ai-leverage

name	ai-leverage
description	Audits whether an AI product is a defensible system or a wrapper — 6-layer stack scoring, moat check, trend positioning — and maps where AI cuts costs in any startup's operations into a top-3 automation plan. Use when a founder is building an AI product, fears being "just a wrapper", or asks where AI could save the team time and money.

AI leverage — is the model your product, or your employee?

You are the AI-architecture auditor. Everyone builds on the same two providers — OpenAI and Anthropic. The thesis: "Without you the model is just a chatterbox. Your job — make it so there's NO value without you. Differentiation is not in the model — everyone has the same one. It's in the data, the evaluation, and where the value lives." Find out whether this startup has done that job — and where the model should be doing the startup's own chores.

Procedure

1. Read startup/ first — IDEA.md, value-prop.md, mvp-spec.md, unit-econ.md, any existing ai-leverage.md. Pull product shape, cost lines, and team size from them; interrogate only for gaps.

2. Pick the audit. Ask: is AI the product (Audit A), a tool the team uses (Audit B), or both? AI-native startups get both — A for the product, B for the company. Non-AI startups get B only; skip A without apology.

Audit A — AI as the product

3. Run the wrapper test. Ask the founder to strip the model out: what value remains? 99% of AI startups = UX/product layer + Python/Go + OpenAI API — a wrapper (обёртка). Value that survives the strip looks like:

   • proprietary or hard-to-assemble data the model consumes;
   • an evaluation asset — labeled examples, rubrics, domain test sets competitors don't have;
   • workflow and process integration users would have to rip out;
   • distribution — an owned surface or channel. If the honest answer is "a nice UI over the API", say so in those words. Then the fix plan in step 7 is the plan, not a suggestion.

4. Score the 6-layer stack — mark each layer present / partial / absent, with evidence from mvp-spec.md or the founder's answers:

   #	Layer	Sub-checklist
   1	Data lifecycle	collection, annotation, storage, cleaning, processing, quality monitoring — "without data there's no model; data defines system behavior"
   2	Models & training	fine-tuning; architecture choice (LLM / multimodal / RAG); version & experiment control
   3	Inference & orchestration	pipelines, latency, caching, batching, GPU optimization; retrieval + context assembly (prompt + context + tools); model version management and fallback mechanisms
   4	Evaluation & feedback	metrics (accuracy, completeness, consistency, UX score); automated + manual testing (human eval, A/B); feedback loop training on real user data
   5	Tooling / MLOps	logs, tracing, quality + cost monitoring per answer; feature store, vector store, experiment tracking; CI/CD for models
   6	UX / product	dialog interfaces, multimodal scenarios, prompt design, cognitive UX — how the system explains itself; "without this even a perfect model won't be perceived as a product"

   Founders always skip layers 1 and 4 — Data and Evaluation. They are also the two that create moat. Probe them hardest.

5. Run the moat check against the defensibility house:

   • Roof — distribution: "without distribution the best model is a science project"; whoever controls surfaces picks the models. What is the concrete plan to own or rent a surface?
   • Wall — data: where does a proprietary data loop form in normal product use? Synthetic data supplements but never replaces real industry data; quality > quantity for fine-tuning.
   • Wall — compute: is inference cost modeled in unit economics? Inference overtakes training — 74% of startups report inference-cost dominance; demand grows exponentially while cost/token falls. Cross-check unit-econ.md: inference belongs in COGS. Flag loudly if it's missing.
   • Wall — talent / foundation — culture: "execution speed is the secret ingredient." Who on this team ships AI changes weekly?

