Your AI hub can't do the 70%

Seventy percent of AI's value comes from changing how people work. That is BCG's number, published last month, and it is the strongest argument against the thing BCG is selling.

The piece is Why Your Organization Needs an AI Hub, 25 August 2026. It is a good piece, and I want to be fair to it. An AI hub, in its definition, is "a dedicated entity with the mandate to coordinate all AI initiatives throughout the organization," anywhere from 10 people to more than 100, holding "explicit authority to oversee all of the company's AI efforts." Coherence, speed, governance, adoption. Four phases running from hub-led to AI-first. If you are a CIO looking at forty disconnected pilots and no way to tell which ones are real, you will read it and feel seen.

Then you reach the arithmetic and the article argues against itself.

The bureaucracy objection is the weak one

The reflex is to say an AI hub is a center of excellence in a new lanyard. Scott Sambucci gave that reflex its best line in September 2025: "Committees are where AI goes to die." He was describing institutions that were still standing up an AI Center of Excellence nearly three years after ChatGPT shipped. It is nearly four now.

Fair objection. Wrong one.

Centralization is not the failure mode. Plenty of things belong in exactly one place, and I would fight to put them there. One identity model, so an agent inherits a real person's permissions instead of a service account with god rights. One audit trail. One eval harness, because your evals are the moat and nobody builds a second one well. One procurement path, so four teams do not sign four contracts with the same vendor at four prices. Every team that builds those alone builds them badly and alone.

The problem with the hub is not that it centralizes. It is that BCG tells you precisely where the value is, and it is nowhere the hub can reach.

The 70% has been a slogan for seven years

Here is the sentence the article rests on: "In BCG's experience, just 10% of the value from AI comes from algorithms and 20% from data, while the remaining 70% comes from changes to the operating model and new ways of working."

Ten, twenty, seventy. BCG has been publishing those three digits since at least March 2019. The digits have never moved. Everything around them has.

Every phrase is BCG's own, from the publication named below. Use the arrow keys, or click a year.

2019201920202020202320242026

22 March 2019

In Travel, It's Time to Push AI Beyond the Pilot Phase

10

data science and the algorithms themselves

20

enabling technology infrastructure and data engineering

70

embedding AI into business processes

Dividingthe challenge of implementation“the 10-20-70 problem”

In 2019 it was "the 10-20-70 problem of machine learning." By November 2020 BCG's own account was posting that companies "must apply the 10, 20, 70 rule." A problem is something you observe. A rule is something you follow. Nobody announced the promotion.

Chase the provenance and it loops. The one place BCG points at a source is an April 2020 article: "According to joint BCG and MIT research, a good rule of thumb is to dedicate around 10% of AI investment to algorithms, 20% to technology, and 70% to business process transformation." The words "a good rule of thumb" are a hyperlink. Follow it and you land on BCG's October 2019 report drawn from the MIT Sloan Management Review survey of more than 2,500 executives, where the sentence reads: "A good rule of thumb is to consider AI to be 10% about algorithms, 20% about technology, and 70% about business process transformation." Same rule of thumb, sitting in a paragraph of advice rather than in the findings. The citation is a mirror.

Be precise about what that does and does not mean. It does not mean the 70% is wrong. I think it is directionally right, and everything below depends on it being right. It means nobody has measured it, and a number nobody measured is a strange foundation for a hundred-person org.

The report everyone cites to justify a hub concludes the opposite

Every enterprise AI deck of the last year carries the MIT number. Ninety-five percent. It comes from MIT Project NANDA's "The GenAI Divide," July 2025, and the real claim is narrower than the meme: 95% of organizations getting zero measurable return, not 95% of pilots failing.

Read to the conclusion. Organizations that cross the divide "buy rather than build, empower line managers rather than central labs, and select tools that integrate deeply while adapting over time."

Line managers rather than central labs. Section 6.2 is more specific still: the winners "sourced AI initiatives from frontline managers, not central labs," instead of "relying on a centralized AI function to identify use cases." Section 6.1 names the barrier outright. Not integration, not budget. Organizational design. Companies succeed when they "decentralize implementation authority but retain accountability."

The statistic being used to justify the hub comes from the document telling you not to build one.

