AI adoption

The question this topic tracks: what actually binds AI adoption inside organizations once the tools are available? The early evidence points away from access and toward discovery — knowing where and how AI creates value in a specific production process. A second layer is ai-readiness: whether the organization has the strategic alignment, resources, knowledge, culture, and data conditions needed for the specific AI adoption purpose it is pursuing.

2026-kim-mapping-ai-into-production names the discovery constraint directly. In a preregistered field experiment across 515 startups that all received identical AI tools, credits, and training, the only firms that pulled ahead were those given case studies of how other firms had reorganized production around AI. That single informational nudge produced 44% more discovered use cases and, downstream, more completed tasks, more paying customers, and ~1.9× revenue. The authors frame the binding constraint as a managerial/cognitive search problem — discovering where to deploy AI — not access to the technology and not general managerial or technical skill.

This reframes “adoption” away from a tooling/provisioning event and toward an ongoing organizational search over the production process — the mapping problem: locating where AI creates value in a specific firm’s workflows. It connects adoption to work-redesign (the gains showed up as reorganized internal work) and to automation-and-substitution (output rose while labor demand held flat and capital demand fell).

2026-ermakov-designing-post-role-enterprise pushes the adoption constraint from discovery into architecture. Where 2026-kim-mapping-ai-into-production shows that firms need help discovering where AI creates value, Ermakov argues that discovery is still insufficient if work remains bundled into static job descriptions. The failure mode is “pilot purgatory”: broad experimentation with AI tools but little durable operating-model change because workflows, accountability, skills intelligence, and governance remain organized around pre-agentic roles.

2025-lee-generative-ai-critical-thinking adds a user-level quality constraint: adoption is not just whether knowledge workers use GenAI, but whether confidence in the tool suppresses the critical thinking needed to verify, integrate, and steward AI output. In their survey, higher confidence in GenAI was associated with less critical thinking, while higher self-confidence was associated with more critical thinking.

2026-lee-making-invisible-visible adds a worker-centered implementation constraint. In interviews and workplace demonstrations with 16 workers using decision-support AI in healthcare, finance, and management, adoption failures showed up less as generic resistance and more as misalignment between organizational goals and the actual work system: poor interoperability, hierarchy-level disagreement, AI placed on tasks workers valued or needed control over, weak explanations, and limited mechanisms for worker correction or verification. This makes adoption a sociotechnical fit problem, not just a search problem.

2026-dellacqua-jagged-technological-frontier makes the mapping problem task-level and empirical. In a randomized field experiment with 758 BCG consultants, GPT-4 improved speed, volume, and quality on tasks inside the AI capability frontier, but made consultants 19 percentage points less likely to solve a task selected to be outside it. Adoption quality therefore depends on whether workers and managers can map where AI helps versus harms inside a workflow, not merely whether the tool is available or used.

2025-brynjolfsson-generative-ai-at-work adds a production-rollout case where adoption was embedded directly into the workflow. In a staggered deployment across 5,172 customer-support agents, GenAI access raised issues resolved per hour by 15% on average, with the largest gains for less experienced and lower-skilled workers. This suggests adoption can succeed when the tool distributes situated best practices at the point of work, not just when workers receive generic access or training.

2025-handa-economic-tasks-ai adds an observed-use baseline for the adoption map. Instead of asking which jobs are exposed to AI or which workers report using it, the paper maps more than four million Claude.ai conversations to O*NET tasks and occupations. Usage was concentrated in software development and writing, but not confined there: about 36% of occupations showed AI use for at least a quarter of their associated tasks. This makes adoption visible as uneven task-level diffusion, not simply tool availability.

2025-agarwal-what-work-ai-doing sharpens that observed-use baseline into task-level-ai-adoption. In another analysis of four million Claude interactions mapped to O*NET tasks, AI usage clustered around non-routine, cognitively demanding, creative, and complex work, with the top 5% of tasks accounting for 59% of interactions. That turns the mapping problem into a task-characteristics problem: organizations need to identify which combinations of task demands invite AI use, not just which occupations are exposed.

2021-johnk-ready-or-not-ai-comes gives adoption a readiness frame. In interviews with 25 AI experts, the paper identifies five organizational readiness categories — strategic alignment, resources, knowledge, culture, and data — and argues that AI readiness must be assessed against the particular AI adoption purpose. This complements the mapping-problem evidence: even when a firm knows possible use cases, adoption still depends on whether processes, data, skills, culture, governance, and resources can support the intended deployment.

2026-reich-work-design-multidimensional-ai-threat adds the work-design predictor layer. In a cross-sectional survey of 2,257 employees, skill variety and autonomy were the most consistent positive predictors of AI adoption, while AI-threat dimensions showed mixed and differentiated associations with frequency and duration. This turns adoption from a simple attitude or access problem into a job-characteristics problem: roles with more discretion and varied cognitive work may create more opportunity to experiment with and embed AI.

