McKinsey + OpenAI (Lilli): When the Consulting Firm Owns the Memory, Not the Model
Lilli Adoption
Peak
0 (Jun 2023)
Now
70%+ of firm weekly
Layer Scoring
Sublayer Impact Map
Which of the 50 sublayers this case actually touches, and at what magnitude.
Intelligence Cube · 2D
Footprint across Functions × Verticals × Layers, the three axes that determine structural fate.
Layers × Verticals
18 cells · 6×3
Layers × Functions
24 cells · 6×4
Two 2D projections of the Intelligence Cube (Functions × Verticals × Layers). Filled cells = this move occupies that intersection.
Timeline
Mid-2023
Lilli launches internally, built on OpenAI, grounded in McKinsey's corpus.
Late 2023
Adoption ramps. Partners start citing Lilli output in client decks.
Mid-2024
70%+ of the firm uses Lilli weekly. Becomes standard tooling.
2025
Consulting firms emerge as some of the largest enterprise LLM customers, opposite of the predicted disruption.
2026
The model layer keeps commoditizing. McKinsey's L1+L8 advantage compounds. Pure-play AI strategy startups struggle to find buyers.
- Who Wins
- McKinsey, BCG, Bain. L1 (decades of proprietary client work) and L8 (compounding institutional memory), the two layers AI can't reproduce.
- OpenAI / Anthropic. Consulting firms became some of their largest enterprise customers, not their competitors.
- Any firm with deep proprietary archives. Law firms, accounting firms, hospitals, banks, the L1+L8 pattern generalizes.
- Who Loses
- Pure-play AI strategy startups. No L1, no L8, no client trust. The model is the cheapest layer to own, the others are the business.
- Junior consulting headcount (slower growth). Drafting work compresses. The pyramid narrows. Entry-level path changes shape.
- The 'AI will kill consulting' thesis. Falsified in real time. The opposite happened, consulting became one of the biggest beneficiaries.
- Steelman: The Counter-Thesis
The counter is that Lilli's moat is overstated because the *actual* output of consulting (the synthesized 60-page deck, the executive narrative, the client-relationship judgment) was always the value, not the underlying corpus. If frontier models continue to improve at long-form reasoning and synthesis, a well-prompted GPT-6 with public data may produce a McKinsey-grade strategy doc for a tenth the cost. McKinsey's corpus is large, but most of it is dated or sector-specific in ways that a generalist model can fluently approximate. The honest read: L1 protects them for 3–5 years; the open question is whether L8 (the compounding loop) can build a durable advantage before model capability closes the gap on the synthesis layer itself.
The most misread AI story in professional services. Everyone asked "will ChatGPT kill McKinsey?" The structural answer was always no, and Lilli is the proof.
The setup. In mid-2023 McKinsey shipped Lilli, an internal generative AI assistant built on OpenAI's models but grounded in McKinsey's own corpus: 100,000+ proprietary documents, 70 years of consulting studies, frameworks, interview notes, expert profiles. Within a year, the majority of the firm uses it weekly.
The structural read:
• L2 (model): rented from OpenAI. Commoditizing fast. McKinsey pays nothing to defend it.
• L1 (proprietary data): 70 years of client work, expert memos, sector benchmarks. Cannot be reproduced by any model. This is the moat.
• L6 (orchestration): retrieval over the corpus, citation, expert routing. McKinsey-built.
• L8 (institutional memory): every engagement adds to Lilli. Compounding. Every consultant's output becomes future fuel.
Law I, intelligence commoditizes downward. OpenAI's model is now interchangeable with Anthropic's or Google's at this task. McKinsey can swap the L2 underneath Lilli in a weekend. The L1 and L8 above it are untouched.
Law III, value migrates to the scarcest layer. The scarce thing is not the ability to write a strategy memo. It's the 70 years of which strategies worked for which clients in which sector cycles. That data only exists inside McKinsey.
The inverted lesson for everyone else. Most enterprises panicked about "AI displacing consulting." The opposite happened: consulting firms became some of the largest OpenAI/Anthropic customers, because they had the L1 and L8 to make the model valuable. The model layer is a tool. The data and memory layers are the business.
What this means for your firm. If your competitive position is "we know things and we remember things," AI is an amplifier, not a threat, provided you own L1 and L8. If your position is "we can write good documents," the model now reproduces most of that work.
Public reporting; figures approximate.
What This Means for You
Product Leader
Map your product to the layers it actually owns vs. rents. The rented ones are where the counter-move work belongs.
Investor
Underwrite layer ownership, not feature count. The Cube footprint is the moat.
Operator
Audit your stack against Supply Chain of Intelligence. Anything sitting only at L7 is the layer to watch.
Anand Arivukkarasu
Ex-Meta product leader. Creator of Supply Chain of Intelligence™. Writes about where AI value accrues, and who can fire your product. LinkedIn
Get the next teardown in your inbox.
One issue when something structurally important happens, usually weekly. No spam, no filler, unsubscribe anytime.
Worth sharing? Pull-quote: "L1 + L8 OVER L2"