About the author
Anand Arivukkarasu
Product Leader · Angel Investor · Ex-Meta (Instagram) · AI Product Architect
Mind mapping, shared freely
JTBD, Wardley Maps, and Christensen's work were given away by their authors. This framework is mine, in the same spirit, free to use, cite, argue with, and improve.
More on the disclaimer.
I'm a product leader and architect focused on designing and scaling AI-first products, from 0→1 foundations to the growth systems that hold up at scale. I spent a decade shipping product across consumer and B2B SaaS, including three years at Meta leading product and growth for the Messenger business platform and Instagram monetization surfaces.
After Meta I ran product at Vungle, Pinsight Media, GRIN, and Refersion. Today my primary role is at Ideas2IT as a product leader. Across a decade of building AI-first products, the same question kept surfacing - "what does AI mean for this surface?", and the honest answer was usually "we don't have a framework that explains it." So I started writing one down, on evenings and weekends, as a personal project. This site is that project.
JTBD told us what users want. It never told us whether a model release, a hyperscaler bundle, or a productivity-suite plugin would erase the entire feature six months later. After watching Jasper collapse, Chegg lose 99%, Stack Overflow bleed traffic, and Grammarly get squeezed by Copilot, all predictable structurally, none predictable by demand alone, I started writing this framework down.
Supply Chain of Intelligence™ is that framework. Ten layers, fifty sublayers, four structural laws, one diagnostic cube. It is opinionated, it is portable across categories, and it is free, a give-back to the product community.
I spent two decades building products inside Meta, Vungle, and others, and the frameworks I leaned on most, JTBD, Wardley Maps, the Innovator's Dilemma, were all given away by their authors. This is my contribution back. Use it, cite it, fork it, disagree with it in public. That's the whole point.
I grew up a competitive chess player, a former higher-level junior champion, and that is how I read the AI market. It is a board, not a forecast. Every company on this site is sitting on a square. The framework names the square. The juggernaut still has a move. The Predictions page tracks who saw the fork and who didn't.
I write it for the audience I wish had it when I was building: founders, product leaders, boards, and investors who need to decide "is this layer ours, or are we renting it from someone bigger?" before they commit a roadmap or a check.
How the framework relates to JTBD
JTBD is the length of the need.
The Supply Chain is the depth of the answer.
Credit to Bill Leece (Ex-Google product leader) for the sharpest one-line framing of this. JTBD tells you what job the customer is hiring the product to do. Supply Chain of Intelligence tells you how many layers of the AI stack you have to own to deliver that job durably. Same job can be answered shallow (one layer, fast, fragile) or deep (multiple layers, slow, defensible). Surface looks identical. Fate is not.
| Customer job (JTBD) | Shallow answer (feature) | Deep answer (chain layer) | Why depth wins |
|---|---|---|---|
| "Trust what the AI generated." | An L7 "verifier" widget bolted onto the output. | Bake an L3 Gatekeeping layer into the pipeline, provenance, citation, policy, audit trail. | Any competitor can ship the widget in a weekend. Almost none can ship a gate that regulators and buyers accept. |
| "Let the AI actually do the thing." | A button that opens a confirmation modal. | Own L4 Access + L5 Execution, auth, identity, write-permissions into the system of record. | Execution requires earned trust with the underlying system. That is a contract, not a feature flag. |
| "Remember me. Get smarter for me." | Local chat history in the sidebar. | Build an L8 Memory layer, user, org, and network-level state that compounds across sessions. | The shallow version resets every time the model resets. The deep version becomes switching cost. |
| "Give me a better answer than ChatGPT." | Better prompt template on top of GPT-5. | Combine L1b proprietary data with an L5 execution loop fine-tuned on your domain. | A prompt is reproducible. A data + execution flywheel is not. |
Rule of thumb: when a customer need shows up, do not just ask "what feature ships this?", ask "which layer of the chain do we have to own to make this durable?" Most AI products die because they answered the right job at the wrong depth.
Why I'm really doing this
The AI industry doesn't have a shared vocabulary yet.
That's the gap I'm trying to close.
Walk into any product review, board meeting, or investor call in 2026 and you'll hear the same three words doing all the work: "it's a wrapper", "it's an agent", "it's a copilot". That's not analysis. That's a shrug. Two companies called "agents" can sit on completely different layers of the stack, with completely different defensibility, and the word tells you nothing about which one survives the next platform release.
