The Human Layer Framework
A working theory of why AI adoption succeeds or fails — and the science underneath it.
The tools were never the problem. Organizations have spent two years buying AI that works and watching it go unused. The bottleneck is not the technology. It is the human layer — the set of beliefs, conditions, and capabilities that determine whether people will actually take up a new tool and get a return from it. This is a working theory of that layer, built on published behavioral research and the discipline of operationalizing it.
The problem, stated plainly
The gap between AI's promise and its actual use is now well documented. A 2025 report from MIT's NANDA initiative estimates that 95% of organizations are getting no measurable financial return from their generative AI projects, despite billions in investment (MIT NANDA, 2025). A 2026 study in the International Journal of Information Management reports that while around 75% of organizations had experimented with large language models, only 9% had adopted them at scale. The authors name this pattern high interest, low adoption (Wang et al., 2026).
Most responses to this gap target the technology — better models, better integration, better prompts. But the evidence points elsewhere. The failures cluster at the human layer: people who don't feel safe trying, tools matched to the wrong jobs, capabilities that never transfer into real work, returns that never materialize even after people start.
The Human Layer Framework is a structured account of where, and why, adoption breaks at that layer — and what to do about each break.
The first principle: psychological safety is the precondition
Before a person will adopt an AI tool, they have to be willing to do something uncomfortable in front of others — ask a basic question, produce a bad first draft, admit they don't understand the output. That willingness has a name in organizational psychology: psychological safety, the belief that you can experiment and make mistakes without penalty.
The evidence here is strong and recent. In a study of 2,257 employees in a global consulting firm, psychological safety reliably predicted whether employees adopted AI tools at all; each one-unit increase in psychological safety increased the odds of adoption by about 30%, and this effect held across experience levels, seniority, and geographic region (Reich et al., 2025).
Psychological safety is the gate. Without it, adoption does not begin. This is the foundation of the framework — and the part most rollouts skip entirely, because it is not a technology problem and most implementation playbooks have no tools for it.
The overlooked finding: adoption is necessary but not sufficient
Here is the result that reorganizes everything. The same study found that psychological safety predicted whether employees started using AI — but it did not predict how often or how long they used it once they had begun. In other words, psychological safety is a necessary condition for adoption, but it is not sufficient to drive the deep, sustained use that actually produces return.
This is why so many rollouts described as successful still produce no ROI. The team adopted the tool — leadership sees the logins, the licenses, the early enthusiasm — and concludes the job is done. But usage stays shallow, the capability never deepens, and the return never arrives. The organization solved the adoption problem and assumed it had solved the return problem. It hadn't. Those are separate.
The research that proved safety drives adoption was explicit that it did not measure what drives the second stage. The authors named, among the variables they did not control for, individual self-efficacy. That is the open seam. And it is where this framework makes its contribution.
The contribution: self-efficacy is what carries adoption to return
If psychological safety is the socialpermission to start, what governs whether a person becomes genuinely capable — using the tool well enough, persistently enough, to produce a return? That is not a question about interpersonal climate. It is a question about an individual's belief in their own capability with a specific, difficult task. In behavioral science, that belief is self-efficacy.
Adopting AI is not a one-time decision. It is the acquisition of a complex skill under uncertainty — precisely the conditions under which self-efficacy governs whether people persist through early failure, recover from errors, and transfer a new capability into the actual work. Decades of research have established two things that matter here: self-efficacy is measurable and it is buildable, including in technology domains — from classic computer self-efficacy scales to newer AI-specific self-efficacy instruments.
But it is not built by accident, and this is the part that makes the framework actionable rather than merely diagnostic. Self-efficacy has antecedents that can be engineered. Individual differences — how people differ in emotional stability and disposition — shape how well they attend during training. Attention is what allows self-efficacy to form in the first place. And training design is the lever that gets attention there. Design the training badly and attention scatters; without attention, self-efficacy never develops; without self-efficacy, adoption never becomes return. Design it well — accounting for who the learners actually are — and you build the belief that carries people across the sufficiency gap.
This is the lens I am trained to bring. I am Mario, trained in industrial–organizational psychology, and my master's research at the University at Albany examined self-efficacy, attention, and individual differences during training and skill acquisition in a virtual complex simulation environment1 — the science of how people actually learn to use difficult new things, and how training can be designed so that learning holds. AI adoption is that same problem in a new domain. The claim of this framework is therefore specific and testable: the sufficiency gap — the space between adopted and profitable — is largely a self-efficacy gap; self-efficacy is something we know how to measure and build; and the building is a training-design problem we know how to work.
