Speaking on the Possible podcast with Reid Hoffman in 2026, Satya Nadella made an argument that consulting firm partners should read twice: a company's real intellectual property is not merely its data — it is the proprietary learning system built from its workflows, domain expertise and accumulated judgement.
"What is the tacit knowledge of an enterprise or a firm? It's the unique ways that you are able to operate, pass judgment, have taste. Now when it comes to AI, the model is able to extract that tacit knowledge through human trajectories and encode it in weights. If you leak it, it's a one-way door. You're done."
Satya Nadella, CEO, Microsoft — Possible podcast with Reid Hoffman, 2026
He was making this point in the context of large enterprises. But it applies with even greater force to boutique consulting firms, where the knowledge that matters most — the frameworks that work, the client dynamics that shape what's possible, the hypotheses tested and why — almost never makes it into any system at all.
Two types of AI adoption — only one of which builds anything
There are two ways to use AI in a consulting firm, and they lead to fundamentally different outcomes.
The first is task-level productivity: AI to write faster, summarise quicker, format cleaner. The firm saves time. But at the end of each engagement, what has it gained? Nothing the AI can remember. Nothing that informs the next client. The starting point resets to zero.
The second is institutional intelligence: AI that accumulates knowledge — methodologies, client dynamics, patterns across sectors — into a private learning system that grows richer with every engagement. The output is not just better work today. It is a compounding asset.
The BCG finding captures exactly this problem. Most firms are capturing the time savings — the 42%. They are not capturing the institutional intelligence opportunity. The value leaks out because the AI sessions are stateless, the context resets, and each engagement begins as if the previous one never happened.
What knowledge actually leaves when a consultant does
A senior consultant leaves. They were the primary relationship holder on three accounts. Three years of context: the client's internal politics, the initiatives that failed, the language that resonates with the decision-maker, the competitive dynamics. None of this is written down. All of it lived in one person's head.
The new person starts from a standing start. The client notices. Some clients don't wait.
McKinsey can report these numbers because they built Lilli as an institutional system — not individual productivity tooling. The knowledge compounds because it is captured at the firm level, not at the user level. That is the structural difference.
Compounding memory — and why it only works cited
The AI capability that changes this is not generation. It is retention, accumulation, and attribution.
Memory that accumulates without source attribution creates its own professional risk. Over time, you cannot distinguish between what your AI learned from your client's documents and what it is asserting from its general training. Cited memory resolves this: every layer of accumulated context is attributed to its source — the specific document, the engagement it belongs to, or an explicit label where the insight draws on model training rather than your materials.
Model-agnostic memory: the condition that makes the IP yours
Consumer AI tools have memory features. But they store that memory inside the model provider's interface. Switch models — the memory doesn't travel. Provider restructures plans — the memory is held hostage.
Worse: many consumer AI tools store memory in environments where the underlying model is trained on that data. Your client materials, your frameworks, your proprietary thinking may be informing a public model used by your competitors. This is the default unless you have a specific contractual exclusion.
"The model is able to extract that tacit knowledge through human trajectories and encode it in weights. If you look at how the model companies learn, they're essentially setting up these gyms with rewards which are employing employees who worked at your company previously. That should tell you everything about what should not be happening."
Satya Nadella, CEO, Microsoft — Possible podcast with Reid Hoffman, 2026
The learning system that creates durable competitive advantage must be two things simultaneously: cited, so every insight is verifiable; and model-agnostic, so it persists regardless of which model handled each task. If the memory belongs to your AI vendor, it is not your IP. It is theirs.
The compounding return — measured in years
After 12 months: a structured, verifiable record of every engagement, every client dynamic, every framework tested — with every insight cited to its source. New hires access the firm's accumulated intelligence on day one instead of spending six months rebuilding context.
After 3 years: cross-client pattern recognition that currently exists only at firms five times the size. Built from your engagements, your methodology, your IP. Fully attributed, fully private, portable across models.
After 5 years: a proprietary learning system that is genuinely difficult for a competitor to replicate, because it has been built from your specific work, cited throughout, and compounding with every engagement your firm takes on.
PAL builds the institutional memory Nadella describes
Four memory layers — session, workstream, project and client. Every response cited to its source. Memory persists across model switches — Claude, GPT, Gemini, Llama, Mistral — so it belongs to your firm, not your AI vendor. Private by architecture.