💻 computer science

Industry Practitioners Perspectives on AI Model Quality: Perceptions, Challenges, and Solutions

Through interviews with 15 practitioners and a survey of 50, this study identifies nine key AI model quality attributes, revealing that their prioritization is context-dependent and that data imbalance remains a critical challenge, thereby offering insights to help researchers align their work with industry needs.

Original authors: Chenyu Wang, Zhou Yang, Yunbo Lyu, Ze Shi Li, Daniela Damian, David Lo

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chenyu Wang, Zhou Yang, Yunbo Lyu, Ze Shi Li, Daniela Damian, David Lo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are building a high-tech robot chef. In the old days of software, you wrote a strict recipe: "If the user asks for soup, add water, salt, and carrots." The robot followed the rules perfectly.

But today's AI is different. Instead of a strict recipe, you feed the robot a million photos of soups and let it figure out the pattern itself. It's like teaching a child by showing them pictures rather than giving them a manual. This is powerful, but it's also messy. The robot might get confused, make weird mistakes, or learn bad habits from the photos you showed it.

This paper is like a group of master chefs (industry practitioners) sitting down to discuss: "How do we make sure our robot chefs are actually good at their jobs, and what's stopping them?"

The researchers interviewed 15 experts from big companies (like banks, tech giants, and e-commerce sites) and then asked 50 more to confirm their findings. Here is what they discovered, translated into everyday language:

1. The "Speed vs. Perfection" Trade-off

The Finding: In real-time apps (like a live recommendation system or fraud detector), speed is king. If your AI is 1% more accurate but takes 2 seconds to answer, users will just leave.

The Analogy: Imagine a race car driver. If you give them a car that is slightly faster but takes 10 seconds to start the engine, they will lose the race. In the real world, a "fast and good enough" answer is often better than a "perfect but slow" one.
The Lesson: Companies often accept a slightly less perfect model if it means the user gets an answer instantly.

2. The "Moving the Heavy Lifting" Shift

The Finding: A few years ago, developers worried a lot about how to fit their AI onto a server or how to handle millions of users at once. Now, they don't worry about that as much.

The Analogy: Think of building a house. In the past, the architect had to design the foundation, the plumbing, and the electrical grid all at once. Today, the architect just designs the rooms, and a specialized construction company (the cloud platform) handles the heavy foundation and plumbing.
The Lesson: AI developers now focus on the "brain" of the model, while the "muscle" (scaling and deployment) is handled by automated infrastructure.

3. The "Lawyer in the Room"

The Finding: Laws and regulations are the biggest reason why certain AI qualities matter. If you are in finance, you must explain your decisions. If you are in healthcare, you must be safe.

The Analogy: Imagine a student taking a test. In a normal class, they just need a high score. But if they are taking a Bar Exam (to become a lawyer), they can't just get the right answer; they have to show their work and prove they didn't cheat.
The Lesson:
- Finance: Needs Explainability (Why did you deny my loan?).
- Safety: Needs Robustness (Don't let a hacker trick the self-driving car).
- Privacy: Needs Data Protection (Don't leak my medical records).

4. The "Data Imbalance" Problem

The Finding: AI models struggle when they don't see enough examples of rare things (like fraud or fake documents).

The Analogy: Imagine teaching a kid to spot a rare blue bird. If you show them 1,000 pictures of sparrows and only 5 pictures of blue birds, the kid will think everything is a sparrow.
The Solution:
- Active Learning: Instead of guessing which pictures to show the kid next, you ask the kid, "Which of these 10 pictures are you confused about?" and then show them those specific ones. This is the most effective way to learn.
- Data Augmentation: If you don't have enough blue bird photos, you take the 5 you have and flip them, zoom in, or change the colors to make 50 new photos. The experts prefer these "old-school" tricks over fancy new AI generators because they have more control.

5. The "Drift" Dilemma

The Finding: The world changes. A model trained on 2020 data might fail in 2026 because people's behavior changed.

