The Production AI Playbook: Deploying Agents at Enterprise Scale — Sandipan Bhaumik, Databricks

TL;DR

• 1

  Model choice should come late, not first: Bhaumik says most teams start with GPT vs. Claude debates, but in one 8-week banking POC his team picked the model in week seven after building evals, tracing, and data plumbing first.

• 2

  Three gaps kill production AI: He frames the recurring failures as an observability gap, an evaluation gap, and a governance gap, meaning teams cannot see what the system did, cannot measure what matters, and do not know who owns failures at 3 a.m.

• 3

  Evaluation needs three layers: Deterministic checks catch formats and PII, semantic checks use LLM-as-a-judge for groundedness and relevance, and behavioral evals catch expensive issues like an agent making three duplicate database calls for one account-balance answer.

• 4

  Data quality becomes unforgiving with agents: Bhaumik says humans tolerate bad data in reports, but agents do not, because they return wrong answers confidently, which is why he spends about 60% of project time on question data and tracking data foundations.

• 5

  Tracing turns AI from a black box into an operable system: In a banking dispute flow, traces exposed every step from intent classification to policy retrieval to guardrails, and later helped diagnose a CSAT drop caused by an outdated policy document in the vector database.

• 6

  Governance includes prompts, models, and incident response: He argues prompt versioning should follow enterprise change management, model updates must be tested against your own eval set, and AI incidents need a playbook of detect, diagnose, contain, and fix tied into ITSM systems.