Executives are under pressure to deploy AI, but many efforts stall before they start. Successful AI adoption is not about algorithms but the foundations of data, infrastructure, and people.
The current wave of AI adoption in manufacturing comes with high expectations. The possibilities appear limitless, from predictive maintenance and autonomous supply chains to AI-generated reports and robotic programming via natural language. However, a significant gap exists between building a prototype and embedding AI into daily operations across a global organisation.
For Matt Iliffe, CEO of cloud and AI consultancy Beyond, the difference comes down to foundations. “AI has been around for decades,” he explains, “What is changing is accessibility at scale. Cloud platforms, particularly Google Cloud, allow organisations to operationalise AI because they offer data access, quality, and automation tools within a cohesive infrastructure.”
This coherence matters. Many manufacturers already operate dozens, if not hundreds, of digital systems across production, logistics, compliance, and reporting. The promise of AI is not simply better data but faster decisions. That only happens when infrastructure is unified, models are repeatable, and insights are delivered directly into business workflows.
Choosing a platform that understands AI
Google Cloud has emerged as a strong candidate for AI workloads among the big cloud platforms. The reason, according to Iliffe, is convenience and integration. “Google Cloud’s workflow infrastructure, from BigQuery to Vertex AI, simplifies the path from raw data to actionable insight,” he says. “The platform bakes in MLOps, supports data quality controls, and makes it easier to test, tune, and deploy models without needing to stitch together disconnected tools.”
This consistency is particularly valuable for manufacturers navigating complex environments. Google Cloud enables organisations to run these workflows end-to-end rather than rely on manual processes to prepare data, move models between environments, and assess compliance risks.
Beyond, as a Google Cloud partner, often embeds MLOps teams into client organisations. This helps operationalise locally developed models and accelerate deployment cycles.
“Some organisations were taking weeks or months to get a model from test to production,” Iliffe explains. “With Google Cloud’s MLOps tools, we can bring that down to days. Continuous training, multivariate testing, and live-data staging environments make this possible.”
Generative AI and the rise of conversational manufacturing
Traditional AI is not going anywhere. Optimising a machine’s performance, forecasting demand, or monitoring product quality still relies on tried-and-tested machine learning approaches. However, the recent leap in generative AI adds a new layer of usability that lowers the entry barrier.
“The key shift is that generative AI gives people a new interface to engage with manufacturing. Instead of writing a SQL query or configuring a dashboard, a production manager can ask a question in plain language and receive insight instantly,” Iliffe says.
In one case, a robotics firm in San Francisco deployed generative AI to enable voice-based commands for robotic operations. This allowed non-specialist staff to interact with manufacturing robots, reducing reliance on programming expertise and speeding up onboarding in small-scale production settings.
Beyond uses the term ‘workplace solutions’ to describe this layer of enablement, tools that augment staff capability, automate reports, or provide access to mission-critical knowledge without requiring technical fluency. These tools are not just for efficiency but for empowerment. “Generative AI offers conversational insight, real-time data interrogation, and the ability to embed automation into workflows,” Iliffe adds. “Whether that means generating compliance reports, logging incidents, or retrieving specifications, it makes information accessible in a way that is both immediate and useful.”
Why most AI projects still fail
Despite this potential, failure remains the norm. Many projects do not move past the pilot stage and those that do often deliver disappointing returns. Iliffe points to five key challenges that separate success from stagnation: fragmented infrastructure, poor data quality, unsafe deployment methods, internal skills gaps, and inadequate proof of value.
“Infrastructure is a frequent bottleneck,” Iliffe explains. “We worked with a client with over 200 separate data repositories across its brands. You cannot build cohesive AI services on top of that. You need modern architectures, consolidated data stores, and a clear path from ingestion to insight.”
Manufacturing firms often have volumes of historical data, but much was never collected with AI in mind. It lacks context, consistency, or labelling. “The old rule still applies, garbage in, garbage out,” Iliffe adds. “Machine learning only works when the data reflects the reality you want to predict,”
Then, there is the challenge of safe deployment. In regulated sectors or even contractual guarantees exist, models cannot simply be pushed into production. Controls around access, reliability, and explainability must be in place. This is where mature MLOps pipelines make a difference, using staging environments with live data and structured governance to reduce risk.
The fourth obstacle is humans. Many organisations lack the internal expertise to move fast. Hiring is slow, upskilling takes time, and new tools appear faster than teams can absorb them. Beyond’s model aims to bridge this gap by embedding experts within clients while training internal teams.
Finally, there is the question of proof. Too many pilots are executed without assessing operational value. “Proof of concept is not the goal,” Iliffe adds. “Proof of value is. If you do not start with a clear view of the benefit, a project may technically succeed but never get adopted. You need to measure against real outcomes.”
Focus on fit, not hype
AI should not be deployed for its own sake. Nor should it be embedded indiscriminately into every process. The danger with AI’s current visibility is that it gets used to tick boxes rather than solve problems. “AI is often marketed as an all-purpose solution,” Iliffe says. “But what really matters is the specific value it unlocks. You do not need AI in every process; you need it in the right process, with the right data and infrastructure behind it.”
This targeted approach is critical when moving beyond experimentation. AI in production is about stability, maintainability, and integration. That means using technology like microservices, APIs, and containerisation to ensure AI models are not isolated components but embedded in business systems. In one example, a client migrated its legacy applications to Google Cloud, adopted microservices, and moved its data into BigQuery. This foundational work enabled the development of AI-powered recommendation engines and price prediction tools that would have been impossible otherwise.
Multimodal and adaptive intelligence is next
Iliffe identifies two trends likely to reshape manufacturing in the next five years. The first is multimodal AI, which can process and respond to inputs across voice, visual data, and text. The second is adaptive AI, a model that tunes itself to individual users or workflows. “These tools will act more like co-pilots,” he adds. “In manufacturing, that means enabling workers to access knowledge on demand, interact with machines through natural input, and respond to real-time issues. It changes how knowledge is transferred, how training is delivered, and how decisions are made.”
As large language models continue to improve, smaller and more specialised models are also gaining ground. These can be fine-tuned for sector-specific tasks, improving accuracy while reducing computational cost.
Iliffe offers simple but firm advice for manufacturers looking to begin this journey. “Do not rush in,” he concludes. “Start by modernising your infrastructure. Move to the cloud. Clean your data. That gives you the foundation to experiment with AI and scale what works.”
Behind the buzzwords, the real story of AI in manufacturing is about building something durable that goes beyond prototypes and into operations. AI is not the solution to everything. But with the proper foundation, it can become the architecture behind a smarter, faster, more adaptive factory.
