Psychological Safety → Adoption → Self-Efficacy → ROI
The research logic behind the Human Layer Framework.
This page expands the Human Layer Framework and shows the logic underneath it: how psychological safety, adoption, self-efficacy, and ROI fit together, and why we treat AI adoption as a complex skill-acquisition problem rather than as a software rollout.
1 · The empirical baseline: high interest, low adoption, near-zero ROI
Two independent research programs frame the problem:
- MIT NANDA, The GenAI Divide (2025). An analysis of over 300 public AI initiatives, 52 organizational interviews, and 153 senior leaders concludes that about 95% of organizations are getting no measurable P&L return from their generative AI projects. Just 5% of integrated AI pilotsgenerate meaningful business value; the rest stall in experimentation or limited pilots. The report labels this pattern the “GenAI Divide.”
- Wang et al. (2026), International Journal of Information Management. In a survey of organizations adopting large language models, the authors find that around 75% had experimented with LLMs, but only 9% had adopted them widely. They describe this as high interest but low adoption and show that LLM implementation is a multi-stage, socio-technical process rather than a single decision.
The methodology below is our answer to that specific translation problem.
2 · Stage one: psychological safety as the gateway
What psychological safety is
Psychological safety is a construct from organizational psychology. In Edmondson's original work it is defined as a shared belief that a team is safe for interpersonal risk-taking: safe to ask questions, admit errors, and propose ideas without fear of embarrassment or punishment. A large meta-analysis shows that psychological safety predicts information sharing, learning behaviors (asking for help, reporting errors), and creativity — the same behaviors people need during early AI adoption: asking basic questions, showing incomplete prompts, exposing misunderstandings.
What the AI-specific evidence shows
A 2025 study titled Safety First: Psychological Safety as the Key to AI Transformation surveyed 2,257 employees in a global consulting firm, measuring psychological safety, AI adoption (yes/no), and usage frequency and duration among adopters. The findings:
- Psychological safety reliably predicted whether employees adopted AI at all. A one-unit increase was associated with roughly a 30% increase in the odds of adoption, even after accounting for role, experience, and region.
- Among employees who had already adopted AI, psychological safety did not significantly predict how often or how long they used the tools.
We draw two methodological conclusions:
- Psychological safety is a gateway condition. It determines who is willing to step into AI experimentation at all.
- It is not the main driver of depth of use.The mechanisms that govern “who starts” are not the same as those that govern “who persists and gets value.”
Our work treats psychological safety as the non-negotiable precondition for any adoption effort. We do not run capability-building programs in environments where people cannot safely fail in public.
3 · Stage two: adoption as a socio-technical design problem
The Wang et al. (2026) study uses Rogers' innovation-decision process and activity theory to analyze how organizations implement LLMs across five stages — agenda-setting, matching, redefining & restructuring, clarifying, and routinizing. Across those stages, their interview data surfaces recurring contradictions: tools that technically work but don't match frontline workflows; data, privacy, or security tensions that stall pilots; and a lack of clear ownership and governance over time. Their recommendations include tiered rollouts, trial/decision-support platforms where users can safely experiment, agile modular architectures, and accountable governance.
Within our framework, once psychological safety is established, we look at adoption as a fit and flow problem:
- Fit — Is the AI system matched to a job people actually have, with clear, local success criteria?
- Flow — Is it woven into the way the work actually happens, including constraints, hand-offs, and existing tools?
We use the Wang et al. stages to diagnose where organizations are stuck and design rollout sequences that match real activity systems instead of idealized process maps.
4 · Stage three: the sufficiency gap and self-efficacy
That gap — adoption has happened, but meaningful return has not — is what we call the sufficiency gap. The core claim of our methodology is that this gap is, in large part, a self-efficacy gap.
What self-efficacy is and why it matters here
Albert Bandura defined self-efficacyas people's judgments of their capabilities to organize and execute the courses of action required to attain specific performances. It predicts whether people initiate a behavior, how much effort they invest, and how long they persist against difficulty. In technology and training research, higher self-efficacy is associated with greater willingness to use new systems, less anxiety, and better transfer of training into real work. A lineage of technology-specific and AI-specific self-efficacy scales shows that self-efficacy is measurable (with validated instruments) and buildable (it changes in response to experience and training design).
