Proposal and scoping

Hand AI the Cadence brief and it will produce a competent context section, clean boilerplate, a sensible document structure, and a first-pass approach - the parts that used to eat an analyst's evening. Take those gains; session 1 called refusing them malpractice. But three parts of the proposal sit outside the frontier, and they happen to be the three parts the client actually judges:

What to EXCLUDE from scope. AI expands scope by default - ask it for an approach and it will cheerfully include everything adjacent. Deciding that this engagement answers enterprise-tier-or-not and does NOT redesign the product and does NOT renegotiate infrastructure costs is a judgment about what six weeks can honestly deliver. The fee. A number that prices your accountability and the client relationship - no model has that context. The promise. The terms section is what the firm signs; session 1's liability rail says every commitment reaching a client has a human owner. That is nowhere more literal than on the signature page.

The exclusions deserve one more beat, because they are the real craft. An unwritten exclusion is an argument waiting for week 4, when the client "assumed pricing redesign was included." A written one costs two sentences in week 1 and settles the argument before it exists. Your drafting exercise today requires three of them, in writing.

Session 1's study had a second-order finding worth re-reading in a proposal context: AI-assisted output homogenized - answers converged on similar ideas. A proposal is a competitive document; if three firms pitch the same client with AI-drafted approaches, the context sections will be interchangeable. What cannot converge is the part drawn from judgment: the one question you chose, the exclusions you dared to write, the specific fact from THEIR business (62% of trial teams hitting a seat cap says more about Cadence than any market overview). Differentiation lives exactly where delegation stops.

The tree reads top-down: "Cadence should launch an enterprise tier" holds IF demand exists, AND the economics favor it, AND the market window is open. Each branch is checkable against evidence you can name today - the seat-cap number and the 210 procurement accounts for demand, the 94-vs-71 retention split and the ARPU math for economics, rival pricing at $38-45 for the window. And each branch spawns a workstream, which is how the tree becomes the approach section: the plan is not "research the market," it is "test branch two by week 4."

AI's role on the tree is precise and valuable: it generates counter-hypotheses and missing branches. "What would make this false?" is a superb prompt - it will suggest that the 210 accounts entered procurement to satisfy security teams rather than to buy, or that individual-plan churn is driving the retention gap rather than enterprise stickiness. Some suggestions will be noise; the two that sting are worth the whole exchange. What AI never does is decide the tree - stating the likely answer is the engagement's central act of judgment, and it belongs to whoever signs the proposal.

The discipline is stating the likely answer so the work can attack it, not defending the answer against the work. Each branch carries an implicit kill condition: if the procurement pipeline turns out to be security theater, branch one falls, and the recommendation moves. The tree you draft today gets rebuilt with real findings in b4, becomes the storyline's skeleton in b5, and gets scored in the readout. Write it in pencil; test it in ink.