How Enterprise Leaders Build and Deploy Custom Glean Assistants Securely

Enterprise teams waste thousands of hours every month searching across fragmented applications for answers that already exist. A support engineer combs through closed Jira tickets and Slack channels to diagnose a recurring bug. A sales executive digs through Google Drive and Salesforce notes to verify pricing exceptions. An HR business partner repeatedly answers the same questions about parental leave and remote work stipends.

Department leaders often attempt to fix this with custom internal bots or generic AI assistants. Most of these projects stall when security and compliance teams raise immediate red flags: How do we prevent employees from querying confidential compensation sheets? How do we ensure proprietary customer data isn't leaked into an external model? How do we stop an automated workflow from executing unapproved actions in core business systems?

Glean solves this dilemma by combining enterprise-grade search across 275+ SaaS connectors with a native "permissions-first" architecture. Rather than requiring engineering teams to build custom retrieval-augmented generation (RAG) pipelines from scratch, Glean empowers department leaders to build custom Glean Assistants and Agents using natural language, curate domain knowledge securely, and automate multi-step workflows while enforcing source-system access control lists (ACLs) on every interaction.

Here is the operational blueprint for enterprise department leaders to design, configure, secure, and deploy custom Glean Assistants across their organizations.

1. The Core Architecture: Why Glean Assistants Are Enterprise-Safe

Generic chatbots treat enterprise knowledge as a monolithic vector database. Once documents are embedded, standard LLM applications struggle to know whether the user asking a question actually has clearance to read the underlying source file.

Glean operates on a fundamentally different foundation:

1. Native ACL Inheritance: Glean integrates directly with identity providers (Okta, Microsoft Entra ID, Google Workspace) and connected applications (Salesforce, Jira, Confluence, SharePoint, ServiceNow, GitHub, Zendesk, Slack). It continuously mirrors permissions so that an assistant only retrieves and synthesizes information the active user is explicitly authorized to view in the source tool.
2. The Enterprise Graph: Glean does not merely match keywords or embeddings; it maps people, content, teams, communication patterns, and document lifecycles to understand organizational context.
3. Multi-LLM Intelligence and Auto-Routing: Glean Intelligence routes user prompts to the most suitable large language models (such as Claude 3.5 Sonnet, GPT-4o, or Gemini 1.5 Pro) or connects to company-hosted models via AWS Bedrock, Azure OpenAI, or GCP Vertex AI.
4. Data Isolation and Zero Retention: Customer data is processed within a single-tenant virtual private cloud (VPC) or customer-managed cloud environment. Under strict enterprise agreements, foundation model providers never retain or train models on customer queries or internal documents.

2. Departmental Use Cases: Where Custom Assistants Deliver Immediate ROI

Before opening the builder, department leaders should target workflows characterized by high frequency, distributed documentation, and clear downstream actions:

3. Step-by-Step: Building a Custom Glean Assistant

Building an assistant in Glean requires no custom code. Department leaders or designated moderators can configure domain-specific assistants through Glean's intuitive builder interface:

Step 1: Define Role, Objective, and Persona

Navigate to Glean Apps / Agents and initiate a new assistant. Configure the system prompt to define the assistant's specific mandate:

• Role & Target Audience: Specify who will interact with the assistant (e.g., "You are the Tier-2 Technical Support Assistant for enterprise customers").
• Tone and Output Formatting: Establish rules for precision (e.g., "Provide concise bulleted steps, cite official product documentation links, and state clearly when an issue requires engineering escalation").
• Negative Guardrails: Define what the assistant must refuse to do (e.g., "Do not estimate bug delivery timelines or discuss unannounced feature roadmaps").

Step 2: Ground in Curated Knowledge Sources

While Glean Assistant has access to the full enterprise graph (subject to permissions), a custom assistant performs best when scoped to relevant domain knowledge:

• Collections & Spaces: Group high-authority documentation into curated Collections (e.g., verified runbooks, customer success playbooks, or compliance policies).
• Application & Folder Scoping: Constrain knowledge retrieval to specific Confluence spaces, Jira projects, Google Drive folders, or Slack channels to reduce noise and prevent hallucinations.
• Dynamic Prioritization: Instruct the assistant to prioritize newer documentation over deprecated archives when conflicts arise.

Step 3: Wire Actions and Workflows

A true assistant does not stop at answering questions; it completes work. Using Glean Actions and pre-built connectors, department leaders can link the assistant to external systems:

• Read & Lookup Tools: Fetch real-time status updates from Salesforce opportunities, ServiceNow incidents, or Datadog alerts.
• Write Actions: Empower the assistant to file Jira tickets, draft follow-up emails, or log notes directly into CRM records.
• Execution Governance ("Always Allow" vs. "Needs Approval"): For low-risk read actions (e.g., querying public ticket status), select Always Allow. For write actions or sensitive operations (e.g., creating a bug ticket, updating an account status, or broadcasting a notification), set the action to Needs Approval, forcing the assistant to display a preview for the user to confirm before execution.

Step 4: Validate in Sandbox Before Deployment

Before publishing to the organization, test the assistant across critical edge cases:

• Permission Boundaries: Test queries using accounts with differing permission levels (e.g., an intern versus a team manager) to verify that restricted data remains completely invisible.
• Ambiguity and Conflicting Data: Prompt the assistant with edge-case scenarios where documentation is contradictory to ensure it transparently highlights discrepancies rather than guessing.
• Prompt Injection Resilience: Attempt adversarial prompts trying to bypass role constraints or exfiltrate system instructions. Glean Protect screens runtime interactions against injection attempts with 96.9% detection accuracy.

