memories/glean-platform-management-2dac9fc1.md

memory

Scope

Access, provisioning, usage/credit monitoring, system configuration, and release/update cycles for the Glean AI platform.

Access and Provisioning

  • User Access: Granted by default; AI Ops manages integrations and admin-level access.
  • API Credentials: Strictly scoped tokens required; global tokens are prohibited. Admins manage self-service tokens; otherwise, AI Ops issues time-bound tokens.
  • Offboarding: Relies on SSO assignment and SCIM deprovisioning; verify current Okta/SCIM configuration for validation.

Credits and Usage

  • Monitoring: AI Ops tracks FlexCredit consumption.
  • Allocation: Centralized visibility is the goal; however, definitive policies for department allocation, recharges, or exception handling are pending definition.
  • Controls: Glean-native dashboards and budgets are recommended for implementation to manage limits.

System Setup

  • Connectors: AI Ops manages identification, configuration, testing, and documentation of data owners/validation.
  • Permissions: Glean respects the querying user's native permissions; agent/document visibility follows user-level access.

Maintenance and Updates

  • Infrastructure: SaaS-level upgrades are managed by Glean.
  • Platform-Specific: AI Ops handles connector changes and sync jobs.
  • Gaps: Currently missing a formal release calendar, upgrade notification process, and roll-back procedure.

Open Items

  • Confirm Okta/SCIM deprovisioning.
  • Define credit allocation and exception handling policies.
  • Establish and publish connector release/rollback procedures.

Agent Catalog Cleanup (2026-08-28)

  • Progress: Shaurya Mittal reduced the agent count from ~1,450 to ~900 agents before departing (last day 2026-08-28).
  • Revenue Team Cleanup: John Kipe compiled a dedicated keep/cut list for Revenue-team agents.
  • Ongoing Governance: The team is moving toward a decentralized maintenance model where individual leads are responsible for periodically pruning their own agents, moving away from relying solely on manual, central cleanup efforts. A categorization and governance document is currently being drafted.

Deprecation and Migration (Jan 2027)

  • Strategy: Migration to internal-build "[[Docebo Knowledge]]" (codename "[[Zive]]").
  • Rationale: Non-renewal of Glean contract effective Jan 2027.
  • Status: Kickoff held Aug 26, 2026. "AI Ops" is customer-zero.
  • Key Gaps: Workflow/agent orchestration (Glean currently ~15k runs/day) and specific connectors (Gong, Salesforce) are not yet in Zive scope; these are critical blockers.
  • Risk: Concurrent cutover with [[UnifyApps]] (also expiring Jan 2027) represents a $377K platform risk; requires staged execution.
  • Visibility: Native Glean dashboard disabled (high credit usage); team currently using Docebo's DUD MCP for usage monitoring.

Glean → Docebo Knowledge Migration (Aug 31, 2026)

  • Gap Analysis: [[Samuel Adams]] and [[Sam Irizarry]] have kicked off a 3-week gap analysis comparing Glean functionality to Docebo Knowledge requirements.
  • Agent Schema Extraction: Exploration underway on extracting full agent schemas (steps/connectors/skills) from Glean to facilitate migration scoping; [[Samuel Adams]] is researching extraction methods.

Credit and Budgetary Outlook (Sep 2, 2026)

  • Cost Status: Consumption has been effectively free to date; however, usage is currently ~2.5x over projection (up from 2x as of July 30).
  • Migration Budget Risk: Moving to [[Docebo Knowledge]] (codename [[Zive]]) will shift these costs to internal budgets, which remain unfunded for 2027. Estimated impact could reach six figures.
  • Future Architecture: Investigating if future compute routes through AWS/Bedrock, with a preference for cheaper/open models as recommended by [[Sam Irizarry]]. [[Francesco Convia]] is the point of contact for sizing AWS-based costs.

Agent Configuration Extraction (Sep 2, 2026)

  • Status: Complete. [[Preet Patel]] has successfully implemented a method to extract full Glean agent configurations by parsing logs from a dedicated S3 bucket.
  • Methodology: Preet created a new Athena table over the S3 bucket logs. Each agent run includes a proto text config field in step 1, which parses into JSON containing steps, branches, and full instructions.
  • Limitations: Coverage limited to agents that have run since Dec 2025; agents inactive in this window are candidates for deletion pending a creation-date check. Tool calls remain abstracted and are not currently extractable.
  • Outcome: Successfully scoping Glean agents for migration to [[Docebo Knowledge]]. Future implementation may involve dumping output to the AI Hub RDS. (Ref: APMT-338, APMT-318).

Known Issues (Sep 2, 2026)

  • Admin Insights 403 Errors: Reported by [[Samuel Adams]] regarding failure to load Glean Admin Insights (403 status). Giuseppe Bonanno suspects an AWS IAM / permissions regression related to data-access method changes.
  • Verification: Pending investigation to determine if this is a new issue or related to the existing disabling of the native Glean dashboard.

Agent Configuration Extraction (Update 2026-09-09)

  • Status: Per-agent summary schema is locked as of 2026-09-09. Schema fields: agent kind, generated step-by-step description, has_branch, has_auto_agent, is_slack_agent, step_types, data_sources, trigger_type, available tool actions, base_model, unique used_models, has_subagents, has_skills, owner, runs observed.
  • Methodology Rule: The parsed export is the source of truth, as Glean Native drops steps for autonomous agents. actions.json must be used to hydrate action IDs into readable names.
  • Coverage: ~650 agents on portal; ~560 in filtered summary; ~80–100 dropped for zero observed runs.
  • Next Steps: Flatten to CSV for pivot analysis; prioritize deterministic parsing; LLM-based parsing requires evaluation.
  • Known Gaps: Action names lack target system specificity (e.g., "Search learning items" = Docebo LMS); base_model is occasionally missing; Glean-authored tool skills are non-extractable.
  • Governance: Tracks to APMT-338 and gitlab.com/docebo/ai-operations/glean-export (agent-parsing branch).

Credit and Budgetary Outlook (Update 2026-09-09)

  • List-Price Consumption: Athena token usage analysis shows ~$350K–$356K in consumption (at list price) from Nov 2025–Sep 2026. This represents 2.5–3x the current contract value ($100–150K).
  • Breakdown: [[GPT-5.4]] is the primary driver; Sonnet variants account for ~$148K combined. 18 models remain unpriced or low-confidence, making this a floor estimate.
  • Budgetary Risk: This confirms the previously estimated 'six-figure' phantom line item for 2027 that will shift to internal budgets post-migration to [[Docebo Knowledge]].
  • Next Steps: [[Samuel Adams]] plans to review these figures with [[Nitin Chopra]].
  • Ref: Drive "Glean LLM Usage — List-Price ROI Backing Data (fixed)" (2026-09-09); Granola "Preet / Samuel — Glean Agent Audit + Sync".