the tool calls and knowledge lookups made and the coordinator's routing reasoning.

Voice tests vary line quality, language detection, voice, end of speech sensitivity, speaking speed and phrase lists. Scored Agent Evaluation is a separate preview for select customers.

Sourcesupport.talkdesk.com/hc/en-us/articles/52487140094875-ai-agent-platform-simulatorread 2026-10-20

External work runs through configured API actions, workflow functions and tools on customer hosted MCP servers. Each tool has declared inputs and structured outputs, and the remote server executes it independently. Autopilot's webchat is a customer communication channel, not a browser the agent drives. The agent configuration supplies no hosted browser, desktop session or native click and form filling controls.

Sourcesupport.talkdesk.com/hc/en-us/articles/46063702440475-ai-agent-platform-build-skill-mcpread 2026-10-20