The Adoption Playbook & the Human Constraint

The adoption-vs-value gap

The headline numbers say adoption is solved. Stanford’s AI Index reports organizational AI use jumping to 78% in 2024 from 55% in 2023, with generative-AI use in at least one business function more than doubling from 33% to 71%.⁹ McKinsey’s November-2025 survey put organizational AI use at 88% and generative-AI use at 79%.⁸ At the worker level, Microsoft and LinkedIn found 75% of global knowledge workers already using generative AI at work, 78% of them bringing their own tools.¹¹ Diffusion speed is genuinely unprecedented: a nationally representative US survey found work adoption of generative AI “as fast as the personal computer” and overall adoption faster than either PCs or the internet.¹⁷

The value numbers say something else. In McKinsey’s March-2025 edition, more than 80% of companies reported no material contribution to earnings from their generative-AI initiatives.⁸ The Stanford AI Index is blunt: among firms that report any financial impact, most put it “at low levels” — cost savings of less than 10% and revenue gains of less than 5% are the modal outcomes.⁹ The same nationally representative survey that found record adoption speed also found that only 1–5% of total work hours are actually assisted by generative AI, with reported time savings equivalent to about 1.4% of hours.¹⁷ Reach is wide; depth is thin. MIT’s Project NANDA crystallized the gap into a contested but widely-cited figure — roughly 95% of enterprise generative-AI pilots delivering no measurable P&L impact — and located the cause in an organizational “learning gap,” the failure to integrate AI into workflows and structures, not a deficiency in the models.¹⁵

Two practitioner-credible syntheses name the cause directly. McKinsey’s own finding is that “the redesign of workflows has the biggest effect on an organization’s ability to see EBIT impact from its use of gen AI.”⁸ Harvard Business Review’s 2025 article opens with the cleanest statement of the thesis: most firms struggle to capture value from AI “not because the technology fails—but because their people, processes, and politics do.”²² The rest of this paper takes that claim apart — first the economics of why the lag is expected, then the diffusion and behavioral mechanisms that produce it, and finally the levers that close it.

The productivity J-curve

The gap between visible capability and invisible payoff is not new, and it has a name. Brynjolfsson, Rock and Syverson opened their 2017 paper on “the modern productivity paradox” with the observation that systems using AI “match or surpass human level performance in more and more domains” while “measured productivity growth has declined by half over the past decade.”¹ They called it a redux of the Solow paradox — “we see transformative new technologies everywhere but in the productivity statistics” — and evaluated four explanations: false hopes, mismeasurement, redistribution, and implementation lags.¹ Their verdict was that lags are “the biggest contributor”: the most capable AI “has not yet diffused widely,” and like other general-purpose technologies its full effect “won’t be realized until waves of complementary innovations are developed and implemented.”¹

The decisive move is that these complements — “the required adjustment costs, organizational changes, and new skills” — behave like a kind of intangible capital.¹ Building that capital consumes resources that depress measured output now, before the hidden asset begins to pay out later. In the dedicated 2018 paper, the authors formalized this as the Productivity J-Curve: as firms adopt a new GPT, “total factor productivity growth will initially be underestimated because capital and labor are used to accumulate unmeasured intangible capital stocks,” and later “measured productivity growth overestimates true productivity growth” as those hidden stocks finally generate measurable output.² The error in measured productivity therefore traces a J: down, then up. Their headline calibration: an intangibles-adjusted TFP measure runs 11.3% above the official measure at the end of 2004 and 15.9% above it at the end of 2017.² The historical analogies are pointed — electrification “took a generation for the nature of factory layouts to be re-invented” before its benefits appeared, and the British industrial revolution produced a half-century “Engels’ Pause” of wage stagnation during the capital build-out.² Their sharpest line for any leader impatient for ROI: “the more transformative the new technology, the more likely its productivity effects will initially be underestimated.”²

