B2B SaaS의 Vertex AI 연동 메모: 영업 FAQ와 지원 이력 분리

B2B SaaS에서는 영업 FAQ, 도움말, 지원 티켓, 상담 메모, 로드맵, Slack 답변이 모두 AI context처럼 보입니다. 고객명, 계약 조건, 장애 기록, 미공개 기능이 같은 index에 들어가면 FAQ가 정보 유출 경로가 됩니다.

이 글은 구현 전에 Claude Code로 Vertex AI 연동 메모를 만드는 흐름입니다. 목적은 영업 FAQ와 지원 이력을 분리하고, 도입 전 불안을 줄이면서 고객 데이터를 넓은 AI context에 넣지 않는 것입니다.

이 글의 요점

Keep sales FAQ, help content, support history, deal notes, and roadmap data in separate buckets.
Use Vertex AI RAG, Grounding, Agent Search, and safety filters after data boundaries are clear.
Claude Code drafts classification tables, test questions, prompts, and operations notes.
Humans decide customer names, contracts, incidents, unpublished features, legal wording, and pricing answers.
Send implementation-heavy readers to training because every SaaS has different data rules.

SaaS AI FAQ 프로젝트가 실패하는 지점

먼저 공개 영업 FAQ, 지원 FAQ, 내부 확인, AI에 보내지 않을 정보 네 가지로 나눕니다. Vertex AI RAG, Grounding, Agent Search data store, safety filters는 근거 있는 답변을 돕지만 어떤 SaaS 데이터가 공개 답변에 맞는지는 대신 결정하지 않습니다.

Official references: RAG Engine, Grounding overview, Grounding with Google Search, Agent Search data stores, Safety filters, and Responsible AI.

흐름: Vertex AI 전에 데이터를 나누기

| Source | AI use | Human review |

| --- | --- | --- |

| Sales FAQ | public features, public pricing, buying questions | legal wording and pricing examples |

| Help center | operations, settings, roles, errors | old UI and deprecated features |

| Support history | anonymized patterns and frequent errors | customer names, contracts, incidents |

| Deal notes | usually no direct AI use | budget, discount, roadmap, buyer name |

Claude Code가 할 일과 사람이 결정할 일

3가지 Use case

Use case 1: Split sales FAQ and support history

Input: public FAQ, help articles, support CSV columns, ticket tags, pre-sales questions.
Output: four-bucket table: sales FAQ, support FAQ, internal review, do-not-send-to-AI.
Human review: check customer names, contracts, incident history, unpublished features, and special pricing.

Use case 2: Build test questions before Vertex AI

Input: buyer persona, pricing page, help articles, prohibited answers, escalation rules.
Output: allowed questions, handoff questions, and questions requiring source links.
Human review: decide boundaries for pricing, contracts, security, incidents, cancellation, and custom requests.

Use case 3: Create operations notes for wrong answers

Input: wrong answers, ticket logs, support corrections, FAQ update history.
Output: issue class, data source to fix, retest question, and approver.
Human review: legal wording, refund terms, incident explanations, customer impact, and release timing.

복사해서 쓰는 프롬프트


Create a Vertex AI integration note for a B2B SaaS. Goal: separate sales FAQ from support history and avoid sending customer data into broad AI context. Create four buckets, twenty test questions, handoff rules, source-link rules, and operations notes. Do not paste real customer values. Main CTA: /training/.

실행 가능한 확인 코드


const faqCandidates = [{ source: 'support_history', title: 'Client incident response', bucket: 'sales_faq', fields: ['customerName', 'incidentId', 'contractNote'] }];

const sensitiveHints = [/customer/i, /contract/i, /incident/i, /discount/i, /account/i];

const findings = faqCandidates.flatMap((item) => {

  const sensitiveFields = item.fields.filter((field) => sensitiveHints.some((pattern) => pattern.test(field)));

  return item.bucket === 'sales_faq' && sensitiveFields.length ? [{ title: item.title, issue: sensitiveFields.join(', ') }] : [];

});

console.table(findings);

if (findings.length) process.exitCode = 1;

Pitfall: RAG가 잘못된 데이터를 그럴듯하게 만든다

Cause: RAG brings answers closer to the indexed data. If the indexed data is wrong for sales use, the answer becomes confidently wrong or unsafe. Fix: split data before ingestion, use anonymized support patterns, and keep contracts, incidents, and roadmap notes out of public FAQ context.

FAQ

Q. What should come first? A. A data classification table, not a model setting.

Q. Should all support history be removed? A. No. Patterns and tags can help, but customer-specific details need separation.

Q. Why training? A. Vertex AI, Agent Search, RAG, data boundaries, and security review differ by SaaS.

상담으로 이어가기

ClaudeCodeLab training can help create FAQ buckets, test questions, and operations notes before Vertex AI or Agent Search implementation.

실제로 확인한 결과

This article checks frontmatter, official links, three use cases, input-output-human review labels, code fences, and the /training/ CTA. The JavaScript sample detects customer, contract, and incident fields mixed into sales FAQ candidates.