Why we treat AI adoption as complex skill acquisition
AI adoption is not “turning on a feature.” It is learning to translate messy domain problems into prompts, interpret uncertain outputs, chain tools and workflows under time pressure, and do all of that in a visible, sometimes evaluative social context. Methodologically, this is identical to learning any complex cognitive skill under uncertainty.
My own thesis work at the University at Albany examined self-efficacy, attention, and individual differences in a virtual complex simulation environment: a high-fidelity, air-traffic-control-style simulator used for training. The main findings:
- Mind wandering (off-task thought) reliably predicted lower post-training self-efficacy, even when performance was controlled.
- Trait emotional stability (neuroticism) did not significantly predict self-efficacy in that setting.
- Training sequence (easy-to-hard vs hard-to-easy) did not eliminate the negative impact of attentional lapses.
We therefore treat AI capability-building as a self-efficacy engineering problem: designing experiences so that learners come away believing, accurately, “I can do this.”
5 · Stage four: the Adoption-to-ROI Loop
The final stage links the human constructs above to the business outcomes organizations care about. At a minimum we track four families of measures:
- Psychological safety — adapted from Edmondson-style scales, focused on AI-related interpersonal risk.
- AI self-efficacy — domain-specific items capturing confidence in using AI for real tasks (informed by existing computer/AI self-efficacy instruments).
- Behavioral adoption — who is using AI, on what tasks, and with what depth (one-shot prompting vs multi-step workflows).
- Workflow-level outcomes — cycle times, error rates, rework, token usage per unit of output, and other process metrics tied to specific AI-enabled workflows.
The loop runs as:
Establish baselines for safety, self-efficacy, behaviors, and workflow outcomes.
Identify the bottleneck: safety, fit/flow, self-efficacy, or workflow design.
Design targeted interventions — training redesign, sandbox experiences, workflow changes, leadership signals — aimed at that bottleneck.
Assess changes in self-efficacy and behaviors, track downstream workflow and token/effort efficiency, then iterate and expand to new workflows.
Over time, the loop produces a data-backed picture of how human factors and workflow design interact in your specific context, a progressive shift from surface-level use to deep integration, and a traceable line from psychological interventions to concrete business metrics.
6 · What is established vs what is proposed
We are explicit about the line between what the literature already shows and what we are adding as a working theory.
Established by current research
- Most organizations exhibit high AI interest but low adoption and minimal ROI at scale.
- Psychological safety predicts whether people adopt AI, but not how deeply they use it once adopted.
- AI implementation is a multi-stage, socio-technical process with recurring contradictions at each stage.
- Self-efficacy robustly predicts behavior, effort, and persistence, and can be measured and built, including in technology domains.
- Poorly designed training and attentional lapses undermine self-efficacy in complex skill acquisition.
Proposed by this framework
- The sufficiency gap between adoption and ROI is largely a self-efficacy gap.
- AI adoption should be treated as a complex skill-acquisition and training-transfer problem, not just a technology rollout.
- A structured Adoption-to-ROI Loop, centered on measuring and building self-efficacy, is a practical way to turn adoption into return while generating the longitudinal evidence the field currently lacks.
I am Mario, trained in industrial–organizational psychology, and my work at Rebel Minds takes these research strands and turns them into designs, diagnostics, and experiments that live inside real organizations. The Human Layer Framework is deliberately framed as a working theory: precise enough to test, transparent enough to critique, and flexible enough to evolve as new data arrives.
7 · How this connects to the Framework
The Human Layer Framework on the main page is the narrative: Psychological Safety → Adoption → Self-Efficacy → ROI. This methodology page is the scaffolding underneath it — why safety is the gateway, how adoption depends on socio-technical fit and flow, why the sufficiency gap is fundamentally a self-efficacy problem, and how the Adoption-to-ROI Loop turns those constructs into measurable change.
También disponible en español: La Metodología
Last updated: June 2026