Gainsight

Breadth across classes, with authenticated write-back rather than ingestion alone. CRM and data: Salesforce, HubSpot, Pipedrive, Zendesk, ServiceNow and Snowflake. Communication signals Staircase reads: Gmail, Outlook, Zoom, Slack, Microsoft Teams, Gong, Intercom and Jira. Enterprise search: Glean, which the vendor names as holding a dedicated native Gainsight integration building a unified index.

Outward the platform writes into the customer's systems and, through MCP, creates and updates CTAs, Success Plans and Timeline Activities on instruction from an external assistant. That spans CRM, ticketing, ITSM, data warehouse, conferencing, chat, conversation intelligence, issue tracking and search, which is spread rather than depth in one ecosystem.

Sourcegainsight.com/mcp and support.gainsight.com Gainsight CS MCP FAQread 2026-09-04

Multi-step motions running unattended across five products.

Journey Orchestrator and playbooks fire health alerts, escalation routing and renewal nudges without CSM input, the Atlas agent layer runs outreach, personalization, follow-up and end-to-end renewal motions for long-tail accounts, and specialized agents hand work along: Staircase analysts surface risk and expansion signals, the Insight Agent turns communications into sub-scores that feed scoring, the Community Moderation Agent triages violations and escalates the rest, and the Skilljar Tutor operates inside courses.

The MCP layer extends the chain outward, letting an assistant generate a Success Plan from account data and create it directly in Gainsight from a single prompt, which is orchestration initiated outside the product and completed inside it.

Sourcegainsight.com/press agentic stack launch and support.gainsight.com Gainsight CS MCP FAQread 2026-09-04

Grounded on the customer's own conversations rather than on structured fields alone, which is the distinctive claim here. Staircase analyses qualitative customer data across every communication channel, emails, meetings and support interactions, to produce relationship, sentiment and stakeholder intelligence and sub-scores on a hundred-point scale, so the grounding corpus is the unstructured record most platforms cannot read.

Alongside it, the Skilljar AI Tutor is course-aware, answering from the specific lesson a learner is in rather than from general knowledge, and MCP answers are described as grounded in live customer data from Staircase and Gainsight CS rather than in copied notes or uploaded files. No citation or source attribution on generated insights is documented, so a CSM reads a sub-score or a risk flag without seeing which email or call produced it, which matters more here than for most vendors because the inputs are private conversations. The longitudinal customer record these analyses draw on is a business record that grounds the agents rather than agent memory.

Sourcegainsight.com/staircase-ai and communities.gainsight.com MCP announcementread 2026-09-04

The autonomy boundary is published as a product decision, and one agent mechanizes it. Gainsight states which classes of work agents own and which stay with people: playbooks, health alerts, escalation routing and renewal nudges run unattended, while strategic account decisions, executive engagement and expansion conversations remain human led.

The Community Moderation Agent turns that into a mechanism rather than a policy, handling clear-cut violations itself and escalating judgment calls, which is a confidence-based hand-back rather than a scope statement.

On the agent-to-assistant path, permissions are enforced per user and the MCP server publishes an explicit list of operations it will not perform, including user management, connector configuration and Rules Engine changes, so the blast radius is bounded by design.

Outside the moderation agent, no approval queue, review step or configurable confidence threshold is documented for the CS agents, so the boundary is enforced by what the product will attempt rather than by a gate a customer operates. The Atlas layer that runs end to end renewal motions for long tail accounts is a service Gainsight staff deliver rather than part of the platform.

Sourcegainsight.com/press agentic stack launch and support.gainsight.com MCP FAQread 2026-09-04

Access and compliance are both documented, and the access surface reaches into the agent layer rather than stopping at the console.

Single sign on runs through the organization's configured Google or Microsoft provider, alongside role based access control, and the vendor's own documentation states that MCP queries respect existing permissions, that each user authenticates individually with no shared organization level tokens or API keys, and that sessions cannot be transferred between users, so an assistant sees only what its operator may already see. Automatic PII anonymization is documented alongside.

The vendor's security pages publish SOC 2 Type II, ISO 27001, HIPAA and GDPR across products, plus CCPA, and ISO/IEC 42001 for Staircase AI, the AI management system standard, which is unusual. Protections against prompt injection on the MCP path are documented.

Sourcegainsight.com/security/staircase-ai and support.gainsight.com Staircase AI MCP Security and Complianceread 2026-09-04

Inspectable scoring, and no record of what the agents did with it. Health scoring is weighted across usage, support, NPS, engagement, billing and relationship data with configurable inputs, sub-scores are exposed individually on a hundred-point scale so a CSM sees which driver moved a number, and the windows are set per customer tier.

The MCP server exposes read-only configuration inspection of field schemas, picklist values and template definitions. None of that reaches agent behavior: there is no trace of what the Staircase analysts read before flagging a risk, no log of which playbook an agent fired against which account and why, and no configuration change trail or retention control over such a record.

Sourcegainsight.com/staircase-ai and support.gainsight.com MCP documentationread 2026-10-01

A longitudinal customer record, not agent memory. CustomerOS holds a unified customer record across the post-sales journey, health sub-scores are computed over tier-specific rolling windows, and stakeholder relationship history is tracked. Those are business records rather than memory, and they ground the agents. No memory construct the agents write and read back as their own state, no per agent recall across sessions and no retention or forgetting control at the agent level is documented.

