The firm's AI operating model: build, buy, or wrap

McKinsey's Lilli - launched 2023, named for Lillian Dombrowski, the firm's first professional woman hire - is the most instructive case because it is the least about the model. Lilli searches 40+ internal knowledge sources and 100k+ documents: decades of engagements, methods, and expert profiles nobody else has. Roughly 72-75% of the firm's ~43,000 people use it, at 500k+ prompts a month, and McKinsey claims about 30% time saved on information gathering - their claim, and label it that way when you repeat it.

Strip the scale away and the design decision remains: McKinsey did not build a better model. It pointed a model at the one asset no competitor can copy - its own accumulated knowledge. That is the transferable lesson, and it transfers down to a twelve-person boutique: the question is never "can we build Lilli?" but "what is our version of the thing Lilli searches?"

Every number in this part came from the firm itself or from press coverage of the firm's statements. That does not make them false - it makes them marketing-shaped. BCG's ~90% employee usage and its Deckster and GENE tools are press-reported BCG statements. EY's US$1.4B and KPMG's US$2B Microsoft commitment are announced investment figures, not audited outcomes. None of these firms publishes the number you actually want: profit impact per dollar spent. Read announcements the way a4 taught you to read an AI draft - plausible, fluent, and unverified until someone accountable shows the source. The discipline you apply to their numbers is the discipline clients will apply to yours.

The assemble tier tempts every firm with a file server full of old decks, and the temptation is legitimate - retrieval over your own past work is exactly what made Lilli valuable, and the knowledge base matters far more than the model you put in front of it. But notice what you are proposing: making two hundred past engagements searchable by everyone in the firm. Client A's confidential pricing study is now one query away from the team serving Client A's competitor.

So the prerequisite is not a vector database - it is a sentence: which past-client material may enter the knowledge base, under what anonymization, and who decides. Write that rule before the pilot, not after the incident. It is the same confidentiality rail from a3, pointed inward: the Samsung lesson applies to your own knowledge base just as much as to a vendor's chat window. A firm that cannot write this rule in a week is not ready for tier 2, and that is useful information, not failure.

Run the copyability test on everything in your operating model. Prompt libraries: copyable in an afternoon - any competitor with the same tools converges on similar prompts. Policies: copyable - and should be, since good policy is not a secret. Enterprise subscriptions: identical to everyone else's by definition. Now the other column: your 200 past engagements are not copyable. Neither are your benchmarks, your sector relationships, or the pattern library in your partners' heads. The moat was never the AI layer; it is what the AI layer gets pointed at. Which is why tier 2 - done with the governance rule written first - is where a boutique's durable advantage lives, and why tier 1 alone, however necessary, differentiates nobody.

Three rituals carry adoption. Usage reviews: look at whether the tools are actually used, by team, monthly - not to punish, but because silence tells you where the playbook does not fit the work. Playbook rituals: the prompt library is a living document with a named owner; teams add what worked and retire what did not, or it dies as a PDF. Celebrating catches: when a reviewer catches an AI-invented fact before it reaches a client, praise it publicly - that is the a4 review discipline paying out, and what gets celebrated gets repeated.

And the budget truth that reframes every pricing conversation with your own partners: tool spend is trivial next to partner review time. The subscriptions for an entire boutique cost less than a few days of one partner's time. If the operating-model debate is consuming partner hours arguing about tool costs, the debate itself is the expensive part. Decide, write it down, put a review date on it, and move.