2025-tully-ai-literacy-receptivity adds a pre-adoption relationship between ai-literacy and ai-receptivity from consumer and student settings. Across cross-country data and six preregistered studies, lower AI literacy often predicted greater receptivity, mediated by perceptions of AI as magical and feelings of awe for tasks associated with distinctly human attributes. That complicates adoption programs that assume knowing more about AI and wanting to use AI are the same curve: demystifying AI may improve calibration while reducing some forms of awe-driven appeal.

2025-kellogg-novice-risk-work adds a reverse-mentoring failure mode. In interviews with 78 junior consultants after GPT-4 task use, juniors expected to educate seniors but often proposed risk mitigations grounded in limited GenAI understanding, human-routine changes, and project-level interventions. That makes adoption quality depend not only on bottom-up energy or tool exposure, but on whether early users know which risks require system design, deployer controls, or ecosystem-level governance.

2026-almog-barriers-ai-adoption adds an evaluative-visibility constraint. In a field experiment with 450 U.S.-based online labor-market workers, making AI reliance visible to an HR evaluator reduced AI recommendation following and lowered final accuracy, even when evaluators were instructed to assess expected accuracy on the same AI-assisted task. Adoption can therefore fail because workers fear what AI use signals about their judgment, not because they lack access, use cases, literacy, or evidence that the AI helps.

2026-zhu-choose-your-agent adds an agentic-delegation version of the adoption problem. Participants preferred the more controllable Advisor mode, while Delegate access produced the strongest group-welfare gains and possible spillovers to non-users. Adoption quality therefore depends on whether the interaction mode is adoption-compatible: a technically superior delegate can fail to improve outcomes if users avoid it, distrust it, or prefer low-effort agency over autonomous execution.

2025-friis-performance-principle adds a principle-based adoption constraint at the occupation level. Better AI performance can convert much resistance to automation, but not where AI replacement violates the perceived moral meaning of the role. That makes moral-boundaries-of-ai-automation a distinct adoption barrier from access, discovery, readiness, literacy, image concerns, or task fit.

2025-hautamaki-fully-leveraging-ai-b2b-sales adds a managerial-capability version of adoption. In grounded-theory interviews with 32 top-level B2B sales managers’ supervisors, AI exploitation depended on sales managers’ data-based human capital, social capital for knowledge-sharing culture, and transformative AI-positive mindset. That extends the adoption map from “find use cases” to “build the managerial and knowledge-process capabilities that let a sales organization turn use cases into shared practice.”

2025-schweitzer-leading-ai-adoption adds a behavioral change-management layer. The article argues that AI adoption fails when leaders treat implementation as an engineering exercise and ignore how employees, designers, and managers react to AI through bounded rationality, distrust, autonomy concerns, loss aversion, vivid failure narratives, overconfidence, and escalation of commitment. This makes behavioral-human-centered-ai a peer topic to adoption rather than a footnote: adoption quality depends on whether implementation actively designs for human behavior across design, adoption, and management stages.

2018-fugate-change-compliance-championing adds a foundational change-support distinction behind AI adoption metrics. In a two-wave merger study, change management support predicted challenge appraisal, and challenge appraisal predicted both compliance and championing; threat appraisal predicted lower championing but not lower compliance. For AI adoption, this means usage, training completion, or policy compliance can overstate transformation if employees are following instructions while withholding discretionary problem solving, peer advocacy, and local repair work.

1997-goh-benchmarking-learning-capability adds the organizational-learning version of the adoption problem. If AI adoption is partly a search over use cases and workflow changes, organizations need routines for experimentation, knowledge transfer, and team problem solving so local AI discoveries can become shared practice. Without that learning capability, adoption can remain a portfolio of disconnected pilots even when tools, training, and managerial interest are present.

2026-ali-ai-adoption-multinational-workforce grounds the worker-centered fit problem in a live HR-system transition. Employees did not simply migrate from legacy to GenAI search: they selected among the GenAI system, legacy search, source documents, colleagues, and HR according to task and uncertainty. The case makes formal access a weak adoption indicator: device access, language, tenure, organizational terminology, and local work context determined whether workers could obtain useful support, while aggregate usage could hide excluded groups. It also makes adoption depend on the knowledge system around the model—current, localized content; tagging; guidance; and escalation routes—not the interface alone.

2026-lunich-why-workers-use-ai-less-often adds a post-access question: why do some workers reduce use after AI is already available? In its 2026 German survey, reduced use was uncommon, but the workers who reported it most often pointed to low practical value, unreliable results, uncertainty about data protection, and unclear workplace rules. This does not show that those factors caused the change, but it does make a useful distinction: adoption is not complete when a worker tries a tool. It has to keep earning a place in the work through useful, reliable output and conditions that make appropriate use understandable.