The leading skeptical counterpoint comes from Daron Acemoglu, who drew on AI-exposure data shared by J-curve co-author Daniel Rock but reaches the opposite magnitude. Working from a task-based model and a version of Hulten’s theorem, he estimates the macroeconomic upside of AI as “nontrivial but modest — no more than a 0.71% increase in total factor productivity over 10 years,” falling below 0.55% once hard-to-learn tasks (which lack the objective feedback AI needs to learn from) are accounted for.¹⁶ The contrast frames the genuine open question. Brynjolfsson sees a large payoff delayed by the complement-building lag; Acemoglu sees a payoff that is modest even at full diffusion. Both can be partly right: the J-curve says the gap we observe today is expected and not evidence of failure, while Acemoglu warns that the eventual top of the J may be lower than the hype implies. For an operator, the practical reading is the same in either case — the determinant of value is the complementary organizational work, and that work is where the human constraint lives.

Diffusion: why reach runs ahead of depth

If the economics explains why the payoff lags, Everett Rogers’ diffusion theory explains the shape of adoption and why it stratifies. Rogers defines diffusion as “the process in which an innovation is communicated through certain channels over time among the members of a social system,” and identifies five perceived attributes of an innovation that predict how fast it spreads: relative advantage, compatibility, complexity, trialability, and observability.⁷ Four push adoption up; complexity — “the degree to which an innovation is perceived as relatively difficult to understand and use” — is the only one negatively correlated with adoption.⁷ Relative advantage is “the strongest predictor.”⁷

Run generative AI through this rubric and the adoption-vs-value gap stops being a surprise. Relative advantage is large and visible for some tasks (drafting, summarizing, coding assistance) but weak or unproven for others — and the field-experiment evidence (below) shows it is near-zero for experienced experts on familiar work.¹³¹⁴ Trialability is unusually high — anyone can open a chatbot — which helps explain the record top-line adoption speed.¹⁷ But observability is low and, worse, negative: Rogers notes software already has “a low level of observability, [so] its rate of adoption is quite slow,“⁷ and as the identity section shows, with AI the observability of use actively triggers a social penalty, so people hide it. Rogers’ adopter categories — innovators (2.5%), early adopters (13.5%), early majority (34%), late majority (34%), laggards (16%) — also reframe the BCG finding that frontline adoption has stalled at 51% while overall regular use is 72%:¹⁰ the organization is not one adopter but a distribution, and the early/late-majority crossover is exactly where adoption efforts stall without opinion-leader mediation and reduced uncertainty.⁷

Trust, and the calibration problem

Diffusion theory says observability and relative advantage matter; the trust literature says why they convert to use, and the surprising answer is that the goal is not maximal trust. John Lee and Katrina See’s foundational review establishes that “because people respond to technology socially, trust influences reliance on automation,” and that trust matters most “when complexity and unanticipated situations make a complete understanding of the automation impractical.”⁶ Their central construct is appropriate reliance — trust calibrated to the system’s true capability — and they name the two symmetric failure modes: disuse, rejecting capable automation, and misuse, over-relying on imperfect automation.⁶ Both are forms of miscalibrated trust. The adoption-playbook implication is precise: the target is not to make people trust AI more, but to make them trust it exactly where it is reliable and override it where it is not.

Disuse has a well-documented behavioral engine. In the canonical study of algorithm aversion, Dietvorst, Simmons and Massey ran five incentivized experiments and found that people “are especially averse to algorithmic forecasters after seeing them perform, even when they see them outperform a human forecaster.”³ The mechanism is asymmetric confidence: “people more quickly lose confidence in algorithmic than human forecasters after seeing them make the same mistake.”³ This holds despite the algorithms’ real superiority — a meta-analysis of 136 studies they cite found algorithms outperformed human forecasters by about 10% on average.³ The aversion is triggered by seeing the algorithm err, not by abstract distrust, which is why a single visible AI mistake can collapse adoption even when the system is winning on average. This connects directly to the most concrete enterprise data point in the corpus: among workers who tried Microsoft Copilot and stopped, 44.2% cited distrust of answers as the reason.²¹