That does not settle it, and I will not pretend it does. NANDA is 52 organizations and 153 leaders recruited at four conferences, self-reported, version 0.1, and its own authors call the figures directional. There is a better detail buried in the appendix. The interview script asked "Who leads implementation efforts (e.g., IT, line managers, AI CoE)?" and the report never publishes the answer distribution. "Center of Excellence" appears exactly once in twenty-six pages, in that question. They collected the data that would settle the argument and printed the conclusion instead.

Almost nobody hands the hub the authority the definition demands

BCG is unambiguous that a hub must hold "explicit authority to oversee all of the company's AI efforts." Heidrick & Struggles asked 242 organizations who actually owns AI strategy.

Heidrick & Struggles asked 242 organizations. Published 24 February 2026. Try each region.

OverallUnited StatesEurope

• Chief information, technology, or digital officer30%
• Chief data & analytics officer23%
• Chief AI officer or most senior AI executive21%
• An AI center of excellencethe body being prescribed7%
• CEO6%
• CFO4%
• The executive leadership team3%
• The board2%

Seven percent. One point above the CEO, and the CEO is not doing this full time. The mandate sits with the CIO, the CTO, the CDAO, whoever already had the platform and the headcount, which is roughly where it sat before anyone said "hub."

In the same research, nearly half of the AI leaders surveyed said their organization "has simply reclassified existing positions to include AI responsibilities." That is the honest description of most AI hubs. A renamed team, an inherited budget, and a mandate to coordinate people who do not report to them.

The variable that moves EBIT is the one only the workflow owner can touch

McKinsey tested 25 organizational attributes against EBIT impact. The winner was not structure, sponsorship or centralization. It was workflow redesign: "the redesign of workflows has the biggest effect on an organization's ability to see EBIT impact from its use of gen AI."

The 2026 follow-up has the number that should stop a boardroom. AI high performers "remained flat at about 6 percent of all respondents." EBIT attribution sat at 37%, which McKinsey calls "essentially unchanged from 2025" despite growth in the share of organizations scaling AI. And in the same survey: "Eighty percent of respondents report that AI has improved their individual productivity."

McKinsey, the state of AI in 2026. Both panels share one scale.

The gap

• Individuals who say AI improved their own productivity80%
• Companies that count as AI high performers6%

Two years of organizational scaffolding moved the lower bar by zero. EBIT attribution sat at 37%, which McKinsey calls essentially unchanged from 2025.

What separates the 6%

• High performers who fundamentally redesigned workflowsnearly three-quarters
• Everyone elsejust one-quarter

Not structure. Not sponsorship. Not centralization. The one variable that separates them is the one a central body cannot reach into a business unit and perform.

Eighty percent of people are faster. Six percent of companies can find it in the P&L. Two years of intensive organizational scaffolding moved that number by zero.

That gap does not live in an org chart. It lives in the fifty small decisions inside one team's week. Which review step is now redundant. Which handoff existed only because a human needed a queue. Which weekly report nobody has opened since 2023. Robert Glaser said it better than I can: "the adoption unit is no longer the organization, and maybe not even the team. It is the loop inside the work."

You cannot coordinate your way into that. A central body cannot know which of your handoffs is vestigial. The person doing the handoff barely knows.

We already ran this experiment on data teams and published the results

Here is the part that unsettled me. The best post-mortems on centralized technical teams are six to eleven years old, they were written about data science rather than AI, and they all say the same thing. If you are standing up a hub in 2026, they were written about you.

Riley Newman, Airbnb's first data scientist, in 2015: the centralized model was tempting, and then "we became viewed as a resource and, as a result, our work became reactive." Eric Colson at Stitch Fix, in 2019, on what a central team sounds like from the inside: "waiting on ETL changes" and "waiting on ML Eng resources" are the visible symptom, and "the more insidious impact lies in what you don't hear, because you can't lament what you haven't yet learned." Monzo spent three years dissolving its central ML team and landed on a sentence that reads like an epitaph for the category: "machine learning is an established tool that is used by many disciplines to help teams reach their goals."

I went looking for the 2026 equivalent, written about an AI hub instead of a data team, and could not find one. Not because hubs are working. Because none of them is old enough yet.

Newman is also the reason not to over-read any of this. Airbnb decentralized and deliberately stopped short: "by not fully shifting toward an embedded model we're able to maintain a vantage point over every piece of the business." He kept the hub. He stopped putting the work in it.

Nothing in this argument, mine included, has been measured

I have just spent a thousand words using statistics to argue against an article built on statistics. So here is the ledger.

Including the ones I used above.