2026-shoghli-persona-migration-expectation-recalibration adds longitudinal evidence that post-use experience can recalibrate adoption rather than simply confirm pre-use intent. In a matched two-wave sample of 124 employees during an eight-week Microsoft 365 Copilot pilot at a state transportation agency, perceived usefulness declined significantly after use while aggregate ease of use, behavioral intention, and trust changed little; beneath those averages, 40% of baseline Skeptics became Cautiously Positive and 68% of baseline Champions moved to less enthusiastic personas. This suggests that adoption support cannot treat initial enthusiasm or resistance as stable: task-relevant examples, verification routines, and trust calibration may need to adjust as workers encounter the tool in practice.

2023-mkhize-examining-barriers-enablers-ai-adoption adds a grounded, pre-GenAI implementation case: the adoption condition is not abstract acceptance but a chain of operational fit. AI-enabled subtitling and content work gained support where speed and throughput were visible, but short tasks could remain faster manually; accuracy, data quality, integration, and traceability constrained trust; contractual restrictions and approval gates set hard limits on autonomous use. This single case does not establish which condition is generally binding, but it makes the mapping problem concrete: a viable use case still has to fit the task, the surrounding systems, and the organization’s obligations.

The report also raises an equity question without settling it. Higher-skill occupational groups were more likely to report increasing their AI use, while lower-skill groups were comparatively more likely to report decreasing it. That pattern could reflect unequal task fit, access to support, prior skill, or broader differences in work conditions; the survey cannot separate them. Still, adoption programs should not read stable or rising overall use as proof that the same conditions work for everyone.

Connections

• The search framing is information-processing / bounded-rationality flavored — an I-O and organizational-cognition lens on adoption, not just an IT-diffusion one.
• If discovery is the bottleneck, interventions that spread patterns of use (case libraries, cross-firm examples) may matter more than tool provisioning or generic training.
• If worker-centered fit is the bottleneck, adoption interventions have to make workers visible before full deployment: pilot testing, feedback sessions, task-fit diagnosis, explanation design, and worker control over outputs that affect accountable decisions.
• If job design is the bottleneck, adoption work has to change the opportunity structure of jobs — autonomy, skill variety, feedback, and workload — not just raise awareness or reduce anxiety.
• If novice risk work is the bottleneck, adoption programs need to support bottom-up learning without mistaking novice fluency for calibrated governance expertise.
• If image concerns are the bottleneck, adoption programs need to change evaluation norms and visibility structures, not just persuade workers that AI is useful.
• If managerial capabilities are the bottleneck, adoption programs need to develop local context-sensing, data literacy, use-case sharing routines, and cross-functional relationship-building — not just give frontline employees tools.
• If behavioral human-centeredness is the bottleneck, adoption programs need to diagnose predictable reactions to AI — distrust, loss aversion, image concerns, autonomy threat, overconfidence, and escalation — rather than assuming technically better systems will be accepted.
• If employee change support is the bottleneck, adoption programs need to distinguish compliance from championing: workers can meet rollout expectations while still not doing the discretionary advocacy, experimentation, and problem solving that make AI use durable.
• If organizational learning capability is the bottleneck, adoption programs need survey-feedback, knowledge-transfer, and remeasurement loops that convert experimentation into durable changes in routines, not just more tool trials.
• ai-receptivity should be treated as a time-dependent adoption input rather than a fixed pre-rollout trait: post-use persona migration and usefulness decline can coexist within the same pilot, so support should be recalibrated after workers encounter the tool in practice 2026-shoghli-persona-migration-expectation-recalibration.
• task-level-ai-adoption should stay analytically separate from organizational adoption: conversation traces reveal where AI is being tried, but not whether the work has been redesigned, governed, or connected to outcomes.