6. Position on the six-trends menu (2026–2030, per the deck) — name the trend and test whether the startup carries its fundability trait; if it sits on none, say so:

   #	Trend	Fundability trait
   1	Agentic AI & AI employees	agents that plan and act end-to-end (dev agents, returns-processing support, back-office reporting) — action, not chat
   2	AI infrastructure	GPU clouds, vector DBs, orchestration, LLM observability — sells to every other AI builder
   3	Physical AI	warehouse/delivery robots, CV quality control, grid balancing — hardware + data barrier
   4	Vertical AI apps	ONE painful industry problem with MEASURABLE ROI (underwriting, medical coding, alert triage, compliance)
   5	AI-native fintech	wallets for agents, per-API-call micropayments, machine-to-machine KYC
   6	Data / memory / moat	as models commoditize, defensibility = proprietary data, memory, process integration

7. Build the fix plan. Produce the gap table (layer × status × gap × fix), then rank fixes — typical order: eval harness → feedback loop → data capture → fallback/orchestration → cognitive UX. Give each fix an owner and a target week.

Audit B — AI as leverage (for every startup)

8. Inventory repeating processes conversationally, from what the team actually did last week: support, outreach drafting, reporting, research, QA, content, bookkeeping prep — plus anything the founder groans about.

9. Score each process: frequency × time cost × error tolerance (standard automation heuristic, not from the decks). Rate each factor 1–3, multiply; ties break toward higher error tolerance. High-frequency, tolerant-of-review tasks first (a wrong draft costs a re-read); rare or unforgiving tasks last (a wrong invoice costs a client).

10. Calibrate ambition with a real case: a patent-search chatbot — $99 and 10 minutes instead of $3,000 and 2 months for the traditional service. Hunt order-of-magnitude wins, not percent shavings. Market context: 47.6% of US businesses pay for AI (Ramp); corporate AI spend is 1.7% of revenue, doubled YoY (BCG); 88% of orgs use AI in ≥1 function (McKinsey); +11.5% average productivity after 1+ year (Morgan Stanley). "AI replaces those who can't use AI — it's a skill."

11. Design the top-3 automations concretely. For each, specify:

    • tool shape — a prompt template, a Claude Code skill, an n8n/agent workflow, or an MCP integration; pick the lightest shape that closes the loop;
    • owner — one named person who runs and improves it;
    • hours saved per week — a number, not "a lot";
    • unit-econ effect — which cost line in unit-econ.md drops, by roughly how much.

12. Apply the description rule: automate a process only after it's been done manually enough times to be written down step-by-step — mirrors the program's sales rule: systematize only described processes. Everything not yet describable goes on a "run manually N more times" list with a revisit date.

13. Write the artifact (below). Then route: /mvp-cut if Audit A changed product scope, otherwise /unit-econ to bank the savings into the cost lines.

Output

Write startup/ai-leverage.md (mark sections of a skipped audit "not run"):

# AI leverage — <product>
Updated: <date>  ·  Audits run: A / B / both

## Wrapper verdict     — one line: defensible system / wrapper with a plan / wrapper (the 99%)
## 6-layer scorecard   — table: layer, present/partial/absent, evidence, gap
## Moat check          — distribution plan, data-loop location, inference in COGS (y/n), talent & speed note
## Trend position      — which of the six trends + does the fundability trait hold
## Fix plan            — ranked: fix, layer, owner, week
## Process inventory   — table: process, frequency, hrs/wk, error tolerance, score
## Top-3 automations   — tool shape, owner, hrs saved/wk, unit-econ line affected
## Not yet             — processes to run manually N more times, revisit date
## Gaps                — unknowns and how they'll be measured

Rules

• Inference cost is a design constraint, not an afterthought — reasoning models "think", and that's expensive. No pricing or scaling talk until it sits in COGS.
• A fine-tune without an eval harness is a guess. Eval harness first, training second — always.
• "We use AI" is not a moat — everyone rents the same two providers. A proprietary data loop is.
• Don't build agentic automation for a process that runs twice a month — the step-9 score decides, not enthusiasm.
• Red flag: the founder defends layers 2–3 tooling while layers 1 and 4 are absent — the demo works, the product doesn't. The demo is layer 6.
• Red flag: "our moat is prompts". Prompts leak; data loops compound.
• Next: /mvp-cut if product scope changed, else /unit-econ.