5 Challenges UX Designers Face When Scaling AI in Enterprises and How to Address Them

Distrust of AI Conclusions

Because of users’ distrust of AI conclusions, users rarely accept even highly accurate outputs without supporting evidence. To be credible, AI systems must provide additional information such as confidence scores, reasoning steps, or explanations of how the AI generated specific results. However, this creates a design trade-off: the greater the transparency, the more technical and complex users perceive a product to be.

The design challenge is to balance credibility with simplicity. One effective approach is progressive disclosure, which keeps core user interfaces clean and at a high level, but allows users to drill down when they want to. A financial-forecasting tool, for instance, might keep the main, high-level dashboard clean, but let users click to view specific projections or confidence scores and see which data points influenced the results.

2. Clarifying AI Capabilities and Limitations

Users often misunderstand what AI technology can and cannot do. This leads to unrealistic expectations, frustration, and reduced adoption. The solution is making AI’s potentials and limitations visible from the start by providing the following:

• onboarding tutorials, presenting specific AI scenarios
• contextual tips and progressive disclosure of advanced AI features
• system controls and suggested inputs that demonstrate AI capabilities
• previews of system outputs that show the range of possible results
• role-tailored explanations for each stakeholder—for example, executives, AI governance, affected users, business users, regulators or auditors, and developers. These explanations can be calibrated according to their technical depth, format—for example, ToolTips, dashboards, or audit logs—and their timing—that is, real time versus retrospective.

To strengthen users’ trust, be explicit in stating that AI is probabilistic and might not always perform perfectly. Don’t hide the AI’s limitations. Communicate them by doing the following, so users can calibrate their expectations and make informed choices:

• Reflect uncertainty in language. Use wording that matches the appropriate level of confidence—that is may, might, or likely—rather than stating absolutes.
• Match numbers to reality. Don’t show greater precision than the system has. Use rounded estimates when appropriate—for example, “about 5 minutes,” not “4.83 minutes.”
• Report performance in plain terms. Share meaningful indicators regarding accuracy, false positive/negative rates, or context-specific levels of performance. Present these in terms that users can understand and explain what they mean for the user.
• Call out the system’s limitations and risks. Tell users where a model might struggle—for example, with missing data or rare cases—and adjust alerts’ tone to reflect the stakes—for example, a gentle heads-up or an urgent warning.

While the probabilistic nature of AI is inherent in how most models operate—producing outputs that are based on statistical patterns rather than deterministic rules—the aim of active research is reducing this unpredictability. Techniques such as deterministic inference pipelines, fixed random seeds, and controlled hardware environments are under development to make model behavior more stable and reproducible. These advances are especially valuable within enterprise contexts where auditability and consistency are critical.

3. Driving User Adoption

User adoption is as much about psychology as technology. In many cases, users resist AI from the start. During early research on the implementation of AI systems, employees often obscure information about workflows, avoid mapping steps in plain terms, downplay what AI could automate, or frame tasks as too complex or uniquely human to be implemented using AI.

Users’ resistance creates a familiar paradox: AI promises easier, more efficient, even more satisfying work, yet the very people AI is supposed to help often push back against its implementation. This reaction reflects the reality that we are still in the early stages of adopting AI at scale, and many users are unsure how to engage with AI.

From our experience, two approaches to driving adoption are consistently helpful:

1. Lead with outcomes, not AI. Avoid using terms like AI, ML (machine learning), and automation at first. Frame early conversations around clear benefits of AI such as fewer routine tasks, faster decision-making, and higher-quality work. This small shift in language can reduce users’ anxiety, open up more honest conversations about AI, and keep the focus on what matters: making work better for people.
2. Prioritize high-utility use cases. Ship features that solve concrete, repetitive problems, not gimmicks that destroy users’ trust. To provide a concrete example, AI-generated LinkedIn response suggestions often feel like a gimmick, while an integrated grammar checker provides clear everyday value.

4. Ensuring Strong Solution Performance

Performance isn’t just about solving latency issues; it’s also about ensuring the high quality of outputs.

