Spend half the budget on the part nobody demos.

There is a number buried in the research on companies that actually scaled AI, and it reads like a budget line, because it is one: nearly 90% of the successful cohort spent more than half of their analytics budgets on adoption, workflow redesign, communication, training, while only 23% of everyone else did (McKinsey, in Harvard Business Review, 2019). Not on models. Not on infrastructure. On the unglamorous work of changing how people operate. The finding is old enough to predate the current wave entirely, which makes its continued neglect harder to excuse.

The pattern persists because the two halves of the program have opposite sales dynamics. The technology demos beautifully in a forty-five-minute meeting. Role redesign, certification, and operating cadence demo never. So procurement gravitates toward the half that can be seen, and the budget quietly takes the shape of the demo rather than the shape of the outcome.

Budget structure predicts outcome

Read the 90%-versus-23% split as a diagnostic, because that is what it is. Before a single agent is deployed, the allocation already announces which cohort an organization has joined: the scalers, who treat the program as an operating-model change with software inside it, or the rest, who treat it as a software purchase with change management sprinkled on top. The same research found the broader practice gap to be brutal, only 8% of firms engage in the core practices that widespread adoption requires (McKinsey, in Harvard Business Review, 2019).

Eight percent is a remarkable figure to sit beside another one: 95% of corporate AI pilots deliver no measurable return (MIT, 2025). The two numbers are not a coincidence; they are the same fact measured at different ends of the pipeline. Organizations skip the practices, then record the absence of results, then blame the model. The budget told the story before the pilot started.

The unglamorous line items decide the return

What does the adoption half of the budget actually buy? Four things, none of which will ever headline a vendor keynote. Certification: every operator who works alongside the fleet is trained and assessed, not merely licensed. Operating cadence: a weekly rhythm in which agent output is reviewed, corrected, and folded back into how the function runs. Role redesign: job descriptions rewritten so the agent's work has an owner and the saved hours have a destination other than re-absorption. Pairing: each named agent attached to a senior person whose judgment it feeds and whose accountability it inherits.

Each line item looks like overhead in the spreadsheet, and each is the mechanism by which capability becomes P&L. A drafting tool with no cadence produces drafts nobody reviews; an agent with no paired owner produces output nobody trusts; a function with no role redesign refills every saved hour within a quarter. The technology supplies potential at roughly constant quality across the market now, every competitor can buy the same models on the same day at the same price. The adoption work is the entire remaining variable, the only part of the program a rival cannot order from the same catalog, and that is why it deserves the majority of the money.

Eighteen to thirty-six months, mostly headwind

The duration data explains why the soft line items cannot be deferred. Typical AI transformations run 18 to 36 months end to end (McKinsey, in Harvard Business Review, 2019), and across transformations generally, 96% encounter challenges severe enough to derail the entire program (Harvard Business Review, August 2024). A program that long, with a base rate that hostile, is not won by the quality of its software. It is won by whether the organization around the software holds, whether the cadence survives the third reorganization, whether the certified operators stay certified, whether anyone still owns the outcome in month twenty.

Milton's answer is to productize precisely the part nobody demos. The ladder runs assessment, lighthouse, rollout, managed operations, and the lighthouse stage carries a 30–60% function-level improvement design target with a credit policy if the target is missed (internal operating record). A credit policy concentrates the mind: it is only rational to offer one if the adoption machinery, not the model, is what you control. That is the bet, stated in commercial terms.

The technology is the cheapest part. Budget accordingly.

Here is the planning heuristic for an operator building next year's program: take the proposed AI line, and check what fraction survives if every software invoice is removed. If the answer is under half, the budget is structurally aligned with the 23%, the cohort that buys capability and never converts it. The successful pattern is the inverted one, where workflow redesign, training, cadence, and ownership consume the majority, and the technology rides along as the cheapest, most replaceable component in the stack.

Nobody will demo that budget. No screenshot exists of an operating cadence, and no keynote has ever featured a rewritten job description. But the research has been consistent since 2019, across two distinct technology generations: the money spent on the invisible half is the money that decides whether the visible half ever shows up in the numbers (McKinsey, in Harvard Business Review, 2019). Spend half the budget on the part nobody demos. It is the only part the outcome has ever depended on.