In recent years, survey after survey has found that adoption of, and often confidence in, AI remains uneven. Avnet’s newly released 2026 Insights survey is the latest in this vein. The study, which offers a view from the engineering perspective, mirrors the growing traction for AI products, which has gone from niche to mainstream in a handful of years. Among 1,200 engineers surveyed worldwide, 56% said they are shipping products with AI incorporated, up from 42% a year ago. Yet the same respondents flagged persistent hurdles: 46% cited data quality as a top design challenge, and 54% named continuous learning and maintenance as the top operational challenge.
The results are broadly similar to findings in software. For instance, Stack Overflow’s 2025 Developer Survey of 49,000 developers worldwide found that 84% of developers now use AI tools in their workflows, yet only 33% trust the accuracy of what those tools produce.
Avnet’s data also helps clarify what “AI in products” looks like right now. Among the most common embedded deployments, engineers cite use cases like process automation (42%), predictive maintenance (28%) and fault or anomaly detection (28%). At the same time, more teams are mixing approaches: 57% say they are prioritizing Edge AI and ML equally.
That mix of rising adoption and persistent friction helps explain a shift in where leverage sits. Engineers do use mainstream LLM tools, but their preferences point toward more specialized, professionally tuned systems: only 16% say they would prefer a publicly available LLM for technical questions, while 47% would rather use an LLM trained by engineers outside their organization. The companies that win share may be the ones that make AI easier to verify, easier to maintain, and better aligned with engineering workflows, rather than the ones with the biggest launches.
Below, we recap 10 themes from Avnet’s research and place them in context with other recent surveys on AI adoption, trust, and operational readiness.
1. Adoption is rising faster than confidence
Engineers report a step up in shipping AI-enabled products (56% vs. 42% last year), a year-over-year increase that signals the market is moving from pilots to default inclusion. But the survey’s top challenges suggest rollout can outpace readiness, especially as AI systems move into production environments where teams must monitor performance, manage drift and maintain models over time. In the consumer sector, AI adoption is no monolith. Menlo Ventures found in a survey of 5,031 U.S. adults that while 61% used AI in past 6 months, only 3% pay for premium services. Meanwhile, Capgemini (12,000 consumers) found in January 2025 that 71% want GenAI integrated into shopping and that 58% replaced traditional search with GenAI for their go-to for product/service recommendations.
2. Data quality remains the dominant constraint
In Avnet’s survey, 46% of engineers cited data quality as a top design-level challenge—the highest-ranked barrier to AI deployment. The finding reinforces an old principle: garbage in, garbage out. And Avnet’s number may fall on the conservative side. KPMG’s Q3 2025 enterprise survey found data quality concerns jumped from 56% to 82% in a single quarter as companies attempted to scale AI initiatives. Qlik reported that 81% of AI professionals still face significant data quality issues, with 85% saying leadership is not addressing the problem.
3. Maintenance is emerging as the core operational burden
Continuous learning and maintenance is the top operational challenge (54%), ranking higher than cost. This is where AI roadmaps can slip: monitoring for drift, triggering retraining, catching regressions, and navigating governance requirements. The 2023 to 2024 period rewarded prototypes; 2025 to 2026 is about keeping systems stable after launch.
Q&A: Alex Iuorio on the Survey Findings
Alex Iuorio is Senior Vice President of Global Supplier Development at Avnet.
Adoption jumped 33% in a year, but 46% cite data quality as a top challenge. Is the industry integrating AI faster than it can handle reliably?
Iuorio: As with any new technology, there’s going to be a learning curve. Our results found that applications predicted to have the highest adoption rates were the same for last year and this year: process automation, predictive maintenance and fault/anomaly detection. This showcases a trend that engineers are figuring out the best use cases and getting over the initial learning curve. However, the technology is still new and ever-changing, presenting strong potential for growth.
Gemini usage hit 57%. How much of that comes from Google’s Android bundling versus engineers choosing it for performance?
Iuorio: This is not something our survey looked at, but what I can tell you is that engineers would prefer to be using an LLM trained by their peers outside of their organization to answer technical questions, as opposed to a publicly trained LLM.
Avnet sits in the middle of the supply chain. Have you seen evidence that poor data quality or rushed integration is causing project failures?
Iuorio: These are not new challenges; there is a “rush” in all aspects of product design and launch. However, these issues are at the forefront now because of the very nature of AI. The data sets engineers are working with are massive, and the higher the quality of that data, the more precise these outcomes will be. Almost half of the engineers surveyed being aware of the data quality challenge is a positive: knowing there’s a problem is an important step in solving it. Coupling this with the 33% increase in products shipping with AI functionality suggests these challenges are not insurmountable.
