Resistance Is Rational: The Denominator Under Every AI Mandate

John Cutler’s formula for AI adoption fits on a single line:

(Machine Understanding x Problem Understanding x Practice Evolution) / Social Contract

Three terms above the line, one below. The three above multiply, so any of them at zero takes the whole numerator with it. That much is intuitive to anyone who has watched a rollout stall. The term below the line is the one that rarely appears in a rollout plan at all, and Cutler is blunt about its status: “The denominator is not optional. It is the precondition for everything above the line.”

Read the division as a foundation rather than as arithmetic. Whatever you stack above the line rests on what sits below it.

Six Ways the Numerator Dies

Cutler enumerates six named failure modes, each a specific combination of zeros. Two will be recognizable to anyone who has run an AI pilot.

Machine Understanding present, Problem Understanding absent, Practice Evolution present produces “Fast iteration in the wrong direction.” The team knows what the model can do, has rewired how it works, and ships confidently against a problem nobody framed correctly.

Invert the pattern and you get the other one. Machine Understanding absent, Problem Understanding present, Practice Evolution absent produces “You know it isn’t working, but have no idea why.” The team watches the outcome degrade with no vocabulary for the mechanism.

The remaining four are permutations of the same arithmetic. Their value is diagnostic. When adoption stalls, the useful question is which term sits at zero, and “more training” answers only one of them. Cutler’s read on the standard corporate response is unsentimental: “Everyone must use AI is a Practice Evolution mandate that assumes Machine Understanding and Problem Understanding will take care of themselves. They won’t.”

What Sits Below the Line

Cutler defines Social Contract as “the organizational commitment that improvement gains will be used in line with a shared agreement, not against the people who generated them.”

Most adoption programs treat holdouts as a training problem. We have argued a version of that ourselves. Our piece on developer identity framed resistance as identity work to be worked through, and mandates meeting mental models treated the mismatch as something a better rollout could repair. Cutler inverts it: “Without that contract, you don’t get resistance out of ignorance. You get resistance out of rational self-preservation.” Resistance under a broken contract is an accurate reading of thin evidence, not a psychological obstacle to be trained away.

The employee doing that reading is running a small underwriting exercise. The gains are measurable and near-term. Who captures them is unstated. The downside sits entirely on one side of the table. An engineer who automates 30% of their own work and cannot name what happens to the remaining 70% holds an uninsured position. Declining to automate is a hedge, and it is a cheap one.

The Experiment Already Ran, Under Another Name

Design systems ran this scenario before AI mandates arrived, and someone published the results.

Karolina Szczur’s case against design system federation is that record. Federated governance distributes system ownership across product teams instead of concentrating it in a dedicated one. Per the ZeroHeight Design System Report she cites, it is used by 13% of teams, up from 9% a year prior. Adoption rose. Satisfaction moved the other way: the How We Document report “showed the highest level of dissatisfaction among the federated cohort.”

The context of that rise matters more than the number. Szczur cites layoffs.fyi for 226 companies letting over 120,000 employees go this year, and observes that “the skills that are the backbone of design systems (and lasting product design) are now first to go.” Federation arrived as the cost answer to losing the dedicated team, dressed in the language of shared ownership. The people who lived through it read the framing correctly.

What federation produced is the accountability version of a broken denominator. Szczur quotes Albert Bandura: “Where everyone is responsible, no one is really responsible.” Her worked example of how that shows up in an artifact is a system carrying 1,500 components, duplication accumulating without reuse. The governance model said everyone owned the system. In practice nobody held enough standing to say no to a component.

Her line about contribution volume lands differently now that a share of the contributors are machines: “Whether the output is erroneous, its quality acceptable, or it fits into the system is another story.” Contribution volume was never the scarce resource. Review capacity is, and federation spreads review authority thin enough that nobody holds enough of it to exercise.

The Rule That Transfers

You cannot distribute accountability and then act surprised when nobody exercises it.

That is the design-system finding, and it has the same shape as the mandate finding. A mandate distributes the obligation to adopt while leaving the question of who captures the gains centralized and unanswered. Federation distributed the obligation to maintain while leaving no one with the authority to refuse bad work. Both models moved the duty downward and kept the decision rights somewhere else. We traced the same mechanism in Kubernetes contribution governance and in how design systems documented their own AI governance. It repeats across domains because the mechanism is organizational, and organizations are where the denominator lives.

Write the Denominator Down Before the Next Mandate

Before the next “everyone must use AI” email goes out, write one paragraph and circulate it alongside the mandate. It should answer four questions in language a skeptical senior engineer would accept:

Where do the recovered hours go? If a team’s cycle time drops 25%, name what fills the capacity. “More output at the same headcount” is a legitimate answer. Silence is not an answer.

Who is accountable for the output? Name a role, not a committee. If review authority is already federated across every team, you have Bandura’s problem before you have an AI problem.

What is the headcount commitment, and through when? A dated commitment beats a warm assurance. If you cannot make one, say that instead, and say what you can commit to.

What happens to someone whose measured productivity drops during the learning curve? Every adoption curve has a trough. Whoever writes the performance review during that trough is the real policy.

If those four have no answers, the resistance you are about to meet is your own policy read back to you. Run the numerator diagnostics after the contract is written, in that order. A team that trusts the contract and still stalls is stalling on Machine Understanding or Problem Understanding, and both are solvable with work you already know how to do. A team that distrusts the contract will never give you a clean read on anything above the line.

---

This analysis synthesizes TBM 431: The Denominator That Matters (The Beautiful Mess, July 2026), Against design system federation (Karolina Szczur, July 2026).

Victorino Group helps leadership teams write the social contract that AI adoption programs assume and rarely state. Let’s talk.