The Productivity Harness: Why the Value of AI Sits Outside the Model
BCG’s 2026 workforce study puts a number on it. In the firms that actually convert AI pilots into cash returns, roughly 10% of the value comes from the algorithm itself, 20% from technology and data plumbing, and 70% from people, process and change management. BCG calls this the 10-20-70 rule and treats it as the single most consistent finding across its transformation practice. Almost none of the enterprise value sits in the frontier LLM. Nearly all of it sits in retrieval, workflow, permissions, evaluation, retraining and the operating model wrapped around the model.
This is not a theoretical concern. MIT’s NANDA initiative found that 95% of enterprise generative-AI pilots produce no measurable P&L impact, and named the “learning gap” — the harness — as the binding constraint. McKinsey’s State of AI 2025 reports that even as adoption goes near-universal, only a minority of firms redesign workflows around AI, and only those firms see enterprise-wide EBIT effects. The Stanford AI Index 2025 sits on the same finding: model quality is converging while diffusion, integration and organizational adoption are where value now diverges.
The proof points are unambiguous when the harness is in place. Morgan Stanley reports 98% adoption across its financial advisor workforce and a shift in document access from about 20% to 80% for its research corpus — a change driven not by the model but by the retrieval architecture, permissioning and prompt scaffolding around it. Klarna is the more instructive story precisely because it went both ways: initial customer service consolidation with dramatic productivity gains, followed in 2025 by a partial rehiring cycle when quality regressed. The lesson is not that AI failed; it is that the human loop, escalation paths and quality assurance are part of the harness, and removing them prematurely destroys the very value the model created.
For boards, this reframes the capital allocation question. The economics of AI reward operators who invest disproportionately outside the model. That means:
- Retrieval architecture and data rights, so the model has clean context.
- Workflow redesign, so the marginal cost of using AI approaches zero for the user.
- Evaluation harnesses and human-in-the-loop, so quality is monitored, not assumed.
- Change management, so the 70% of value BCG identifies actually accrues.
AIMG View. Model prices are collapsing. The premium in AI is not in what you buy, it is in what you build around it. Compute is the commodity; the harness is the moat.
AIMG Bottom Line. Ask your teams three questions this quarter: which workflows are we redesigning, not merely accelerating? Who owns the evaluation loop? And where in our operating model does the 70% actually live? If the answers are vague, your AI budget is subsidizing someone else’s compute bill.