But the story is not one-directional, and over-claiming a universal “people reject AI” narrative would be wrong. Logg, Minson and Moore document the opposite tendency — algorithm appreciation — across six experiments: lay people “adhere more to advice when they think it comes from an algorithm than from a person.”⁵ The reconciliation is the key analytic move. Appreciation dominates for lay people and unfamiliar tasks; it “waned when people chose between an algorithm’s estimate and their own” and, critically, among experts — “experienced professionals, who make forecasts on a regular basis, relied less on algorithmic advice than lay people did, which hurt their accuracy.”⁵ Aversion and appreciation are not contradictory findings but a map of when each appears, and that map has a clear axis: expertise and the comparison being to one’s own judgment.

The expert / novice asymmetry

The trust map predicts that AI’s value — and the resistance to it — should vary sharply by skill. The field-experiment evidence confirms it, and the contrast between two rigorous studies is the cleanest illustration in the literature. In “Generative AI at Work,” Brynjolfsson, Li and Raymond studied the staggered rollout of a conversational AI assistant across 5,179 customer-support agents and found productivity (issues resolved per hour) rose 14% on average — but with “a 34% improvement for novice and low-skilled workers” and “minimal impact on experienced and highly skilled workers.”¹³ The mechanism is that AI “disseminates the best practices of more able workers and helps newer workers move down the experience curve.”¹³ AI’s value here is largely in lifting the bottom of the skill distribution.

Now the counterpoint. METR ran a randomized controlled trial with 16 experienced open-source developers on mature repositories they averaged five years of experience on, randomizing 246 real tasks to allow or disallow early-2025 AI tools. The developers forecast a 24% speedup beforehand and estimated a 20% speedup afterward — but the measured result was a 19% slowdown.¹⁴ Outside experts were even more wrong, predicting 38–39% speedups.¹⁴ All three vantage points erred in the same direction. This is the perception-vs-reality gap at the heart of the human constraint, made vivid: experts confidently believed they were faster while measurably being slower. (The authors are careful — small N, expert contributors on familiar complex code; it is explicitly not a claim about novices or greenfield work.)

The asymmetry resolves the apparent contradiction with the algorithm literature: experts under-rely (aversion, observed slowdown) exactly where lay novices appreciate and gain. It also reframes who resists and why. If AI’s value is largest for novices and smallest — sometimes negative — for experts, then the most skilled, highest-status people in an organization are precisely those for whom the relative advantage is weakest and the threat to professional identity strongest. Resistance is not irrational technophobia; for a senior expert it can be a locally accurate read of both the productivity math and the status math.

Resistance rooted in professional identity

The most underweighted barrier in most adoption plans is not capability or training — it is identity and social evaluation. The survey evidence is stark. Slack’s Workforce Index found that just 7% of desk workers consider AI outputs “completely trustworthy,” with 35% calling them only slightly or not at all trustworthy.¹² But the identity finding is sharper than the trust finding: in the Fall-2024 wave, 48% of desk workers said they would be uncomfortable telling their manager they had used AI for a common task, and the top reasons were that using AI feels like “cheating” (47%), fear of being seen as “less competent” (46%), and fear of being seen as “lazy” (46%).¹² Adoption is, for nearly half the workforce, something privately rational to hide.

Two studies move this from correlation to cause. In four preregistered experiments (N = 4,439), Reif, Larrick and Soll found “people who use AI at work anticipate and receive negative evaluations regarding their competence and motivation” — and that these evaluations “affect assessments of job candidates.”¹⁹ The belief that you will be judged for using AI is both held and justified. The behavioral cost is then pinned down by David Almog’s field experiment: hiring 450 remote workers for an AI-assisted task, he found that “workers adopt AI recommendations at lower rates when their reliance on AI is visible to the evaluator, resulting in a measurable decline in task performance.”²⁰ Specifically, making AI use observable cut reliance from 30.5% to 26.2% (a 14% relative drop) and accuracy from 79.1% to 76.4% — and reassuring the evaluator about the workers’ track record did not remove the effect.²⁰ The elicited belief: workers fear that heavy AI reliance “signals a lack of confidence in their own judgment.”²⁰ This is the human constraint in its purest form — a social/incentive barrier, not a capability barrier, that suppresses use and degrades measured performance, while simultaneously distorting the adoption statistics that leaders rely on. Note the cruel interaction with Rogers: the one diffusion attribute that should drive uptake through social proof — observability — is the very thing that, for AI, triggers the penalty.