The claim	Sample	What kind of evidence
70% of AI value is ways of working BCG, 25 Aug 2026	none given	A rule of thumb whose only citation links to a rule of thumb
1.7x revenue growth, 3.6x three-year TSR BCG Build for the Future 2025	1,250 CxOs	Top self-assessment band against the bottom 60%. The methodology says overall company performance has multiple drivers beyond AI
72% of CEOs are the main decision maker on AI BCG AI Radar 2026	640 CEOs	Executives rating their own authority
Half of workers say governance is unclear BCG AI at Work, 3 Jun 2026	11,749 workers	Perception of governance, which is not the same as governance
95% of organizations get zero return MIT Project NANDA, Jul 2025	52 organizations	Interview synthesis, version 0.1, which its authors call directional
Workflow redesign has the biggest effect on EBIT McKinsey, Mar 2025	1,491 respondents	Self-reported and cross-sectional, so association rather than cause
AI high performers flat at about 6% McKinsey, 2026	1,719 respondents	Self-reported, but a time series against itself
An AI center of excellence owns strategy at 7% Heidrick & Struggles, Feb 2026	242 organizations	Self-reported org chart, which is the one thing self-report is good at
Centralized vs federated vs hub and spoke, tested against outcomes	Empty. I went looking and found consultancy and vendor taxonomies, and no study comparing the topologies against a measured result.

Not one measured financial outcome in the set. Every number in this debate is an executive describing their own program, and the taxonomy the debate runs on, centralized against federated against hub and spoke, has never been tested against an outcome by anyone I can find. That absence is the most interesting fact in the whole file, and it should make everybody in the argument quieter, me included.

Centralize the substrate, decentralize the loop

So build the hub. Just be honest about which number it is chasing.

The 30% is real and it is genuinely centralizable. Identity, permissions, data access, model access, evals, audit, procurement, the boring plumbing that every team needs and no team wants to own. Build that once, well, and put your scarcest people on it. That is architecture, and it is worth your innovation tokens.

The 70% is a workflow problem, and workflows belong to the people inside them. What a hub can do there is narrow and unglamorous: make it cheap for a team to try something, cheap to measure whether it worked, and safe to delete the step that stopped mattering. Not run the project. Lower the cost of the team running it.

BCG's own maturity model already says this, though not in those words. Read the four phases end to end and it is not a maturity model. It is a shrink schedule.

Objective, talent and control text is BCG's own. Use the arrow keys, or click a phase.

Phase 1Hub-ledPhase 2Center and podPhase 3Business-unit-ledPhase 4The AI-first organization

Share of AI talent held by the huba reading of BCG's talent column, not a BCG figure

Objective

Demonstrate credible value quickly, concentrate scarce resources, and establish responsible AI delivery discipline.

Talent model

Centralize scarce AI talent in the hub.

Risk and control model

The hub owns delivery controls and approvals.

What BCG says must be true to reach phase 2

Nothing stated. The phase is described, and the condition for leaving it is not.

BCG's three founding questions are who leads the hub, what the biggest stumbling block is, and where to start. Ask a fourth one, on day one, before anyone is hired. What does this hub stop doing when it works?

A hub that cannot answer that is not a hub. It is a department. Departments do not dissolve. They budget.

Glossary

Terms

CDAO

Chief data and analytics officer: the executive responsible for a company's data and, increasingly, its AI efforts.

center of excellence

A central team of specialists that sets standards and does shared work for a topic such as AI, instead of each business unit building its own skills.

EBIT

Earnings before interest and taxes: a company's operating profit. Surveys use it to ask whether AI spending is actually showing up in profit.

ETL

Extract, transform, load: the routine of pulling data from source systems, reshaping it and loading it into a place where analysts and models can use it.

eval harness

Code that runs a system against a fixed set of test cases and records the scores, so different approaches can be compared fairly and repeatably.

Evals

Repeatable tests for an AI system: feed it inputs, score the outputs against a standard, and compare runs. They show whether a change actually helped or hurt.

Federated model

An organizing approach where each team or business unit runs its own AI work under shared standards, instead of one central group doing everything. It trades control for local fit.

Hub and spoke

An organizing model with a central core team that sets standards and tools, and surrounding teams that do the day-to-day work. It sits between fully centralized and fully federated.

Tools

ChatGPT

OpenAI's conversational AI assistant, available as a web app and mobile apps.