Contradictions & open questions

• Who benefits — floor or ceiling? Here, gains concentrate in the upper tail and show no heterogeneity by baseline performance. This sits in tension with task-level GenAI studies where lower performers gain most. Firm-level discovery effects and individual-level task-execution effects may simply be different mechanisms — surfaced, not resolved. Needs a second source on each side before this is a real claim.
• Does it last? The evidence is a 3-month window; whether discovered use cases compound or decay is untested.
• Does it generalize past young firms? Early-stage, low-inertia, self-selected accelerator startups may relieve the mapping problem far more cheaply than incumbents with entrenched processes.
• Discovery vs. architecture. 2026-kim-mapping-ai-into-production provides causal evidence that exposing firms to AI use-case patterns improves performance; 2026-ermakov-designing-post-role-enterprise argues conceptually that the deeper bottleneck is role architecture. These may be sequential mechanisms — first discover use cases, then redesign the work system — but that relationship is not yet empirically established.
• Use vs. stewardship. 2025-lee-generative-ai-critical-thinking suggests that more confident GenAI use may mean less critical thinking, not better adoption. The wiki should separate adoption volume from adoption quality.
• Use vs. task fit. 2026-dellacqua-jagged-technological-frontier shows the same AI tool can raise or lower performance depending on task location relative to the frontier. Adoption metrics that only count usage can therefore hide misdeployment.
• Floor vs. ceiling. 2025-brynjolfsson-generative-ai-at-work finds the largest productivity and quality gains for lower-skilled and less experienced workers, while 2026-kim-mapping-ai-into-production finds firm-level gains concentrated in the upper tail. Individual skill diffusion and firm-level production discovery may be different mechanisms.
• Discovery vs. worker fit. 2026-kim-mapping-ai-into-production suggests firms can improve adoption by seeing more AI use-case patterns; 2026-lee-making-invisible-visible suggests adoption can still fail when identified use cases violate worker autonomy, task meaning, communication norms, or accountability. These are complementary constraints, but the sequencing is open: discover use cases first, or diagnose worker-centered fit first?
• Observed use vs. organizational adoption. 2025-handa-economic-tasks-ai measures actual Claude.ai conversations, which is stronger than pure exposure modeling but still not the same as durable adoption inside an accountable work system. A task can appear in AI usage data without being formally redesigned, governed, or tied to performance outcomes.
• Task demand vs. adoption quality. 2025-agarwal-what-work-ai-doing shows which task characteristics correlate with observed use, but not whether that use improves quality, learning, stewardship, or downstream performance. High-friction cognitive work may attract AI precisely where verification and accountability matter most.
• Readiness breadth vs. intervention focus. 2021-johnk-ready-or-not-ai-comes defines readiness broadly across strategic alignment, resources, knowledge, culture, and data; 2026-kim-mapping-ai-into-production identifies a narrower causal intervention that improves firms’ discovery of AI use cases. The open question is whether broad readiness assessments predict adoption success beyond targeted mapping support, or whether they mainly name the organizational conditions that make mapping interventions stick.
• Attitude vs. opportunity structure. 2026-reich-work-design-multidimensional-ai-threat suggests AI adoption may depend more on skill variety and autonomy than on undifferentiated AI threat. That complicates adoption programs built mainly around persuasion, communication, or training.
• Literacy vs. receptivity. 2025-tully-ai-literacy-receptivity suggests ai-literacy and ai-receptivity should be modeled as separate constructs: knowing more about AI can reduce some forms of willingness to use it when lower literacy makes AI feel magical or awe-inspiring. That creates a tension between adoption programs that want more AI use and governance programs that want better-calibrated use.
• Reverse mentoring vs. risk governance. 2025-kellogg-novice-risk-work suggests junior or novice AI users may help seniors experiment while still recommending underpowered mitigation tactics. Adoption systems need evidence on when peer teaching is enough and when deployer-level scaffolding is required.
• Private usefulness vs. public signal. 2026-almog-barriers-ai-adoption shows workers may privately benefit from AI recommendations but publicly avoid them when reliance is visible to evaluators. That creates a design tension between making AI use transparent for accountability and making it safe enough that workers do not hide useful reliance.
• Capability vs. moral permission. 2025-friis-performance-principle suggests adoption can remain blocked even after performance concerns are removed. That is a different mechanism from poor mapping, low readiness, low literacy, or image concerns, and it may require redesigning the human role rather than merely improving the AI.
• Tool access vs. managerial knowledge process. 2025-hautamaki-fully-leveraging-ai-b2b-sales suggests wide AI accessibility in B2B sales is not sufficient when managers lack the capability to connect data, context, experimentation, use-case sharing, and cross-functional knowledge transfer. The open question is whether these managerial capabilities predict realized adoption and performance beyond broader ai-readiness factors.
• Technical improvement vs. behavioral adoption. 2025-schweitzer-leading-ai-adoption argues that better technical capability and principled HCAI design are insufficient if designers, employees, and leaders behave predictably around AI mistakes, threats, sunk costs, and uncertainty. The open question is which behavioral interventions improve adoption quality without becoming mere persuasion to accept systems that should be redesigned or rejected.
• Compliance vs. championing. 2018-fugate-change-compliance-championing implies that adoption dashboards may confuse minimal compliance with active support. The open question is which AI rollout indicators capture challenge appraisal, discretionary problem solving, and peer championing rather than mere tool use or policy adherence.
• Pilots vs. learning systems. 1997-goh-benchmarking-learning-capability implies that AI experimentation only compounds when organizations can transfer knowledge, discuss failures, and remeasure progress. The open question is whether learning-capability measures predict which AI pilots scale into work redesign and which remain local anecdotes.