Within many enterprises, legacy systems and analog processes still dominate. Data is scattered across old databases and spreadsheets or just lives in people’s heads. Decision-making often relies on individuals’ judgment rather than structured information.

But the hard truth is that the quality of AI outputs is directly tied to the quality of the input data. While high-quality data doesn’t automatically guarantee strong model performance, poor data quality almost always degrades it. Ensuring data accuracy, completeness, and consistency remains the foundation of any reliable AI system.

On one project, for example, we needed to implement AI to calculate how much product could fit into in-store fixtures—a task that required precise product dimensions, both with and without packaging. But we hit a wall because that data was not centralized or easily accessible. As a result, the project stalled.

The lesson is simple: before AI can deliver value, we must do the unglamorous work of collecting, cleaning, digitizing, and standardizing data. Otherwise, even the most advanced AI system will not perform as expected.

5. Mitigating Security Threats

Organizations often see eliminating security risks as a purely technical problem, but for AI systems, it also presents design problems. Prompt injection, data leakage, unsafe outputs, and excessive autonomy can all emerge from the ways in which users interact with a product. If such issues remain unchecked, they can undermine user trust, adoption, and compliance. To address this issue, embed safeguards directly into product experiences, as follows:

• Define clear boundaries. Make scope limits visible and enforceable—for example, drafting an email message, but not sending it automatically. Use sandboxed workflows, in which AI suggestions require human review before execution.
• Guide users’ inputs. Provide templates or examples that steer users toward creating safe prompts. Provide inline warnings or validations when users are entering sensitive or risky data.
• Reinforce intentional user control. Provide permission dashboards, least-privilege defaults, and progressive disclosure so sensitive actions require clear user consent.
• Enable monitoring and reporting. Provide human-readable logs of what has run, using what data, and why. Include one-click abuse reporting and indicate when data has been anonymized or minimized.

The goal is to balance protection with efficiency: too many restrictions lead to users’ creating workarounds, while too few create vulnerabilities. Calibrate safeguards to eliminate risks by creating lightweight checks for low-impact actions and stronger reviews for high-risk operations.

Real Challenges We’ve Encountered and How We Solved Them

Now, let’s look at some of the real challenges we’ve encountered on our projects and how we solved them.

A Major US Logistics Provider

Problem: The company’s freight-bidding process was fragmented. Opportunities weren’t centralized, operators relied on subjective judgment, carrier reliability was not evaluated consistently, bidding rounds were unlimited, and decisions moved slowly.

Solution: We built a centralized bidding platform that aggregated all opportunities in one place, introduced carrier-evaluation metrics, enforced round limits to prevent drawn-out cycles, and adjusted workflows based on bid size so operators could prioritize high-value opportunities.

Result: The company saw a significant increase in both the number of bids that were submitted and the average monetary value per bid.

A Global Beauty Company

Problem: The manual capacity of teams to evaluate and respond to risks constrained the company’s supply-chain risk-management process. There was little visibility into how or why decisions were made. Execution of mitigation strategies such as transport-mode changes or stock redeployments across multiple systems was slow and error prone.

Solution: We developed a solution that provided full visibility into the state of all locations and distribution centers. The system could instantly model all mitigation strategies, then identify and execute the most effective strategy—for example, creating stock transport orders or rebooking logistics—while syncing updates to arrival times and projecting losses across systems.

Result: Risk mitigation became significantly faster, more transparent, and scalable, with smarter decision-making that didn’t depend on manual effort.

A Global Accounting Firm

Problem: The company needed to comply with a wide range of internal and external regulations, but its audit processes were slow, expensive, and depended heavily on manual effort. Operations generated a mix of structured and unstructured data, making it challenging to consistently verify compliance. As a result, data preparation and auditing consumed a significant share of the auditors’ time.

Solution: We implemented an AI-powered system that could process unstructured data, extract new rules as they emerged, and automatically check all operations for compliance.

