Dialpad

Skills grant an agent specific data operations from connectors, each an API endpoint such as Create Contact, Update Lead or Log Call Note, so the agent books appointments, checks order status, updates records and creates tickets in the systems behind it. Connectors cover Bullhorn, Salesforce, ServiceNow, Zendesk, HubSpot, Google Calendar and Workspace, Outlook, athenahealth, Calendly and Zoho CRM, and administrators enable each one, add a connection and pick which operations a skill may use. Instructions and guardrails inside the skill set when an operation runs and what it must never touch.

Sourcehelp.dialpad.com/hc/en-us/articles/54542536836891-create-an-agentic-ai-agentread 2026-10-19

Workflows combine variables, conditional branches, reusable shared flows and API calls across voice and digital conversations, and a call can be parked while an outside process runs, then picked up again when it finishes. Skills expose chosen flows and data operations to the agent, and a greeting skill can run a lookup on caller ID or launch variables before the normal conversation starts, after which the agent's own orchestration takes over. Changes are saved as drafts and published to connected channels, and the most complex Expert flows are switched on by Professional Services.

Sourcehelp.dialpad.com/hc/en-us/articles/55206515822235-workflowsread 2026-10-19

Each agent draws on its knowledge throughout every conversation, with no skill needed to reach it, and searches it semantically. Zendesk, Document360, Box, Guru and a chosen Google Drive folder sync every 24 hours, a configurable crawler imports whole sites or chosen pages with labels, CSS selectors and crawl delays, and uploaded documents of up to 5MB sit alongside them. Uploaded files do not sync, so a changed document has to be removed and uploaded again, and the knowledge base language is set when the agent is created.

Sourcehelp.dialpad.com/hc/en-us/articles/54542536836891-create-an-agentic-ai-agentread 2026-10-19

A customer who asks for a person is escalated straight into the contact center, from digital channels as well as voice, with the conversation context attached. Guardrails in each skill can require the agent to ask for confirmation before acting, name data it must not change and say when to escalate, and Guardian watches live interactions to keep the agent on task and limit data exposure. Validation flags weak instructions before publishing. Once a skill is live, its actions run without a staff member approving each one.

Sourcedialpad.com/jp/blog/agentic-ai-from-pilot-to-productionread 2026-10-19

Single sign on runs over SAML with SCIM provisioning from Okta, Azure, Google Workspace, OneLogin and other identity providers, permissions apply at company, office, department and user level, contact centers carry role based access control, and only company or office admins can create an agent. The platform holds SOC 2 Type II and annual ISO 27001:2022, 27017 and 27018 certifications, and healthcare customers can sign a BAA with one click. Detailed audit reports sit in the Trust Center behind an access request.

Sourcedialpad.com/trustread 2026-10-19

The AI Agentic Analytics report tracks deflection, escalation, resolution rate, predicted CSAT and handle and resolution times, and a skill leaderboard counts how often each skill started and reached its completion criteria, filtered by contact center, date, channel and agent.

Each conversation in History opens to its transcript and metadata, with filters for workflow use, unanswered questions and feedback, and admin changes land in an exportable event changelog. Which data operation ran inside a given conversation, with what inputs and result, is not shown, despite the product page's promise of end to end decision tracing and rollback.

Sourcehelp.dialpad.com/hc/en-us/articles/54859831644571-ai-agentic-analytics-reportread 2026-10-19

An agent works from launch context, caller ID, the number dialed and launch variables, plus whatever its skills look up in connected systems during the conversation. When it escalates, the contact reason, authentication status, skills attempted and desired outcome travel to the person who picks up. Nothing in the agent settings keeps a memory of its own from one conversation to the next, with a scope, retention period or deletion control separate from CRM and contact center records, and voice sessions close after 90 seconds of caller silence.

Sourcedialpad.com/jp/blog/agentic-ai-from-pilot-to-productionread 2026-10-19

Resolve runs in Dialpad's cloud. Contact lists, call records, recordings and transcripts are stored in the United States on Google Cloud Storage unless a Dialpad administrator has selected a different region, yet no list of regions or way to choose one is given, and the security FAQ still describes US storage with no exception. Nor is there a customer VPC, private tenant or on premises edition for the agent, so residency outside the US rests on that one exception in the privacy terms.

Sourcedialpad.com/legalread 2026-10-19

New skills can start from predefined templates whose instructions, actions and guardrails are then edited to fit, and voice workflow templates cover standard menus, customer context lookups and routing by number. Skills Mining, an early adopter feature for English voice contact centers, reads past calls and turns high impact paths into prefilled draft skills. Templates cover common tasks such as booking or order status, and each still needs the customer's own connector before it can act.

Sourcehelp.dialpad.com/hc/en-us/articles/55203978141979-workflow-templates-voiceread 2026-10-19

Voice routing hands calls from a Contact Center, Department or Main Line to an AI agent during open or closed hours, and every new agent gets its own web chat channel, which can be swapped for another digital channel. A greeting skill fires on each new launch, inbound calls, web chat and outbound bot conversations alike, using caller ID, the number dialed and launch variables, which is how outbound appointment reminders run. Published digital workflows attach to channel events, and Android and iOS bindings bring the agent into native apps.

Sourcehelp.dialpad.com/hc/en-us/articles/54542536836891-create-an-agentic-ai-agentread 2026-10-19

A multi model architecture combines Dialpad's own speech, intent and task models with third party language models such as OpenAI and Google's Vertex AI, vetted under processing terms that bar training on customer data. Admins shape behavior through the task description, tone, skills, guardrails and knowledge, but have no setting to pick a model or bring their own key, so Dialpad decides which model serves each step and can change providers without the customer reconfiguring anything.

Sourcehelp.dialpad.com/hc/en-us/articles/53072795539995-data-security-ai-language-modelsread 2026-10-19

Developers reach the platform through REST APIs authenticated by API key or OAuth, including the AI Virtual Agent and Agent Assist APIs, a Digital Engagement API and event subscriptions for calls, SMS, contacts and agent status.

The agent's own developer site adds versioned REST APIs, a JavaScript widget SDK with callbacks and profiling, and custom data fetchers for unsupported knowledge sources, and Android and iOS bindings ship as public repositories with sample apps.

A hosted MCP server that exposes calls, messages, contacts and transcripts to outside assistants is in an early access program.

Sourcedeveloper.karehq.comread 2026-10-19

Admins check an agent in a Test preview, typing a question and reading the answer, and validation flags weak instructions and guardrails with suggested fixes before publishing. Proving Ground adds simulated runs, 10 to 15 scenarios that probe edge cases and guardrail limits and report a task completion rate before launch, but it is reached from Skills Mining results, a feature still limited to early adopters, and it has no setup guide or plan listing of its own. After launch, skill completion and resolution rates show how changes perform in production.

Sourcedialpad.com/jp/blog/beyond-faq-bots-mastering-agentic-ai-with-omnichannel-skills-miningread 2026-10-19

Agents reach business systems through connectors, data operations and workflow API calls, and a crawler imports website content into knowledge ahead of time, not during a conversation. The web chat widget and the mobile bindings embed the conversation in the customer's site or app. Screen work, such as filling in a form on a portal that has no API, is outside what Resolve agents do; a system without a connector is reached through a custom action that calls its API.

Sourcedialpad.com/resolveread 2026-10-19