Two research passes across the published record, including the cases that argue against us. This page is the short version: what the record supports, where it pushes back, and why almost none of it was measured in a company your size.
It is worth being direct about the evidence base. Nearly all of it comes from enterprises: the surveys of CEOs, the change-management budget ratios, the banks and the pharmaceutical companies. Bank of America has 213,000 employees. DBS built its position on a decade of platform and data-governance investment.
In a company doing $10–150M with one or two IT people and an outsourced provider, the layers those findings describe do not exist. There is no data team to fund at 30%. There is no change-management function to resource. There is no innovation group, no head of AI, and nobody to delegate this to.
So the enterprise playbook does not transfer, and copying it is how these programs fail in the middle market. What transfers is the direction of the findings, not the mechanism: that the constraint is people and process rather than technology, that the pull already exists inside the company, and that redesigning the work has to come before choosing the tool.
An enterprise VP needs three approvals to connect a system. A CEO at this scale owns the admin seat, or knows the one person who does, and can say yes in the meeting. That collapses the single largest source of friction in corporate AI programs, which is why the same work can move faster here than it does in a company a hundred times the size.
The scarce input is not curiosity. It is abundant, and in most companies it is currently unrecorded, unsupported, and happening on personal accounts. That is the single most useful thing to know before starting.
MIT’s Project NANDA documented what it called a shadow AI economy: when the official program stalls, the usage does not stop. It moves off-platform, where nobody can see it, support it, or govern it.
MIT Project NANDA
The same Wharton study found 31% actively working against their company’s AI initiatives, rising to 41% among younger workers, and more than half willing to use AI without approval. The authors are explicit that this is not a training gap: “it’s not simply a training deficit. It is a psychological one.”
This is the most consistent result across the record, and it comes from sources a board already trusts.
From a survey of 1,000 CEOs and their direct reports, with the observation that the algorithmic 10% routinely consumes a disproportionate share of an organization’s time and attention.
BCG
Against roughly 10% for organizations generally. The gap between those two numbers is most of the difference between a program that lands and one that does not.
Gartner
The strongest result on returns is a process result, and it is the one most often skipped because it is slower than buying something.
That is the returns figure. A separate finding in the same research runs alongside it: the high performers are roughly three times more likely to have fundamentally redesigned their workflows in the first place. Redesign rate and return rate are two different measures, and both point the same direction. Only 39% of organizations attribute any EBIT impact to AI at all.
McKinsey
It is tempting to assume the companies building this technology simply absorbed it. They did not. Every one of them had to do the enablement work deliberately, which is the clearest evidence that it cannot be skipped.
Across 132 engineers, 53 interviews and 200,000 Claude Code transcripts: engineers use Claude for 59% of their work, but over half can fully delegate only a fifth of it. Adoption is bimodal, and the company is building an internal AI fluency framework to address it.
A six-person acceleration team built role-specific agents. In their own words: “An agent per role may seem intuitive, but it failed in practice.” What worked was change management: “we found our champions, who ultimately helped scale adoption.”
An automation-native company still needed a declared “code red” and a full week of hackathon to move the number. Every new hire since has had to pass an AI-fluency assessment.
Zapier
A thesis that only collects supporting evidence is not worth much. These are the strongest cases against a people-led reading, and they are the reason the data layer is not optional.
213,000 employees, 95% using its internal assistant, and help-desk calls down 55%, with no mandate and no upskilling campaign. Deliberately not a language model: 700 curated intents across 110 systems. Their CEO on the constraint: “the data has to be perfect.” In banking, no amount of people-readiness clears a data-quality floor.
Peter Cappelli’s critique is the one we find hardest to argue with: “The problem is getting line and operating managers who supervise employees doing the work to get behind it.” Champions are individual contributors. If their managers’ numbers get worse when the team experiments, naming champions changes nothing.
The pull already exists inside your company, so the first job is finding it rather than creating it. The constraint is people and process rather than technology, so that is where the effort goes. The work has to be redesigned before the tool is chosen, which is why process sits at the end of the chain and not the start.
And because none of the enterprise scaffolding exists at this scale, the capability has to end up resident in your own people. There is nobody else to put it in. The framework sets out how that runs, and the PDP Assessment is how it starts.
If you read the record differently, that is a conversation worth having. We would rather be corrected early than sell you a thesis that does not hold up.