Other industries solved this decades ago. Semiconductors have a fab → foundry → fabless taxonomy. Cloud has IaaS → PaaS → SaaS. Logistics has Tier 1 / Tier 2 / Tier 3 suppliers. JTBD became durable not because the idea was uniquely brilliant, but because the vocabulary froze: job, hire, fire, functional, emotional, social. Same words everywhere. That's what lets a PM in Berlin and an investor in Singapore actually talk about the same thing.
Generative AI doesn't have that yet. So a product team ships a feature thinking they own a moat, and an investor funds a "platform" thinking it's defensible, and six months later a hyperscaler ships the same capability as a checkbox, because nobody named which layer the work was actually living on.
Supply Chain of Intelligence™ is my attempt at that naming layer. Ten layers. Fifty sublayers. Four laws. One cube. Precise enough that a Series B founder, a corp-dev lead at a hyperscaler, and a PM at a vertical SaaS can all point at the same square on the board and mean the same thing.
Before
"It's just a GPT wrapper."
Tells you nothing about who absorbs whom, or when.
After
"It's an L7 surface on L2 with no L1b, L5d or L8c."
Now everyone in the room knows it gets absorbed by the model layer in two quarters.
After
"It's L1b + L5a + L8d in Legal."
Now you can argue about price, moat, and exit on the same map.
The ask
If you're a founder, PM, or investor, try using the layer notation in your next memo, review, or pitch. Say L5 instead of "the AI doing the work". Say L8 instead of "it remembers stuff". Say L3 instead of "trust and safety". The framework is free, citable, and intentionally portable. Standards only become standards when enough people use them.
Why this framework exists
Six things the AI conversation cannot do without this lens.
Every row below names a sentence you hear in board rooms, pitch decks, and Twitter threads - and the precise instrument the framework gives you to replace it with. If none of these gaps existed, the framework would not need to exist.
| Without the framework | What the framework gives you | Where to see it |
|---|---|---|
| "It's agentic / AI-native / a wrapper."Vague labels. Two companies with the same label have completely different fates. | A precise vocabulary, 10 layers, 50 sublayers, so "L1b moat + L2 MCP + receding L7" replaces "AI-native". Same words, everywhere, every room. | /framework → |
| "Cool product. Great UX. Strong distribution."Descriptive vibes. No way to test if it survives the next platform release. | A diagnostic instrument, Defensibility Audit + Triangle - that a product can actually fail. Which layer creates value, captures margin, is vulnerable to absorption. | Home audit → |
| "It's an agent."Used for chatbots, copilots, automations, tool-callers, seven different architectures, one word. | The Agent Decoder, every "agent" decomposes into L5Execution + L6Orchestration + optional L4Access / L7Surface / L8Memory. Stops the conflation. | /framework → |
| "All AI gets commoditized."Assumes every layer collapses together. It doesn't. | Generation ≠ Verification (Law IV). Explains why L3Gates stays economically durable, compliance, trust, ranking, fiduciary review - even when generation goes to zero. | Law IV → |
| "AI eats SaaS."Binary verdict. Doesn't say which layer compresses, which compounds, which becomes a choke point. | A layer-by-layer value-migration map, surfaces collapse, infrastructure commoditizes, orchestration becomes strategic, memory compounds. Each layer on its own clock. | /predictions → |
| "Our moat is proprietary data."Over-indexes on datasets. Misses workflow position, memory gravity, gate authority. | A structural classification of companies: "L7-heavy with weak L8", "strong L3 gatekeeper", "owns L5a but dependent on external L2". Comparable, citable, falsifiable. | /analysis → |
If you can replace any row with a sharper existing framework, I want to hear it. Challenge a definition →
Career arc
A decade shipping product. The framework comes from the receipts.
Talks & teaching
The work, in other people's rooms.
How to Build AI Products
Product Management Exercises · AI PM Community Session
A framework for designing and building AI-first products.
10 Metrics Every SaaS PM Should Use
Product School webinar
20,000+ views · the metrics talk that established the lens behind this site.
Principles of Product Growth, with case examples
Glorium Technologies
Five core principles, applied to real growth motions.
- Read & follow
The framework is free. Use it, cite it, push back on it.
No services, no consulting, no paid engagements through this site. Questions and corrections welcome on LinkedIn.