The playbook: align accountability with control

The evidence converges on a small number of levers, and they are organizational, not technical. The first comes directly out of the algorithm-aversion lab. In their follow-up, “Overcoming Algorithm Aversion,” Dietvorst, Simmons and Massey found that “participants were considerably more likely to choose to use an imperfect algorithm when they could modify its forecasts, and they performed better as a result” — and, decisively, “the preference for modifiable algorithms held even when participants were severely restricted in the modifications they could make.”⁴ The effect was “indicative of a desire for some control… not for a desire for greater control”; even a slight, capped ability to adjust the output increased adoption, satisfaction, belief in the algorithm’s superiority, and willingness to use it again.⁴ The operational principle: keep a human in control of the output, and align accountability with that control. People accept tools they can steer and are answerable for; they reject tools that displace their judgment while leaving them responsible for the result.

The second lever is workflow redesign over tool deployment. This is where McKinsey’s own data points — workflow redesign “has the biggest effect on an organization’s ability to see EBIT impact”⁸ — and where BCG locates “the next frontier: from adoption to value with end-to-end redesign,” noting that the roughly half of companies starting to reshape processes “invest more in their people—and it pays off.”¹⁰ MIT NANDA’s “learning gap” is the same point stated as a failure mode: generic tools “stall in enterprise use since they don’t learn from or adapt to workflows.”¹⁵ Buying licenses is the cheap, visible move; the Recon Analytics data shows where that ends — a 35.8% Copilot conversion rate, meaning roughly two-thirds of provisioned licenses sit idle.²¹ A license is access; adoption is what happens after the workflow is rebuilt around it.

The third lever addresses the identity barrier head-on, and it runs through middle managers and training. The training deficit is enormous: Microsoft and LinkedIn found only 39% of people who use AI at work have had any company AI training, and only 25% of companies planned to offer generative-AI training that year, even as 79% of leaders agreed their company must adopt AI to compete and 60% worried leadership lacked a plan.¹¹ BCG found only 36% of employees satisfied with their AI training, and identified “proper training, leadership support, and access to the right tools” as what breaks the frontline ceiling.¹⁰ Managers are the load-bearing layer here for two reasons grounded in the evidence: in Rogers’ terms they are the local opinion leaders who “put their stamp of approval on a new idea by adopting it” and reduce uncertainty for the early and late majority;⁷ and because the identity penalty is administered by the manager’s perceived judgment, only the manager can neutralize it — by making AI use expected and sanctioned rather than something to conceal, directly countering the 47% “feels like cheating” reflex.¹²¹⁹ HBR’s synthesis names the corollary failure: where incentives are misaligned, workers may decline to surface AI-driven time savings out of self-interest, so the value never shows up in the numbers.²² (summarized from HBR’s argument — its body beyond the verbatim dek is paywalled; not a verbatim HBR quote, see verification note)

Put together, the playbook is coherent and falls out of the sources rather than being imposed on them. The J-curve says expect a lag and invest in intangible complements through it.¹² Diffusion theory says manage the adopter distribution through opinion leaders and lower the complexity and the (negative) observability cost.⁷ Trust theory says aim for calibrated reliance, not maximal trust.⁶ The control lever says give people a steering wheel and align accountability to it.⁴ The identity evidence says make use safe and expected, which only managers can do.¹²¹⁹ And the skill-asymmetry evidence says target the people who actually gain — novices and the lower-skill tail — while not pretending experts will see the same lift.¹³¹⁴ The technology is, for now, the part of this that is already solved.

References