The Manifesto Arrives: Auditing McKinsey's 12 Themes

On April 7, 2026, McKinsey published “The AI transformation manifesto” in McKinsey Quarterly. The article is excerpted from the second edition of Rewired and signed by five senior partners. It lists 12 “themes” that, in the authors’ reading, separate companies truly rewired for AI from their peers. Manifestos are public commitments. They deserve to be read carefully.

This is the sixth time we are reading one of these pieces in sequence. The first was McKinsey Measured the Wrong Thing, which showed how executives self-report AI gains that controlled trials cannot reproduce. Then came McKinsey’s Three Diagnoses, which tracked the vocabulary drift from measurement to design to the word “governable” used exactly once. Then the agent factory prescription, which described factories without describing the control surfaces that make a factory an asset rather than a liability. Then the two-paths piece, which framed enterprise architecture as a choice between incremental and comprehensive transformation, with security mentioned zero times.

The manifesto is the best of them. I want to say that plainly before the audit. In the themes that matter most to us, McKinsey has moved further than in any prior piece in the series. That matters. Progress deserves acknowledgment before critique.

What Is New, and Worth Saying So

Three things in the manifesto are genuinely new.

The first is Theme 10: “No trust, no right to deploy AI.” This is the first time in the series that trust, risk, and testing have been named as a full theme rather than buried in a sub-bullet about responsible AI. The authors write that “the excitement for agentic AI may be getting ahead of companies’ ability to manage the more complex risks.” That sentence is the most honest line McKinsey has published on this subject in the last twelve months. It reads like something written by a practitioner who has been in a war room at 2 AM.

The second is Theme 11: “Agentic engineering becomes the next capability to master.” Inside that theme, almost as an aside, the authors list “automating guardrails and controls” as a component of the capability. It is a short phrase. It is also, as far as I can tell, the closest McKinsey has come in print to naming governance infrastructure as a buildable thing rather than a cultural posture.

The third is Theme 6, which introduces the idea of an organization’s “metabolic rate.” Speed is framed not as an aesthetic but as a measurable property of the system, produced by reducing latency from insight to decision and from decision to action. This, too, is new. Metabolic rate is a step toward treating the organization as something you instrument.

Three steps forward. In a tracking series where most chapters moved the line by a single word, three steps is a lot.

What Is Still Missing

Now the audit.

A manifesto is a set of commitments about what the reader should do. For a commitment to become infrastructure, there has to be something you can walk up to and check. Trust, as defined in Theme 10, is not that thing yet. The manifesto describes trust as the confidence stakeholders place in the organization. That is a perception outcome. It is downstream of infrastructure, not infrastructure itself.

Infrastructure has surfaces. You can name them, measure them, assign owners, set thresholds, and trigger alarms. “Digital trust” in the manifesto has none of those properties. It has a list of inputs (data protection, cybersecurity, transparency) and an outcome (stakeholder confidence). Between those two, the scoreboard is missing. Without a scoreboard, trust is still a vibe you manage with public relations, not a signal you govern with instrumentation.

The same gap shows up in Theme 11. “Automating guardrails and controls” is named but not specified. Which controls? Against which behaviors? Audited how often? Owned by whom? The manifesto says leading companies are “rapidly experimenting to codify what works into a repeatable agentic playbook.” But the hardest question in agentic engineering is not how to codify what works; it is how to define what “works” means before you codify it. Agent behavior has no universal evaluation harness. Teams are discovering this the hard way.

Theme 6 has the same problem at a different altitude. Metabolic rate is a useful metaphor, but until you disaggregate it into cycle-time, lead-time, review-time, and revision-rate, it collapses into “go faster,” which is not a governable principle.

The common thread is this: ten of the twelve themes are correct as stated, and ten of the twelve themes are not auditable without a measurement layer that covers humans and AI on the same operating metrics. The manifesto names the themes. It does not name the instrument.

The 12 Themes, Briefly Grouped

To make the audit concrete, here is how the themes land when you sort them by “correct as stated” versus “correct but requires a measurement layer to be auditable”:

Correct as stated. Theme 1 (enduring capabilities beat tooling). Theme 2 (focus on economic leverage points).

Correct, and waiting for instrumentation. Theme 3 (value must move the business) assumes you can attribute EBITDA movement to AI specifically, which requires a controlled baseline the manifesto does not describe. Theme 4 (senior business leaders in the driver’s seat) is observably true in McKinsey’s own engagement filter, but there is no described metric for how you know a leader is “AI-capable” or not. Theme 5 (the 30-70 talent shifts) is a framework from the Rewired book, presented without an external benchmark against which to audit a real workforce. Theme 6 (metabolic rate) needs the disaggregation above. Theme 7 (platforms as strategic assets) is true, and every platform claim requires instrumentation on reuse, cost-per-call, and time-to-onboarding. Theme 8 (data as a consumable product) requires measurement of discoverability and time-to-consume. Theme 9 (design for adoption, build for scale) requires measurement of upstream and downstream coupling, which is exactly the kind of observability the theme describes but stops short of specifying. Theme 10 (trust) needs a scoreboard. Theme 11 (agentic engineering) needs the evaluation harness that tells you what “working” means. Theme 12 (continuous learning) needs measurement of whether the learning changed any decision.

Ten of twelve. The manifesto is directionally right on every theme I would expect it to be right on. It is the measurement layer, not the list, that is missing.

The One Sentence We Would Add

If I were editing this manifesto as a friendly contributor rather than a tracker, I would propose one additional theme and set it in the middle of the list. Something close to:

If you cannot measure humans and AI on the same scoreboard, you cannot govern either of them.

That sentence is not a takedown of the twelve themes. It is the thirteenth theme that makes the other twelve auditable. Without it, the manifesto is a set of statements about what to believe. With it, the manifesto becomes a set of claims you can check.

This is the layer we work on. We build the scoreboard. We do not think the twelve themes are wrong; we think ten of them are unverifiable without the thirteenth. In our own data, the teams that actually ship the behaviors McKinsey is describing share one property: they measure human and AI contributions against the same operating metrics, from the same dashboards, with the same cadence. Everything else follows.

The Trajectory of the Series

Step back from this article and look at the arc. Chapter one was about measurement treated as perception. Chapter two was about design treated as architecture. Chapter three was about factories treated as scale. Chapter four was about enterprise architecture treated as a binary choice. Chapter six, this one, is about a manifesto that, for the first time, names trust and automating guardrails as themes in their own right.

The line keeps moving in the right direction. Each chapter gets one word closer to governance infrastructure. “Governable” became “trust” became “automating guardrails and controls.” The next word in the sequence is scoreboard, or something like it. When that word appears in a McKinsey article, the series will have arrived at its destination, and the argument we keep making will have become the consensus.

Until then, the job is simple. Acknowledge the progress honestly. Audit what is missing carefully. Keep building the layer the manifesto keeps pointing toward without naming. The manifesto does not have to be wrong for the scoreboard to be necessary. Those two things can be true at the same time, and in the case of this article, they are.

This analysis synthesizes McKinsey’s The AI Transformation Manifesto (April 2026), and references our prior tracking of McKinsey Measured the Wrong Thing, McKinsey’s Three Diagnoses, McKinsey’s Agent Factory Prescription, and McKinsey’s Two Paths.

Victorino Group builds the measurement layer that turns 12 themes into auditable capabilities. Let’s talk.