The Analogy: Imagine a weather forecaster who learned the climate of 1990. If they try to predict 2024 weather using those old rules, they will be wrong.
The Solution: Instead of building a complex, expensive sensor to detect exactly when the climate changes, most companies just say, "Let's retrain the model every Monday" or "Let's retrain it once we have 1,000 new examples." It's not perfect, but it's practical.

6. The "AI Helper" for Labeling

The Finding: Labeling data (teaching the AI what is what) is expensive and boring.

The Analogy: Imagine you have to sort a pile of 10,000 letters. It takes you all day. Now, imagine you have a robot that sorts 9,000 of them for you, and you only have to fix the 1,000 it got wrong.
The Lesson: Using AI to help humans label data saves money and can actually be more accurate than humans working alone, because the AI doesn't get tired or make the same mistakes twice.

Summary: What Should We Do?

The paper tells researchers and companies:

Don't just chase the highest accuracy. If it's too slow or too expensive, it's useless.
Listen to the lawyers. Make your AI explainable and fair if the law requires it.
Use the right tools. Don't use a sledgehammer to crack a nut; use "Active Learning" for rare data and "Model Compression" to make AI run on cheaper, faster hardware.
Keep it simple. Sometimes, a scheduled retraining is better than a complex, failing monitoring system.

In short, the industry is moving away from "perfect academic theories" and toward "pragmatic, fast, and legally safe solutions" that actually work in the real world.

1. Problem Statement

While Artificial Intelligence (AI) is ubiquitous, ensuring the quality of AI models remains a critical challenge. Unlike traditional software, which relies on deterministic logic, AI models are data-driven, often functioning as "black boxes" with non-deterministic outputs.

The Gap: Existing academic research often focuses on theoretical correctness and robustness, but these findings may not translate to industrial realities where models are larger, more complex, and subject to specific domain regulations.
The Need: There is a lack of empirical understanding regarding which quality attributes industry practitioners actually prioritize, the specific challenges they face in maintaining these attributes, and the practical solutions they employ. Previous studies often focus on process-level activities (e.g., team collaboration) rather than concrete quality attributes (e.g., fairness, efficiency) linked to business success.

2. Methodology

The study employs a mixed-methods approach combining qualitative interviews and quantitative validation.

Phase 1: Attribute Identification & Interview Study
- Participants: 15 AI practitioners from diverse roles (ML Engineers, Data Scientists, Researchers), industries (Finance, E-commerce, Tech), and company sizes (Startups to Multinationals) across five countries.
- Process: Semi-structured interviews (45–70 mins) were conducted.
- Attribute Selection: The authors refined a list of quality attributes based on literature (ISO/IEC 25059) and pilot interviews, resulting in nine key attributes: Correctness, Robustness, Efficiency, Fairness, Explainability, Privacy, Scalability, Deployability, and Maintainability.
- Analysis: Thematic synthesis using hybrid coding (deductive based on the 9 attributes and inductive for emergent themes). Inter-rater reliability was high (Cohen's kappa = 0.76).
Phase 2: Quantitative Validation
- Survey: A follow-up survey was distributed to 50 AI practitioners (filtered from 61 initial responses to ensure industry experience).
- Goal: To validate the perceptions, challenges, and solutions identified in the interviews. Respondents rated agreement on a 5-point Likert scale.

3. Key Contributions & Findings

The study identifies four key perceptions and four key challenges with corresponding solutions.