The structure
Putting it together, the human layer has a predictable architecture, and rollouts fail at identifiable points within it:
Do people feel safe enough to try and fail in view of others? If not, adoption never starts. (Established: Reich et al., 2025.)
Was the tool matched to a real job people actually have (fit), and woven into how the work actually happens (flow)? Mismatches here stall adoption even when safety is present. (Mapped to the organizational failure stages in Wang et al., 2026.)
Adoption has happened — and return still hasn't. This is the space where shallow, tentative use never deepens into confident, competent practice. The working claim of this framework is that this gap is largely a self-efficacy gap, and that self-efficacy can be measured and built.
The proprietary close: a self-efficacy-driven cycle of measure → diagnose → intervene → re-measure that works the sufficiency gap directly, turning adoption into demonstrated return and generating, over time, the very evidence the field is missing.
Why this is hard to copy — and easy to verify
This framework is published openly because the protection was never secrecy. The vocabulary can be repeated by anyone. What cannot be repeated is the practice underneath it: diagnosing and building self-efficacy during complex-skill acquisition is a trained discipline, not a slide. And the framework is bound to a specific scientific background that a restatement cannot transfer.
It is also built to be checked. Every empirical claim cites the source. Every proposition is marked as a proposition. The boundary between what the research proves and what this framework proposes is kept visible — not as a hedge, but because that boundary is exactly what separates a method from a sales pitch. The work tests its claims; it does not assert them.
Where this goes
The tools will keep improving. That was never the constraint. The organizations that get a return from AI will be the ones that treat the human layer as the real site of the work — safety first, then fit and flow, then the disciplined pursuit of return through self-efficacy. That is the thesis. The rest of this work is the practice of making it usable.
Frequently asked questions
Why do most AI rollouts fail?
Most AI rollouts fail at the human layer, not the technology. A 2025 MIT NANDA report found about 95% of organizations get no measurable financial return from generative AI. The tools work; what is missing is the psychological safety and self-efficacy that turn access into sustained, value-producing use.
What is the human layer of AI adoption?
The human layer is the set of beliefs, conditions, and capabilities that decide whether people actually take up an AI tool and get a return from it — chiefly psychological safety and self-efficacy. It sits below the model, the tooling, and the workflow, and it is where most rollouts quietly break.
Does psychological safety improve AI adoption?
Yes. In a study of 2,257 employees, each one-unit increase in psychological safety raised the odds of AI adoption by roughly 30%, across seniority and region (Reich et al., 2025). But the same study found safety did not predict how deeply or how long people used AI — adoption is necessary, not sufficient.
What is the sufficiency gap in AI adoption?
The sufficiency gap is the space between an AI tool being adopted and it actually producing return. Teams log in and experiment, but usage stays shallow and ROI never arrives. This framework proposes the gap is largely a self-efficacy gap — a problem of capability belief that can be measured and built.
What is self-efficacy in AI adoption?
Self-efficacy is a person's belief in their own capability to perform a specific, difficult task — here, using AI well enough to produce a return. It governs whether people persist through early failure and transfer the skill into real work. Decades of research show self-efficacy is both measurable and buildable.
How do you turn AI adoption into ROI?
Through the Adoption-to-ROI Loop: measure psychological safety, self-efficacy, behavior, and workflow outcomes; diagnose the bottleneck; intervene on it; then re-measure. Repeating the cycle builds the self-efficacy that carries shallow adoption into deep, value-producing use — and generates the longitudinal evidence the field currently lacks.
También disponible en español: El Marco de la Capa Humana
Sources
- Arredondo, M. L. (2022). The neurotic wandering mind and self-efficacy during training.Master's thesis, University at Albany, State University of New York. https://doi.org/10.54014/DKAX-FS1S
- Reich, A., Wolfe, D., Price, M., Choe, A., Kidd, F., & Wagner, H. (2025). Safety First: Psychological Safety as the Key to AI Transformation.
- Wang, X., Zhong, W., Huang, K., & Liang, B. (2026). High interest but low adoption: Navigating organizations' journey towards generative artificial intelligence implementation. International Journal of Information Management, 87, 103009. https://doi.org/10.1016/j.ijinfomgt.2025.103009
- Challapally, A., Pease, C., Raskar, R., & Chari, P. (2025). The GenAI Divide: State of AI in Business 2025. MIT NANDA. — source of the 95%-of-organizations-see-no-return figure.
- Foundational self-efficacy and training-transfer literature (Bandura and successors), applied here to AI adoption as complex-skill acquisition.
1 Arredondo (2022), above — original research on self-efficacy, attention, and individual differences during training in a virtual complex simulation environment.
Last updated: June 2026