The 7% Problem

Companies spend about 93 cents of every AI dollar on technology and 7 on the people and process that turn technology into output. That unspent 7 cents is the most likely explanation for the missing AI returns. A research sidebar off "AI spreads while training lags" (Minnesota Star Tribune).

Read the companion deck (PDF)→

The unspent 7% is the AI return itself

The article reports two figures a few paragraphs apart, that 93% of corporate AI budgets go to the technology and 7% to training and supporting its use, and that "few companies have yet to show a return." The two figures describe the same problem. MIT's Project NANDA tracked $30 to $40 billion of enterprise generative-AI spend and found a funnel where about 80% of organizations explore the tools, 20% launch pilots, and 5% reach production with measurable profit impact. McKinsey's 2025 survey reaches the same place, since only about 39% can link any earnings impact to AI and only about 5.5% see more than 5% of EBIT come from it (Exhibit 1).

Brynjolfsson, Rock and Syverson named the mechanism the productivity J-curve, where general-purpose technologies require complementary intangible investment in process redesign, retraining, and reorganization, all of it expensed and none of it capitalized. Measured output dips first while that capital is built, then climbs once it is in place. The 93/7 split shows firms buying the technology and declining to build the intangible capital, so the curve has nothing to lift it. The strongest objection holds that this is fine and temporary, since low returns now are the dip before the climb, and we take it seriously. However, the curve turns up only for firms that buy the complementary capital; firms that skip it show no measured recovery.

Even if every firm tripled the 7%, the training function could not deliver it.

The article's implied fix, more money for training, runs into a constraint the piece does not address: there are almost no trainers. BLS counts about 452,300 training and development specialists and 46,400 managers, roughly 499,000 dedicated workplace trainers, against about 155.5 million jobs, which is one trainer for every 312 workers. The occupation is one of the faster-growing categories BLS tracks, with specialists projected to grow 11% and managers 6% through 2034 against 3% for all occupations, and even so it adds only about 51,500 people over the decade. AI use went from near zero to roughly 80% of workers in three years, about 124 million people on current employment (our estimate), which works out to roughly 249 AI users per professional trainer (Exhibit 2).

A ratio of 312 to one puts classroom delivery of AI fluency out of reach. The 7% cannot be delivered through a central learning-and-development team, because the headcount to staff one does not exist and will not exist on any plausible hiring path. That pushes the answer toward a structural change in how the teaching gets done, which section 04 takes up.

Skip the enablement and the AI goes underground

KPMG's 2025 work finds 87% of workers using AI at least weekly and 51% daily, while 44% have used it in ways that break company policy, 58% have relied on AI output without checking it, and only 41% work somewhere that has a generative-AI policy at all. Withholding training moves the adoption off the governance grid instead of slowing it. Radha Chavali, the CIO at Affinity Plus, describes the pattern this way, that "employees are using tools on their own, with sensitive data, because the official training didn't meet their actual needs." The unspent 7% becomes a security and compliance question, and IBM's published breach-cost premium for unsanctioned AI runs into the millions per incident (Exhibit 3).

The firms with a return fund the 7% inside the line organization

The companies in the article that show results share one habit that has little to do with the size of a training department, since they push enablement into the workflow. Affinity Plus rolled out Microsoft Copilot only after organization-wide training on how to use it, when to avoid it, and how to verify its output. Thrivent runs peer forums, live demonstrations, office hours, and team-specific sessions, and reports more than 90% of users still active each month and advisers saving about four hours a week. Mayo Clinic pairs fluency courses and hands-on workshops with internal agent-building and a nursing AI competition that drew hundreds of submissions. Each of these programs runs through the line, with peers and managers, rather than through a central classroom (Exhibit 4).

McKinsey frames the capability-building work as three layers, starting with AI literacy, then adoption embedded in redesigned roles and incentives, then domain-specific transformation of the work itself. BCG's 2025 read stresses structured experimentation and tracking measured impact over counting seats. Across both, the 7% buys a change in how work gets done rather than a course catalog.

Minnesota Star Tribune, "AI spreads while training lags" (Hartzog) · Deloitte (93/7 split; CFO Signals / CTO commentary; single-source, finance-framed) · MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025" · McKinsey, "The State of AI" 2025 · Brynjolfsson, Rock and Syverson, "The Productivity J-Curve" (NBER w25148) · KPMG U.S., 2025 ("Shadow AI is already here"; American Worker survey) · BLS OEWS (May 2019, May 2025) and Occupational Outlook Handbook (2024 to 2034 projections, SOC 13-1151, 11-3131); Second Order calculations (data and method in analysis/ai_training_gap/data/) · Carlson Analytics Maturity Model, University of Minnesota.

Exhibits 1 and 3 compile reported survey readings, each cited; Exhibit 2 is a Second Order calculation on BLS data; Exhibit 4 is a synthesis framework, not a computed result. Second Order is an independent research briefing, provided for discussion purposes only; it is not investment, legal, tax, or accounting advice. © 2026.