Performance Expectancy Does Not Predict AI Adoption

In a structural model of 523 U.S. adults, performance expectancy did not significantly predict AI adoption intent (β = -.028, p = .791 in the N=523 sample). The vendors who keep pitching better benchmarks are pitching the wrong lever.

What the Model Found

The AIRS structural model explains 85.2% of variance in behavioral intention to adopt AI tools. Eight factors were tested. Three are supported. Five are not.

Performance expectancy (PE) is the belief that using an AI tool will improve job performance. It is the loudest pitch in enterprise AI sales. It is also the weakest predictor in this data.

The three paths that did significantly predict adoption intent in this sample:

• Price value (PV): β = .505, p < .001 — the strongest predictor, by a wide margin. Users need to believe the tool is worth what they give up: money, time, disruption, the cognitive cost of learning something new.
• Hedonic motivation (HM): β = .217, p = .014 — whether using the tool is enjoyable or interesting. Second strongest.
• Social influence (SI): β = .136, p = .024 — whether people whose opinions matter to the user are already using AI tools. Third strongest.

Five factors did not significantly predict in this sample: performance expectancy, effort expectancy, facilitating conditions, habit, and AI trust. Trust is a separate discussion: the effect was marginal and underpowered at this sample size, not absent. That matters for interpretation, but it does not change the main finding on PE.

Why This Is Counter-Intuitive

Performance expectancy is the backbone of most enterprise AI positioning. Every vendor deck leads with capability. Every ROI model starts with “this tool will make your team X% more productive.” The implicit assumption is that if workers believe the tool improves their performance, they will use it.

The data does not support that assumption in this sample.

This is not a new hypothesis. UTAUT2 (the parent framework AIRS extends) already found that PE effects are weaker in voluntary use contexts. What AIRS adds is a direct measure under current AI conditions, with a sample collected during the period October through November 2025, when AI tool exposure was high but adoption was uneven across job roles and experience levels.

The finding survives a model that explains 85.2% of variance in adoption intent. If performance expectancy were a meaningful driver, the model would have found it.

What Actually Drives Intent

Price value is doing most of the work. The construct measures whether users perceive the benefits of AI adoption as worth the personal costs. That is a fundamentally different framing than “this tool is capable.”

Capability answers the question: can the tool do this? Price value answers the question: is this worth it for me?

Those are different questions. Enterprise AI programs tend to invest heavily in answering the first and underinvest in answering the second.

Hedonic motivation (enjoyment) and social influence (peer adoption signals) are also supported. Neither of them is a capability claim. Enjoyment is about the experience of using the tool. Social influence is about whether the people around you are already using it.

Taken together, the supported predictors point toward a consistent pattern: adoption intent is driven by personal experience and social context, not by performance specifications.

The three predictors are not abstract constructs. Each one names a specific design problem.

Three Predictors, Three Design Inputs

Price value is not a licensing negotiation. It is a user-side calculation that happens whether or not your program accounts for it. The question users answer, consciously or not, is: is this tool worth what I give up? What they give up is concrete: time to learn the new workflow, disruption to the familiar one, the cognitive overhead of working differently, and the risk of looking less competent during the transition period. If your deployment has not addressed that calculation directly, price value stays low regardless of what the benchmark says. The tool can be genuinely capable and still fail this test.

Hedonic motivation is about whether the experience of using the tool is worth having, independent of the outcome. Enterprise programs almost never design for this. The signal appears in second-session rates: does the user come back after the initial onboarding? Low hedonic motivation does not announce itself as “the tool is unpleasant.” It appears as abandonment after week one, low voluntary usage in populations where the tool is not mandated, and support requests that are functionally requests to opt out. In this sample, hedonic motivation was the second strongest predictor of adoption intent (β = .217, p = .014 in the N=523 sample). That is not a UX detail. It is a structural predictor.

Social influence is peer visibility, not executive endorsement. Your governance framework and leadership announcement do not move this predictor. What moves it is whether the person in an adjacent role is visibly using the tool and talking about it. Programs that deploy broadly and expect adoption to propagate on its own work against this predictor. Programs that seed visible peer cohorts first and instrument the social signal work with it. Deploying to twenty high-visibility users before expanding to two thousand is not a slower rollout. It is a different adoption design, and the data supports it.

What This Means for Enterprise Programs

If your AI program is designed to drive adoption by demonstrating capability, you are working on the wrong input.

The implication is not that capability is irrelevant. A tool that does not work will not be adopted regardless of how enjoyable it is. But capability appears to be a floor, not a driver. Once a tool crosses a basic competence threshold, additional capability claims do not move adoption intent in this sample.

What does move it: making the personal cost-benefit calculation favorable (price value), designing for an experience worth having (hedonic motivation), and creating visible social proof within the user’s peer network (social influence).

At the next vendor meeting that opens with benchmark slides, there is one question worth asking before the meeting ends: “What is your evidence that users who believe this tool improves their job performance adopt it at higher rates than users who do not?” The AIRS data puts the performance expectancy path at β = -.028, p = .791 in the N=523 sample. Not statistically significant. The benchmark answers a question that does not predict adoption. That is worth knowing before signing a renewal.

The budget case for your portfolio is not that the tools you are deploying are more capable than the alternatives. The case is that your deployments are designed to make the personal cost-benefit calculation land in favor of adoption, that the experiences are worth returning to, and that adoption is visible inside the peer networks of the users you are trying to move. Those are the three levers the data supports. Capability was not one of them.

Three things to do differently this week:

First, if a vendor renewal is on your calendar this quarter, pull two data points before the meeting. Your internal adoption rate by business unit, with the variance across units visible. And any signal you have on whether users in the low-adoption units believe the tool is worth the disruption cost. The vendor will bring benchmark slides. Those slides answer whether the tool can do the task. Your two numbers answer whether it is being used and why. The budget defense belongs to the second set.

Second, if adoption telemetry in your portfolio only measures usage rate and task completion, add a user-perceived value signal before the next portfolio review. The question is not complicated: do users believe this tool is worth the disruption cost? If you do not measure it, you will not know why adoption stalls when it does, and you will not have a defensible answer in the funding-defense meeting when it does.

Third, at the next team lead meeting, identify the team with the highest voluntary adoption rate and ask them what changed. The answer is unlikely to be “we showed them better benchmarks.” The answer will name the experience, the peer network, or the value calculation. That answer is a design input for the next deployment, not a testimonial.

The research anchor: Dr. Fabio Correa’s DBA dissertation (Touro University Worldwide, defended April 2026) tested eight UTAUT2-based predictors of AI adoption intent across a sample of 523 U.S. adults recruited through Centiment, topic-blinded. The model accounts for 85.2% of variance in behavioral intention. The instrument is publicly available at airs.correax.com.

Funding Meeting Brief

References

• AIRS instrument — AI Readiness Scale — correax.com, 2026-04-01
• AIRS research repository — GitHub, 2026-04-01