RAG returns irrelevant context. Agents fail on edge cases. Costs grow unpredictably. Latency hurts the user experience. Security gaps appear around prompts, tools, and data.
01Unclear scope
Teams start with a model call, a prompt, and a hopeful workflow. Without clear architecture, ownership, baseline, or acceptance criteria, the project expands without becoming reliable.
02RAG failure
Chunking, metadata, ranking, and evaluation are not measured. The team sees plausible answers and cannot explain which evidence was used or why the system failed.
03Agent reliability
The happy path is impressive, but tool permissions, state, retries, fallbacks, and human handoffs are not covered by a regression suite.
04Production gap
There is no release path, cost attribution, observability, runbook, or clear handover. A working prototype becomes an operational risk instead of a product capability.
The missing engineering system
The missing pieces are usually architecture, evaluation, security, optimization, deployment discipline, and an owner for the outcome.
Defined
The system needs clear boundaries, data flows, integrations, and an owner before more features are added.
Measured
A test set and regression loop turn vague quality concerns into engineering decisions.
Hardened
Prompt injection, tool permissions, data leakage, and unsafe fallback behavior need active review.
Operable
Deployment, observability, cost controls, runbooks, and handover determine whether the system can survive production.