Who Updates Knowledge After Go-Live? Minimal Release and Rollback Rules

A post–go-live knowledge release system for a custom AI Agent means: who may edit, who signs off, how production switches, how you roll back, and how you sample—not “anyone with a shared drive can change prod.” Without it, promo prices, shipping insurance, and policy text go stale faster than model upgrades. RAG build: /blog/enterprise-rag-knowledge-base-agent. Ownership: /blog/custom-agent-project-owner-raci. GeonAI ships release flow in delivery and handover; /agents are capability references only (not a public trial of 363+ presets). Email [email protected].

Why “ship and done” fails

• Business rules change faster than model versions—promos, policies, SLAs often weekly
• Vendor demo packs with no takeover → voice drifts; frontline bypasses the Agent
• No versions or rollback → a bad ingest is fixed only by “edit again,” stretching the blast window
• Mixing Prompt/tool allowlist edits with KB hot updates → incidents cannot tell config vs knowledge

Minimal roles (paste into the runbook)

Small teams may combine hats, but keep approval and publish as separate log lines—no self-approve self-publish. Boundary changes still use the three tables: /blog/custom-agent-scope-boundary-in-contract.

Minimal release flow

1. Draft: effective time, which doc it replaces, which channels
2. Approve: business Owner; compliance samples red-line content
3. Pre-prod retrieval sample: gold set + hard cases; new citations, old versions gone
4. Prod switch: bind a KB version ID; log actor and timestamp
5. Weekly sample: sessions cite the new version; wrong promises count as harm metrics

Pilot “do less harm” metrics: /blog/custom-agent-30-day-pilot-metrics. Delivery artifacts should include a knowledge release process: /blog/custom-agent-delivery-artifact-checklist.

Rollback: keep N versions, switch on incidents

• Keep at least the last N switchable KB versions (write N in the runbook; often 3–5)
• On bad policy/price incidents: switch to the prior version, freeze the bad doc, open an incident
• Within 24 hours, ship the corrected version and re-run the gold set
• Never “delete the store with no version”—you cannot prove the restore point

Keep Agent config changes separate

Prompt, tool allowlists, and handoff thresholds must not share one key with KB hot updates. Config uses change tickets/gates; knowledge uses the release flow. Mixing them leaves on-call guessing whether to roll back knowledge or config. Red flags: /blog/ai-agent-project-red-flags.

How to brief GeonAI

Name current doc Owners, desired cadence (daily/weekly/campaign freeze), and whether multi-version rollback is required. Email [email protected], /pricing, or Live chat. We deliver a rollback-ready release flow—not “edit the shared drive in prod.”