4. Enterprise Security and Governance: The AWARE Framework

Deploying AI assistants across enterprise departments requires strict alignment with corporate infosec and compliance requirements. Glean implements the AWARE framework (developed in collaboration with security leaders from Databricks and Palo Alto Networks' Unit 42) to provide comprehensive governance:

1. Actor Intent (Identity & Access)

Every interaction with a Glean Assistant is authenticated against the company's enterprise identity provider (SSO / SAML / SCIM). Glean evaluates both:

• User Identity: The assistant strictly assumes the identity and access privileges of the user interacting with it. An assistant cannot be used as an unintended privilege-escalation backdoor.
• Agent Identity: When independent agents perform background or asynchronous tasks across systems (such as Slack or Jira), they execute under dedicated service identities with explicitly bounded permissions, ensuring clear audit attribution.

2. Work Context (Sensitive Content & Topic Restrictions)

Enterprise data repositories often contain overshared documents, accidentally exposed credentials, or sensitive customer information. Glean Protect safeguards work context through:

• Sensitive Content Policies: Automatically detects and redacts or hides overshared sensitive content, such as PII, payment card data (PCI), financial records, and API credentials.
• Restricted Topics Policies: Department leaders can configure real-time topic guardrails in natural language (e.g., "Block queries regarding employee performance ratings, salary benchmarks, or legal dispute discussions").
• Triage Workflows: Security teams can review sensitive findings, auto-hide overshared files, and synchronize alerts into SIEM/SOAR platforms like Tines or Splunk.

3. Autonomous Guardrails (Action Alignment)

To prevent autonomous agent drift, Glean checks every multi-step workflow before write actions are committed. The agent alignment model compares the proposed action against the user's initial prompt and authorization level, halting execution if the action deviates from acceptable parameters.

4. Real-Time Risk Scoring & Threat Defense

Glean Protect provides active runtime defenses against malicious inputs, prompt injection, and unauthorized data scraping, ensuring that external content (such as email threads or web-indexed pages) cannot manipulate the assistant's behavior.

5. Ecosystem Observability

Every assistant query, retrieved citation, and executed action is logged in an immutable, end-to-end audit trail. Security operations teams can stream audit events directly to corporate SIEM tools, ensuring full compliance with SOC 2 Type II, ISO 27001, ISO 42001, HIPAA, and GDPR regulations.

5. Deployment, Rollout, and Change Management

A successful deployment balances speed of adoption with operational rigor:

[Phase 1: Pilot & Validation]
  └── Department Leadership + 5-10 Power Users
  └── Refine prompt instructions, curate Collections, verify ACL limits
        │
        ▼
[Phase 2: Department-Wide Rollout]
  └── Publish to Department Group via Glean App Library
  └── Embed Assistant in daily workflow surfaces (Slack, Teams, Browser Extension)
  └── Monitor "Needs Approval" write actions and resolution feedback
        │
        ▼
[Phase 3: Cross-Functional Expansion & Measurement]
  └── Analyze usage controls, deflection rates, and query volumes in Admin Console
  └── Iterate Collections, add MCP custom tools, retire outdated assets

Best Practices for Department Leaders

• Start with Read-Heavy, Lookup-Intensive Tasks: Begin with assistants that synthesize answers from complex documentation (e.g., customer onboarding playbooks or IT troubleshooting). Add automated write actions only after initial trust and precision are established.
• Distribute Through Natural Interfaces: Don't force teams to visit a new portal. Deploy your custom Glean Assistant directly into Slack, Microsoft Teams, browser extensions, or intranet portals using Glean's Chat API and container integrations.
• Assign Departmental Knowledge Stewards: Appoint subject matter experts to maintain the underlying Collections. When an assistant cannot answer a query, use Glean's knowledge gap telemetry to draft new documentation and update source repositories.

6. Strategic Takeaway: When to Choose Glean Over Alternative Approaches

Building enterprise AI assistants involves evaluating three distinct paths:

1. Native Single-Ecosystem Copilots (e.g., Microsoft 365 Copilot): Strong within Microsoft 365 apps, but limited in connecting deeply into third-party enterprise tools like Jira, Salesforce, ServiceNow, GitHub, or Zendesk without custom integration work.
2. Custom DIY RAG Pipelines (LangChain / LlamaIndex / Vector DBs): Offers full architectural flexibility for developer teams, but demands massive ongoing maintenance: managing data pipelines, building custom connectors, updating vector embeddings, and rebuilding real-time ACL security filters from scratch. In practice, DIY RAG systems frequently fail corporate security audits due to oversharing risks.
3. Glean Platform & Custom Assistants: Delivers turn-key connectivity across 275+ enterprise applications, zero-code assistant creation, built-in multi-LLM routing, and enterprise-grade security backed by ISO 42001 and the AWARE framework.

For department leaders who need to eliminate repetitive information lookups and accelerate daily workflows immediately—without compromising corporate data security—Glean provides the fastest, most reliable, and most secure path to enterprise AI adoption.

References

• Build Custom Generative AI Apps & Agents for Enterprise
• Enterprise AI Assistant for Work | Glean
• AI Security: Protecting Enterprise Data with Glean
• How to Deploy AI Agents With Confidence: Introducing AWARE and New Glean Protect Capabilities