Result: Audit processes became faster, less costly, and more transparent. Automation reduced compliance overhead, improved operational quality, and increased visibility into financial processes.

The Stingray Model of Innovation

Implementing AI at scale can feel overwhelming. To ensure structure and clarity, UX teams can explore emerging frameworks such as the Stingray Model, an AI-powered innovation approach.

Unlike traditional models such as the Double Diamond, the Stingray model positions AI as a co-innovator, embedding it throughout the innovation process for hypothesis generation, solution development, and testing with synthetic users. Stingray operates as a continuous loop of training, developing, and iterating, an approach that can be particularly useful for complex enterprise AI projects where speed, feasibility, and creativity must coexist.

The Stingray Model is one of the first frameworks to integrate AI deeply into the UX design process, and we expect more models to emerge as the field of AI evolves. Nevertheless, proven structures such as the Double Diamond remain highly relevant and provide a solid foundation for AI-driven innovation.

The Stingray Model comprises three stages, as follows:

1. Train—Define goals and gather intelligence. Instead of jumping straight into user research, teams first define the problems they aim to solve, the constraints that matter, and their success metrics. Then they gather broad inputs such as user behaviors, market trends, internal data, and feasibility considerations. AI helps synthesize this intelligence into prioritized, evidence-backed opportunity areas, reducing blind spots from the start.
2. Develop—Explore problems and solutions in parallel. Rather than treating problem definition and solution design as separate phases, the Develop stage explores both simultaneously. AI generates hypotheses and potential solutions, organizes them into clusters, and surfaces the most promising directions. This stage can be fully AI-driven or occur in combination with expert workshops. Teams can build out early services or product concepts by creating prototypes, visuals, or feasibility assumptions.
3. Iterate—Validate and refine design solutions. Test concepts using traditional methods such as prototyping and user interviews and AI-enhanced tools such as synthetic testing, which predicts user behaviors or evaluates technical fit. AI also helps factor in constraints such as sustainability and cost optimization. This accelerates iteration cycles and supports decision-making that is grounded in real-world viability.

The Stingray Model is especially useful for AI-at-scale projects that involve complex data, high uncertainty, and rapid iteration. To apply Stingray effectively, UX teams must ensure data quality, balance automation with human oversight, and iterate not just for usability, but to test feasibility and business impacts.

Our Final Words

Scaling AI offers significant benefits to enterprises. In 2022, MIT introduced the Enterprise AI Maturity Model, which was based on a survey of 721 companies. The model outlines four stages of AI maturity: Experiment and Prepare, Build Pilots and Capabilities, Develop AI Ways of Working, and Become AI Future Ready. Distinct focus areas and organizational characteristics define each stage, but most importantly, there are measurable differences in growth and profitability at each stage, as Table 1 shows.

Partial AI adoption—by limiting experimentation or staying in pilot mode—can introduce operational costs without delivering meaningful returns. Only when we embed AI into core ways of working do performance indicators turn positive. For enterprises, this requirement reinforces the need to move beyond isolated use cases and invest in scalable, system-wide AI integration.

For UX designers, this shift toward scaled AI creates both opportunity and responsibility. We must develop a deeper understanding of systems, data, and cross-functional collaboration—not only to craft easy-to-use user interfaces, but to influence how we integrate, explain, and govern AI across the enterprise.

When developing AI systems, the UX design role expands beyond surface-level user interactions to shaping users’ trust, aligning AI behaviors with business values, and ensuring that solutions are both usable and viable in real-world conditions. In this AI landscape, UX design is not a finishing layer, but a strategic force for making AI truly work at scale. 

Note: In this article, we’ve used artificial intelligence (AI) as an umbrella term that covers several technologies—from classical, analytical machine learning (ML) to generative models and emerging agent-based systems. When applying specific UX patterns, it’s important to tailor them to the underlying technology. For instance, confidence levels in large language models (LLMs) are often less reliable than in supervised ML systems, so it’s necessary to adjust transparency and feedback mechanisms accordingly.