What Breaks When You Ship AI Without Rollback

Prompts, indexes, and model versions drift like code — but most AI launches have no revert path. Rollback is production hygiene, not pessimism.

Teams ship AI features with the same ceremony as a button color change: merge, deploy, hope. When retrieval quality drops or a model update shifts behavior, there is no known-good version to return to — only firefighting and manual prompt edits in prod.

If you would not ship database migrations without revert, do not ship model and index state without it either.

What needs rollback in AI systems

Rollback targets are not only application binaries:

| Layer | What drifts | Rollback artifact | |-------|-------------|-------------------| | Model | Vendor version, routing | Pinned version + routing table | | Prompt / policy | Template edits | Versioned prompt registry | | Index | Bad ingest, bad chunking | Snapshot or rebuild from known corpus | | Retrieval config | top-k, filters, rerankers | Config version | | Tool permissions | MCP scope creep | Policy version |

Treat these as release artifacts, not tribal knowledge in someone's notebook.

Failure stories (anonymous, common)

The bad ingest. A bulk upload poisoned embeddings. Answers cited deprecated security policy for a week until a customer noticed. No snapshot — full re-index under pressure.

The prompt "hotfix." An engineer patched prod prompt text to stop hallucinations. Fixed the demo path; broke the audit path. Nobody remembered the previous wording.

The vendor upgrade. Model version improved average case; regressed on regulated answers. No pin, no canary — rollback meant "turn it off."

The RAG tweak. Increasing top-k reduced "I don't know" rates — and increased confident wrong answers. Without evals, the tweak looked like success until compliance did not.

Each story is RAG without context engineering meets missing ops discipline.

Rollback pairs with evals

You cannot revert to good if you never defined good:

• Golden questions from support, audit, and operators
• Expected citations or abstain behavior
• Budgets on cost and latency per path

Humans in the loop when stakes are high covers review paths; rollback covers when review was not enough.

Platform patterns

On Services deployments we expect:

• Versioned prompts, policies, and retrieval configs
• Canary or shadow traffic where stakes warrant it
• Runbooks for index rebuild and model pin revert
• Ownership — named on-call for the memory layer, not "the AI guy"

Same bar as any tier-one service.

Minimum viable rollback

Before calling AI "production":

1. Can you name the last known-good model + prompt + index config?
2. Can you revert in one operational step without redeploying unrelated services?
3. Do evals run on every change to retrieval or prompt defaults?
4. Is there an incident record when you rollback — tied to what context changed?

Four nos means you have a feature flag, not production AI.

Next steps

Read failure modes of autonomous agents in production for what happens when rollback meets autonomy.

Scoping rails? Signal on Contact or /memory-infrastructure.

Related: Operational intelligence in regulated environments, Capture once, use anywhere in production.