Week 1–2: one journey, one corpus
Pick the user story that unlocks revenue or retention — onboarding FAQ, sales enablement, support tier-one — not "AI for everything." Inventory documents, note PII, and draft fifty evaluation questions from real tickets or calls.
Freeze non-goals in writing: no fine-tuning, no ten integrations, no autonomous agents until retrieval works. Startups that scope narrowly ship; those that chase parity with ChatGPT do not.
Week 3–4: RAG on staging
Ingest, chunk, embed, and wire a minimal UI or Slack/WhatsApp surface. Add citations for trust. Run daily eval sessions with founders and domain experts marking failures. Fix chunking and prompts before swapping models.
Security basics early: auth, rate limits, no production secrets in prompts. If you are pre-seed, hosted staging with anonymized samples is fine; if enterprise pilots wait, plan VPC migration in phase two.
Week 5–6: harden and narrate
Add logging, admin upload, and a one-page architecture diagram for diligence. Record a demo on real data with failure cases — investors respect honesty about handoff to humans.
Sabrixa's six-week copilot playbook mirrors this rhythm for B2B startups: deliverable is operable software plus eval spreadsheet, not a slide claiming "AI-powered" without metrics.