A. Key Perceptions (RQ1)

Efficiency as the Dominant Constraint in Real-Time AI:
- In real-time applications (e.g., fraud detection, recommendations), Efficiency (latency and cost) often acts as a hard constraint that limits Correctness.
- Practitioners prioritize low latency and cost over marginal gains in accuracy. If a system is too slow, user abandonment rates spike, rendering high accuracy irrelevant.
Infrastructure Absorption of Scalability and Deployability:
- Contrary to older studies citing these as major bottlenecks, practitioners report that Scalability and Deployability are now largely handled by platform infrastructure (Microservices, Kubernetes, Serverless).
- AI teams focus on containerizing models and exposing APIs; the "plumbing" of scaling and deployment is delegated to infrastructure teams.
Compliance as Legislative Specification:
- Compliance is not just a constraint but a rigid specification that dictates which quality attributes are prioritized based on the domain:
  - Safety/Generative AI $\rightarrow$ Robustness (against jailbreaking/harmful content).
  - Finance/Audit $\rightarrow$ Explainability (traceability of decisions).
  - Demographics/HR $\rightarrow$ Fairness (anti-discrimination).
  - Data Handling $\rightarrow$ Privacy (preventing model memorization/leakage).
Dual Purpose of Explainability:
- Explainability serves two masters: Developers (for debugging and maintaining model health) and End-Users (to build trust and aid decision-making). It has shifted from a purely regulatory requirement to a tool for user satisfaction.

B. Key Challenges & Solutions (RQ2)

Challenge: Data Imbalance.
- Solution: Active Learning is highly effective for acquiring scarce positive cases (e.g., fraud). For data synthesis, practitioners prefer traditional resampling/augmentation (e.g., SMOTE, rule-based image manipulation) over complex Generative AI due to better control and stability.
Challenge: Data Drift Monitoring.
- Solution: Complex, real-time drift detection is often bypassed due to high complexity and false positives. Instead, practitioners use deterministic triggers: time-based scheduling (e.g., weekly refresh) or volume-based triggers (retraining when a threshold of new data is collected).
Challenge: Cost vs. Performance Trade-off.
- Solution: Model Compression + CPU Inference. To reduce operational costs (especially for memory-heavy tasks like embeddings in recommender systems), practitioners compress models (pruning, quantization, distillation) to enable efficient inference on CPUs rather than expensive GPUs.
Challenge: Annotation Costs.
- Solution: AI-Assisted Annotation. Using models to pre-label data, with humans only resolving disagreements. This reduces costs and, surprisingly, can improve quality by breaking "human blind spots" and false consensus.

C. Validation Results (RQ3)

Validation: 7 out of 8 findings were "Well-Acknowledged" (average score > 4.0) by the survey respondents.
Nuance: The finding regarding "AI-assisted annotation" was "Marginally Agreeable" (3.98), with some practitioners noting that in highly specialized domains (e.g., rare disease diagnosis), human expertise remains irreplaceable.
Consensus: There is broad agreement that efficiency constraints, infrastructure shifts, and compliance-driven priorities are the current reality of industrial AI.

4. Significance and Implications

For Researchers: The study highlights a misalignment between academic research and industry needs. Researchers should avoid optimizing a single attribute (e.g., pure accuracy) if it compromises attributes practitioners deem critical (e.g., latency or cost). Future work should focus on trade-off optimization and practical solutions for data imbalance and drift.
For Practitioners: The paper provides a validated "playbook" for AI engineering:
- Prioritize Active Learning for data acquisition.
- Adopt Model Compression to leverage cost-effective CPU hardware.
- Rely on Infrastructure for scaling rather than custom code.
- Treat Compliance as a primary driver for attribute selection (e.g., robustness for GenAI).
For the Field: It shifts the narrative from "AI is hard to deploy" to "AI deployment is now an infrastructure problem," allowing AI engineers to focus more on model quality and business value. It also redefines the role of explainability from a regulatory checkbox to a dual-purpose tool for debugging and trust.

5. Conclusion

This research bridges the gap between academic theory and industrial practice by empirically validating how AI quality is perceived and managed in the real world. It demonstrates that while correctness remains foundational, efficiency, compliance, and maintainability are the driving forces shaping modern AI engineering strategies. The identified solutions (Active Learning, CPU inference, pragmatic retraining) offer actionable guidance for building reliable, trustworthy, and economically viable AI systems.