A fair critique of the framework is that it tells you whether your product is durable but not what to ship on Monday. This page closes that gap: a layer-by-layer roadmap translator, a build-vs-rent decision tree, and a discovery question bank for the first week of any new AI product.
- Part 1
Layer → Roadmap translator
For each layer, the exposure if you ignore it and three concrete moves you can add to next quarter's PRD.
L−1 / L0 Resources & Infra
Exposure: You depend on capacity that can be priced or rationed against you.
Lock in 12–24 month compute commitments before launch. Treat GPU as a supply-chain line item, not a cloud cost.
Multi-cloud where the contract size justifies it; otherwise pick one and negotiate volume.
Quantify a 5x token-price increase scenario in the PRD. If gross margin goes negative, you have an L0 exposure to fix.
L1 Data
Exposure: Your differentiation depends on data anyone can scrape.
Identify the one data type you can collect that the model layer cannot license. Make collection a product feature, not a backend job.
Instrument outcome data (L1d), what worked, what didn't, with attribution. This is the most defensible sub-layer.
Negotiate exclusive feeds with design partners before product-market fit, not after.
L2 Models
Exposure: You are renting intelligence from a vendor that can deprecate, reprice, or compete with you.
Abstract model calls behind a router from day one. Never hard-code a vendor.
Fine-tune or post-train only on the slice where generic models measurably underperform, not as a vanity moat.
Build evals before you ship features. Without evals you cannot detect when the next frontier release deletes your edge.
L3 Gatekeeping
Exposure: Your buyer is regulated, audited, or risk-averse, and you have not earned the gate.
Start SOC 2 / ISO / HIPAA / industry-specific compliance in the first 90 days. It is a 9-month process; do not start it in month 9.
Treat editorial / safety review as a product surface, not a compliance burden, it is the layer enterprise buyers actually evaluate.
Publish a public trust page with audit dates, incident history, and data handling. This is sales infrastructure.
L4 Access
Exposure: You sit on top of platforms that own the distribution and the permission model.
Identify the one system of record your customer cannot move off of. Integrate there first, deepest, two-way.
Adopt agent protocols (MCP, etc.) early, they will determine who agents call by default in 2026.
Ship an Identity & Provenance story (L4e) before scaled deployment, enterprises will block agents without it.
L5 Execution
Exposure: You sell 'doing the work' but the work is shallow enough to be absorbed.
Pick one workflow and go three levels deeper than any horizontal model can. Depth is the moat; breadth is the trap.
Codify the decision frameworks (L5b) and playbooks (L5d) of senior practitioners into the product, that is what compounds.
Measure outcome quality (jobs completed, hours saved, revenue produced) not capability (tasks attempted). Outcome data is the moat.
L6 Orchestration
Exposure: Your product is a workflow engine in a category being absorbed into the model layer.
Do not sell orchestration as a product. Sell it as the invisible engine inside a vertical L5 product.
Invest disproportionately in human-in-the-loop (L6b), that is the sub-layer enterprises will pay for and platforms will not bundle.
Build runtime assurance and learning loops (L6e), orchestration without measurement is a demo, not a product.
L7 Surface
Exposure: You are a beautiful interface on top of someone else's intelligence.
Pair every surface investment with at least one bottom-of-stack investment (L1, L5, L8). Surface alone is Law I.
Treat embedded / copilot (L7c) as more defensible than standalone chat (L7a), the surrounding context is the moat.
Resist the urge to expand modalities. Modality is commodity. Context is moat.
L8 Memory
Exposure: Your product does not remember what your users did yesterday.
Add user/entity profile (L8b) before adding new features. Personalization is table stakes by 2026.
Design for institutional knowledge (L8d) explicitly, every team output should become a reusable artifact.
Aggregated network learning (L8c) is the highest-leverage sub-layer, every customer makes the product better for every other customer.
- Part 2
Build vs Rent, per layer
A decision tree. The defaults are opinionated. Deviate only when you have a structural reason.
Layer
Build
Rent
Note
L−1 / L0
Almost never
Always (AWS, GCP, NVIDIA)
Build only if you are a hyperscaler.
L1 Data
Always for L1b/c/d (proprietary, behavioral, outcome)
L1a public data, L1e synthetic
The defensible sub-layers cannot be rented.
L2 Models
Only above ~$100M ARR or a defensible vertical slice
Foundation models, embeddings, routing
Building before scale is a vanity moat.
L3 Gates
Always for editorial/distribution gates (L3d, L3e)
Compliance tooling (Vanta, Drata)
Trust must be earned in-house; tooling can be bought.
L4 Access
API surface, identity, provenance
Generic integration plumbing
Own the integration with your category's system of record.
L5 Execution
Always, this is your moat
Never the core skill; rent only adjacent capabilities
If you rent L5, you have no company.
L6 Orchestration
The parts that encode your judgment
Generic agent loops, retries, queue infra
Bundle orchestration into the L5 product. Do not sell it.
L7 Surface
Only the embedded/contextual sub-layers (L7c, L7d)
Chat surfaces are nearly free to assemble
Surface is a wrapper on whatever you own below it.
L8 Memory
Always, this is your moat
Short-term session memory only
Memory must accrue inside your product, not a vendor's.
- Part 3
Week-1 discovery questions
Ten questions to run in the first week of any new AI product, alone, with your co-founder, or with your team. If you cannot answer at least eight, you are not ready to write the PRD.
Which two layers do we own today? Be honest. 'We use AI' is not a layer.
When the model layer below us ships our top feature for free, what is the user still paying us for?
Which data do we collect that a competitor with the same budget could not replicate in 12 months?
What would the customer have to rebuild if they left us? Name it in one sentence.
Which sub-layer (1 of 50) is our single biggest moat? Which is our single biggest exposure?
If we paste our product description into the framework, do the layers we claim match the layers we actually invest in?
What is our L8 strategy, what does the product remember tomorrow that it did not yesterday?
Which gatekeeper (regulator, platform, editor, distributor) can shut us off? Are we earning that gate or ignoring it?
If our category gets absorbed in 18 months, which adjacent layer do we sprint into?
Which of the Four Laws is most likely to delete us? Write the post-mortem now.