With multi-modal and edge AI prioritized but foundational issues persistent, how has the survey changed your thinking?
Iuorio: The survey has only intensified our belief that Edge AI is a significant and growing opportunity. Over half of respondents said they were prioritizing Edge AI and machine learning models, underscoring the importance of both technologies. This, coupled with the 33% increase in those shipping products with AI incorporated, strengthens this belief.
Unlike traditional software, AI models can degrade even when the code stays the same. A 2025 LLMOps analysis found that models left unchanged for six months or more saw error rates jump 35% as input data drifted from training distributions. Gartner projects that at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, citing costs, governance issues or unclear value as the primary reasons. Only about 48% of AI projects make it to production at all, with many stalling in pilot stages. The implication for engineering teams: the demo is the easy part. The work that follows, drift detection, retraining pipelines, compliance documentation, is where projects live or die.
4. AI in products is concentrated in measurable use cases
The most common embedded deployments are process automation (42%), predictive maintenance (28%), and fault or anomaly detection (28%). These are use cases where value can be measured and failure modes can be contained, such as factory floors, sensor networks, and quality control systems. The pattern aligns with broader industry data. In 2024, predictive maintenance emerged as the leading AI application in manufacturing, according to market research. Adoption tended to be concentrated in automotive, heavy machinery and semiconductor production, sectors where unplanned downtime carries significant cost.
5. Hybrid architectures are becoming standard
Over half of respondents (57%) say they are prioritizing Edge AI and ML equally. In other words, many designs are splitting inference across device and cloud for latency, cost, privacy or resilience reasons. The economics are straightforward: real-time inspection and control applications often generate more data per second than network bandwidth can handle, forcing inference to the edge. After years of next-big-thing positioning, the edge AI market is accelerating, with manufacturing holding the largest share of deployments driven by predictive maintenance, quality inspection and robotics coordination.
6. ‘Multi-modal’ is turning into an integration problem
Avnet frames multi-modal as using more than one type of AI in a product, or AI that can handle more than one inference type, such as text plus image plus sensor fusion. The practical implication is system design: integrating vision, speech, time-series models, and LLM-based interfaces into a single product. IDC predicts that by 2028, 80% of foundation models used for production-grade use cases will include multimodal capabilities. The global multimodal AI market, valued at $1.34 billion in 2023, is projected to grow at a 35.8% CAGR through 2030.
7. Gemini closes the gap with ChatGPT among hardware engineers
In Avnet’s study, engineers reported using mainstream tools (ChatGPT 69%, Gemini 57%, Copilot 50%). The Gemini number marks a shift. Google spent much of 2023 and early 2024 as a latecomer in the LLM race, with Bard widely seen as trailing ChatGPT. That changed in 2025: Gemini 2.5 Pro showed the company could compete, and Gemini 3.0 now ranks at or near the top of several public leaderboards. Google has also embedded Gemini into Android, Chrome and Workspace, reaching users through products they already use. That distribution advantage has drawn regulatory attention: a federal judge recently blocked Google from forcing partners to bundle Gemini with other services, and the EU opened a probe into whether Google gave itself preferential access to content for AI training.
8. There is demand for domain-tuned, engineering-grade tooling
Nearly half of engineers (47%) would prefer an LLM trained by engineers outside their organization: vendor-built, engineering-specific models with stronger evaluation, provenance, and documentation. Only 16% say they would prefer a publicly available LLM for technical questions. The gap suggests a market for tools that map to real constraints: specific stacks, known failure modes, and compliance requirements.
9. Verification is becoming part of the workflow
Pair Avnet’s adoption patterns with Stack Overflow’s trust data and a consistent operating model emerges: AI accelerates drafts, but humans still validate. Only 29% of developers trust AI output accuracy, and 66% cite “almost right but not quite” as their top frustration. The bottleneck is often validation, not generation. The advantage shifts toward teams that can verify faster—through better test coverage, automated evaluation pipelines, structured reviews and staged rollouts. As model outputs become table stakes, the differentiator becomes how quickly and rigorously a team can confirm that the output is correct.
10. Value may shift from model headlines to reliability work
When data quality (46%) and maintenance (54%) dominate the challenge list, differentiation accrues to the unglamorous layers: data pipelines, evaluation frameworks, monitoring, governance, and deployment patterns that reduce operational drag. The frontier model race still gets the headlines, but most engineering teams are not scaling AI projects without headwinds. High-quality data collection remains one of the biggest bottlenecks in deploying AI at scale.
