How to build an AI Enabled Enterprise: The MuShuHaRi method by Himanshu Niranjani
- benjamin. brl
- 16 juin
- 5 min de lecture
I picked up a book last week-end: How to build an AI-enabled enterprise by Himanshu Niranjani , founder of Be Human Capital and a seasoned technology executive with over 25 years of experience building hyper-scale platforms at companies like Amazon, Meta or Microsoft.
I found the most useful thing I have read on enterprise AI in months, I'd like to share with you. A diagnostic, a methodology and a framework.
The diagnostic is ten questions, binary answers, no partial credit. "In progress" scores zero. "Almost" scores zero. "We have a plan for this" scores zero, you can test yourself below:
That one design choice is the whole point. Most companies running AI today are not slightly behind where they think they are. They are at zero, spending real money, convinced they have a strategy.
What state are you actually in today?
The book, by Himanshu Niranjani, pairs each diagnostic question with two things: the Trap, what executives tell themselves, and the Reality, what the answer actually is. Read three of them and you will recognize your own last board update.
On objectives. Can you draw a line from each AI initiative to a named item on your P&L, a number the CFO has agreed will move, by a defined amount, on a defined date? The Trap: "we are building capabilities that will generate significant value across multiple business areas." The Reality: that is not a P&L line. It is a hope.
On production. Do you have at least one model serving real external customers who make real decisions on its outputs? The Trap: "we have a model in internal beta performing very well with our employees." The Reality: internal users are forgiving. They do not churn. A model that has never faced external customers has not been tested.
On return. Has your AI program generated cash, verified by the CFO, that has appeared on your P&L? The Trap: "our models are projected to generate significant savings over the next twelve months." The Reality: projections are not results. A signature on an ROI projection is not confirmation that the money arrived.
Notice what every question routes through. Not the CTO. The CFO. The test refuses to accept technical health as proof of business value. Its sharpest line is on metrics: accuracy, F1 score, and latency measure whether the model is doing its technical job. They do not measure whether that job is moving the P&L.
That is your current state, measured honestly. Now the ladder it places you on.
The framework borrows a Japanese idea called Shuhari: the stages of mastery, from obeying the forms, to breaking them, to transcending them. The author adds one stage in front of it. Mu. The state of unconscious unreadiness.
Four rungs:
Mu is where a company runs disconnected pilots, has no governed data, no dedicated AI team, and believes it is further along than it is. This is where most enterprises actually sit. They are paying what the author calls the Hidden Year Invoice: two to six million on pilots that will never reach production, without knowing the invoice exists.
Shu is the white belt. Foundational discipline. Using AI to cut internal cost, where failure stays contained, and converting those verified savings into the budget for the next stage.
Ha is the green belt. Commercial deployment. AI now touches real customers and real revenue, which is also where failure starts to damage the brand and the infrastructure bar rises sharply.
Ri is the black belt. AI embedded in strategy, governance, and operations at the same time.
The useful claim is not the ladder. Every consultant has a ladder. The useful claim is the rule between the rungs: no stage is funded until the previous one has produced returns the CFO has verified. No board member is asked to bet on a projection. That is the same discipline as the ten-question test, turned into a sequencing rule. It is the rare anti-hype idea in a book with cherry blossoms on the cover.
The different kinds of AI
The framework also splits enterprise AI into five archetypes, which matters because most companies deploy them in the wrong order or don't know how to deploy their strategy step by step, which is also part of reasons for main failures.
OpsAI is the back office: finance, HR, legal, procurement. High-volume, low-judgment work, where failure is contained and which generates the labeled data later stages depend on.
CXAI is customer-facing: support, service, engagement. Reducing cost-to-serve without degrading the experience.
RevAI is the commercial engine: personalization, churn prediction, dynamic pricing. Revenue growth that does not require proportional headcount. This is where the economics get interesting, and where deployment cost jumps to seven or eight figures and time-to-value runs in years, not quarters.
StratAI is the high-stakes layer: capital allocation, market expansion, scenario planning. Turning gut-feel strategy into probability-weighted decisions.
RegAI is governance and compliance: enforcing policy at a speed and scale human oversight cannot match.
The author's sequencing claim is that the first two are white-belt work and the rest are not. RevAI cannot be deployed safely on top of an engineering organization that cannot reliably ship production software. Here is where you have to think for yourself, not follow the belt order. The book assumes the back office is everyone's right starting point because it generates training data and self-funds the rest. For some companies the real leverage is customer-facing from day one. A maturity model describes an average. It does not describe your business. The question is not "which belt am I," it is "where is my first verifiable dollar of AI value, and have I built the foundation to collect it."
So what about the 40x?
The book sums its whole promise up in one chart. Two paths climb the same four stages, Mu to Shu to Ha to Ri. Both start in the same place, a 7x EBITDA multiple. They end far apart.
The conventional path, dotted, drifts up to 28x. It starts with what the author calls the Hidden Year: two to six million paid at the Mu stage for no production value. The disciplined path, solid, reaches 40x. The gap between the two lines at the end is twelve turns of EBITDA, and on a hundred-million business that gap is the entire argument of the book.
It is a clean chart. It is also the author's, not the market's. The two curves are drawn to sell the method, and no enterprise re-rates this smoothly. What the chart does get right is the shape: the early stages move the multiple barely at all, and the real expansion comes late, at the commercial and systemic stages, only after the foundational work is done. You do not earn the multiple by starting. You earn it by finishing.
And the market only pays once you can prove it. AI-defensible software does command high multiples, but buyers now stress-test those claims in dedicated diligence. In 2026, an evidence-backed story about how AI strengthens your economics is a deal-pricing variable, not a branding one. Which is what the ten-question test was telling you all along. The multiple is downstream of one thing: a number the CFO can sign.
So stop asking your CTO whether the AI works. Ask your CFO whether it has changed a number. If the honest answer is "in progress," you have your score.
One question for your next executive committee: of your current AI initiatives, how many would survive a test where "almost" and "in progress" both score zero?
Commentaires