Sourcegainsight.com/staircase-airead 2026-10-01

The customer chooses the region, and the vendor says so plainly. Gainsight's own compliance documentation lists geographic controls under which admins define data processing regions, backed by a Frankfurt data center it describes as ensuring complete EU data residency with all communication analysis remaining within EU borders.

The vendor announced that facility as a second EU data center, building on European infrastructure first established in 2019 for Gainsight CS, so the option spans the platform rather than one product. Standard Contractual Clauses and EU-US Data Privacy Framework certification cover the residual operational flows the vendor acknowledges, which is legal cover rather than residency. The service is multi tenant SaaS throughout, with no self hosted, private cloud, VPC or single tenant option documented.

Sourcesupport.gainsight.com EU compliance documentation and gainsight.com/blog EU datacenter announcementread 2026-09-04

Named agents ship built, and there is a catalog the customer browses. The shipped agents are the Staircase Risk and Expansion Analysts, a Community Moderation Agent, the Skilljar AI Tutor, an Insight Agent producing sub-scores on a hundred-point scale, and the Atlas agent layer.

For selection, Gainsight publishes an MCP Library where a customer installs prebuilt Skills or copies project templates, with worked agent definitions supplied in full, one of them a meeting preparation specialist with its purpose and capabilities written out, alongside admin setup guides per product and an established playbook template library in the core product. Browsing and selecting both exist, so this is more than a set of named agents that simply constitute the product.

Sourcegainsight.com/resource/mcp-use-cases and gainsight.com/press agentic stack launchread 2026-09-04

Signals arrive from the places customer risk actually shows up, and the channels out span the whole post-sales estate. Intake: email, meetings and Zoom calls, Slack, support tickets, product usage through Gainsight PX, in-app engagement, community activity and course completion through Skilljar, with Staircase reading the qualitative channels continuously rather than on a schedule.

Health alerts fire the moment an account crosses a threshold, and sub-scores recompute over rolling windows tuned by customer tier. Channels out: in-app engagement rendered inside the customer's own product, community moderation, course-embedded tutoring, CSM alerts and escalation routing, renewal nudges, and outbound motions run by the Atlas layer, with MCP delivering account intelligence into whichever assistant a CSM already works in.

Sourcegainsight.com/staircase-ai and gainsight.com/press agentic stack launchread 2026-09-04

The customer does not choose the model and no provider is named for what powers the agents. No model list, routing policy, per agent selection or option to bring your own key is published. The MCP page names ChatGPT, Claude and Gemini, and states that MCP requires versions of those tools supporting external tool connections; that is the customer's own assistant calling Gainsight.

The vendor's Staircase security documentation makes the direction of travel explicit in the customer's favor, stating that Staircase returns structured data to the user's local MCP client and does not send data directly to LLM providers, which confirms the assistant is the caller rather than the engine. The provider is undisclosed, so the lack of choice shows as silence rather than as a named single dependency.

Sourcegainsight.com/mcp and support.gainsight.com Staircase AI MCP Security and Complianceread 2026-09-04

A documented, permission-aware MCP estate across five products, and it writes as well as reads. Gainsight publishes an MCP page of its own: Gainsight CS and Staircase servers are live for customer use, with Skilljar, Customer Communities and PX in beta.

The CS server's capability boundary is published rather than implied, listing operational create, read and update on CTAs, Success Plans and Timeline Activities, flexible querying across Company and Relationship objects with filtering, aggregation, sorting and cross-object lookups, and read-only configuration inspection of field schemas, picklist values and templates, alongside an explicit list of what it will not do including user management, connector configuration, Rules Engine changes and schema creation.

Setup is an OAuth app with PKCE created by a Gainsight admin who shares a client ID and secret with the ChatGPT or Claude admin, and a customer can query both servers in a single prompt. The MCP estate is documented across the vendor's own product page, community announcement, support articles and an MCP library of prebuilt skills and templates. No general REST API reference or SDK is published, so extensibility rests on MCP and the OAuth app framework rather than a published endpoint catalog.

Sourcegainsight.com/mcp and support.gainsight.com Gainsight CS MCP FAQread 2026-09-04

Tuning against outcomes, with nothing that tests the agents. Health models are configurable in weight and composition, the rolling window over which sub-scores are computed is set per customer tier, and the vendor describes recommendations sharpening as more signal accumulates, so a team calibrates the model against what actually predicted churn in their book and reads the effect through exposed sub-scores. That is a mechanism with a criterion aimed at the business outcome.

Nothing is aimed at the agents themselves: there is no controlled comparison of one configuration against another, no scored evaluation set, no benchmark, no regression check when a scoring model or agent prompt changes, and no sandbox or dry run before an agent acts against a live account, which matters given the Atlas layer runs renewal motions end to end. ISO/IEC 42001 certifies an AI management system, which is governance of how the vendor builds and operates its AI rather than a test surface the customer runs.

Sourcegainsight.com/staircase-airead 2026-09-04

Everything reaches its destination through an authenticated interface. In-app engagement through Gainsight PX renders guides, surveys and walkthroughs inside the customer's own product using an SDK the customer installs, which is the vendor's own instrumentation in a property the customer owns rather than an agent driving software it does not control.

Signals arrive through native connectors to email, conferencing, chat, CRM, ticketing and warehouse systems, and actions leave through those same connectors or through the MCP server, which the vendor documents as a direct authenticated connection between the user's device and Staircase infrastructure. No browser extension, headless session, screen or click automation, portal navigation or form filling is documented.

Sourcegainsight.com/staircase-ai and support.gainsight.com MCP documentationread 2026-09-04