In our UXR practice at Akraya, we study AI adoption across HITECH, Software Products, and Internet companies. One pattern keeps emerging: AI products fail in predictable ways.
Not random ways. Predictable ones. With observable signals that appear in user research weeks or months before a product launch.
We've organized these failure patterns around what we call the Three Pillars of AI Adoption Risk. Understanding them and knowing how to identify the research signals for each is the core of what Akraya's UXR Decision-Confidence Framework provides to product and technology leaders.
The core question: Can users rely on it?
The risk: AI behavior is non-deterministic. Unlike traditional software that produces the same output from the same input, AI models produce variable outputs, sometimes subtly different, sometimes dramatically so. This variability is fundamentally threatening to professional users who need to stake their credibility on the outputs they act on.
The risks are specific: output inconsistency across sessions; hidden critical errors inside 'mostly correct' responses; accuracy that looks strong in benchmarks but fails under real-world complexity.
THE RESEARCH SIGNAL: Users are spending more time verifying AI output than completing the task.
When we observe this pattern in a usability session, it is a direct predictor of abandonment. Users don't quit immediately. The tool is novel, the expectations are high, and there's a natural inclination to give it a fair chance. But over repeated sessions, the verification overhead accumulates. The tool that was supposed to save time is costing time. And users quietly stop integrating it into their primary workflows.
WHAT TO LOOK FOR IN RESEARCH: Watch for verification behaviors, excessive copy-pasting to other tools for double-checking, running manual processes alongside the AI output, visible pausing, and scrutiny of AI-generated content before acting on it. These are the early signals of Pillar 1 failure.
The core question: Does it fit established workflows?
The risk: AI tools that require context-switching, produce outputs in unfamiliar formats, or disrupt established professional processes will see users revert to manual workflows even when the AI's outputs are technically accurate. The problem isn't the model. It's the fit.
The risks are disruption to expert mental models; friction introduced into trusted, high-stakes workflows; and requirements for context-switching or rework that erode efficiency gains before they can be realized.
THE RESEARCH SIGNAL: Users reverting to manual workflows after initial trial.
This is one of the most common patterns we see in enterprise AI UXR. The initial enthusiasm is real when users try the tool, appreciate the novelty, and engage with it earnestly. But as the novelty fades and real workflow pressure returns, users return to the processes they know. Not because the AI was wrong, but because integrating its outputs created more friction than it removed.
What to look for in research: Study how AI outputs map to existing workflow steps. Where does the AI's output format create additional work for the user? Where does it require reformatting, reverification, or translation before the user can act on it? These friction points are invisible to accuracy benchmarks but immediately visible in a well-designed usability study.
The core question: Will users trust it?
The risk: Trust is the most fragile of the three pillars because once broken, it is the hardest to rebuild. For professional users in high-stakes environments, a single high-impact AI failure a hallucination that damages their credibility, or a confidently wrong answer that leads to a poor decision can permanently alter their relationship with the tool.
The risks are hallucinations that create professional liability, poor error transparency or recovery paths that leave users unable to identify when they can and cannot trust the AI; an overconfident tone in uncertain outputs.
THE RESEARCH SIGNAL: After one high-impact failure, users disengage permanently.
This is the most catastrophic failure mode because it has no recovery path within the same product version, and sometimes no recovery path at all. In our research with senior developers at a major technology company, we found that most had encountered a poor early AI experience, and the majority had simply never returned, even after the company invested significantly in model improvements.
Trust, once broken with a professional user, is notoriously difficult to rebuild. The 'trust elasticity' of AI is fundamentally different from that of traditional software and product teams, and those that don't account for this are gambling their AI investment on a first impression.
WHAT TO LOOK FOR IN RESEARCH: The moment of 'trust failure' in a session. This is usually a clearly visible change in posture, a verbalization of doubt ('I'm not sure I can actually use this for real work'), or a shift from active engagement to passive skepticism. Capturing these moments and understanding what triggered them is the most valuable output of pre-launch AI UXR.
At Akraya, we use the Three Pillars as the backbone of our UXR Decision-Confidence methodology. Each pillar maps to specific research approaches:
Functional Reliability → Usability testing with task completion measurement and verification behavior tracking
Workflow Fit → Contextual inquiry and workflow observation with real users in their real environments
User Trust → Longitudinal trust studies, critical incident technique, and trust calibration measurement
The output is not just a user research report. It is an executive-level risk assessment: which pillar is most at risk, what the research signals indicate, and what specific decisions need to be made and by whom before launch.
That's how UXR transforms an AI launch from a speculative gamble into a calculated strategic move.
→ Download the Full Framework Free Whitepaper [https://www.akraya.com/insights/whitepaper/how-ux-research-reduces-ai-product-risk-before-launch]
→ Watch Akraya's UXR Team Explain This Live Free 50-Min Webinar [https://www.akraya.com/insights/webinars/webinar-ai-product-risk-ux-research]