A planning template, scoring rubric, and scalability guide for generative-AI projects — so you can take an idea from "wouldn't it be cool if" to a governed, shippable plan.
Most GenAI projects don't fail in the model. They fail in the plan — no business case, no data story, no guardrails, no owner for the 70% that is people and process. This repo is the plan.
I benchmarked 15 public GenAI "project books" — 10 from GitHub (engineering scaffolds, RAG/LLMOps templates) and 5 from major consulting firms (BCG, McKinsey, Deloitte, Accenture, KPMG) — against a single scoring rubric.
The finding: not one was a complete, launch-ready project book. GitHub templates nail the build and skip the business case. Consulting frameworks nail strategy and governance and skip the build. Every one tripped at least one gate (cost, PII, secrets, guardrails, compliance, approvals, rollback).
So I took the best ideas from all 15, folded them into one structure, and made it right-sizable — from a two-page startup brief to a full enterprise governance pack. That's what's here. Take it and build your own.
| File | What it is |
|---|---|
GenAI_Project_Book_Template.md |
The plan. An 8-section fill-in template (Overview, Value, Data, Solution Design, Build, Responsible AI, Delivery, Roadmap) + appendices. |
GenAI_Project_Book_SCORING_RUBRIC.md |
The rubric. Score any project book 0–4 per criterion, weighted to 100, with 8 hard gates that cap an unsafe project at "Not ready." |
GenAI_Project_Book_Scalability_Guide.md |
The tiering layer. Pick a tier by business size + risk; see which sections are mandatory, combinable, or skippable. |
GenAI_Project_Book_AI_Authoring_Guide.md |
For AI agents. A machine-followable guide so an agent can draft the project book for you (ACTION / OUTPUT / DONE-WHEN per section). |
BENCHMARK.md |
The benchmark (readable). All 15 examples scored and ranked, rendered for in-browser viewing. |
GenAI_Project_Books_Scored.xlsx |
The benchmark (live). The editable/sortable workbook behind BENCHMARK.md — section scores, rationale, and a 24-idea backlog tagged by category. |
- Project Overview — problem, use case, success + error tolerance, the golden-vs-table-stakes call, stakeholders & skills.
- Value & Business Case — value proposition, cost model, value tree (use case → KPI → $), KPIs.
- Data — sources & ownership, quality & gaps, privacy/PII, preparation.
- Solution Design — approach + model-strategy decision (RAG / fine-tune / train) with cost bands, architecture, components, agent flow & routing.
- Build & Engineering — repo, secrets hygiene, eval + red-team harness with thresholds, model-promotion gate, tests.
- Responsible AI & Governance — guardrails, the quality/harm/honesty/cost/latency metric set, RAI dimensions checklist, compliance (ISO 42001), approvals / review board.
- Delivery & Operations — lifecycle stage gates, CI/CD + rollback + IaC baseline, monitoring, ownership, operating model.
- Roadmap & Risks — milestones, taxonomy-screened risk register, maturity benchmark, decisions log.
- Copy the template. Duplicate
GenAI_Project_Book_Template.mdinto your project. - Pick your tier. Open the Scalability Guide, find your tier from business size + risk (the higher of the two wins; High risk always forces the top tier). It tells you what to keep, combine, or skip.
- Fill it in. Replace the italic prompts. Link out to detailed docs rather than pasting them.
- Score it. Run it through the Scoring Rubric before go/no-go. Every applicable gate must clear.
- (Optional) Let an agent draft it. Hand the AI Authoring Guide + your context to an AI agent to produce a first draft.
| Tier | Who | Shape |
|---|---|---|
| Tier 1 — Lite | Small org, low–moderate risk; pilots & internal tools | A 2-page, 5-block brief — every gate, none of the ceremony |
| Tier 2 — Standard | Mid-market, or any production system at moderate risk | The full 8 sections, lightly |
| Tier 3 — Enterprise | Large org, or any size with high risk / regulated data / low error tolerance | Every sub-section + appendices, all gates live |
The floor never moves: problem & success, cost model, PII, secrets, guardrails, and a basic risk register are mandatory at every tier. You can't scale your way out of a gate that applies.
A project can score well overall and still be "Not ready" if it fails any applicable gate:
Cost model · Privacy / PII · Secrets & config · Model-promotion gate · Guardrails & safety · Compliance & security · Approvals / review board · Deployment & rollback
Anyone shipping with generative AI who wants a plan they can defend — product leads, ML/AI engineers, architects, founders, and the people who have to sign off on the risk. Fork it, rename it, cut what you don't need, and make it yours.
Issues and PRs welcome — new ideas, better criteria, worked examples for specific industries. See CONTRIBUTING.md.
Synthesized from 15 public GenAI project books. Strategy/governance ideas drawn from BCG, McKinsey (Rewired), Deloitte (Trustworthy AI), KPMG (Trusted AI), Accenture, and Microsoft's GenAI lifecycle; engineering/LLMOps patterns from aws-samples, HeyNina101, AmineDjeghri, langchain-ai, callmesora, Azure GenAIOps, Shekswess, fmind, and HimJoe's AI Project Canvas. See the benchmark workbook for full attributions.
MIT — use it, fork it, ship it.