Ardiansyah
Student, STIE YPUP Makassar
Keywords: AI, Innovation, FOMO, Business Strategy.
WIN Media, Opinion – Imagine an automotive component factory in West Java that, at the beginning of 2024, enthusiastically launched an AI project worth tens of billions of rupiah. Management was promised a vision of autonomous robots gliding through warehouse aisles, a system that magically predicts equipment failure three weeks before it happens, and production scheduling that adjusts itself in real-time to demand fluctuations. Yet six months later, the reality is starkly different: the dashboard screen is full of blazing red error reports, the algorithm refuses to work because “data is insufficient,” and instead of reducing workloads, operational staff have to work overtime every night to “fix” the chaotic AI output, an irony where the technology meant to liberate becomes a new set of shackles.
Supporting Facts: Numbers That Cannot Be Ignored
Data from various leading research institutions confirms that the bitter experience above is not an anomaly but a very common condition. Gartner predicts that at least 30% of generative AI projects will be abandoned after the proof of concept (PoC) stage by the end of 2025, primarily due to poor data quality, inadequate risk assessment, and unclear business value. More worryingly, a McKinsey (2024) survey revealed that more than 70% of AI projects in operations fail to reach full scale, with analysis of 50 AI Agent projects across various industries showing that nearly 40% end in total failure. A study by the RAND Corporation even found that 84% of AI implementation failures are caused by leadership and organizational factors, not technical weaknesses in the algorithms.
Thesis: Between Sweet Promises and Bitter Realities
The promise of AI efficiency in operations is very real and has been proven in a number of pioneering companies—from downtime reduction to supply chain optimization. However, the problem is that implementation challenges are systematically underestimated by business leaders lulled by smooth product demos and bombastic marketing claims. The success of AI is not determined by the sophistication of algorithms or the size of the budget allocated, but by the readiness of three fundamental pillars that are most often overlooked: the quality and availability of AI-ready data, the maturity of underlying operational processes, and the readiness of people and organizational culture to adapt.
Article Premise: Exploring the Three Main Gaps
This article will deeply explore the three main gaps that cause operational AI projects to get stuck in what analysts call “pilot purgatory”—a phase where the project never truly dies but also never reaches meaningful production scale. First, the data gap: why data previously considered “good enough” for reporting turns out to be completely unsuitable for training AI models. Second, the process gap: how automation applied on top of chaotic processes only produces automated chaos. Third, the people gap: why operational staff resistance is often a rational response to existential threat, not merely “resistance to change.” At the end, we will offer a practical roadmap to bridge these three gaps.
What Do AI Vendors and Consultants Promise?
Imagine a factory manager listening to a presentation from a leading technology consultant: “With AI, your production machines will be able to predict their own failures, a 2024 study proves that predictive maintenance can reduce downtime by up to 50% and cut annual maintenance costs by 37%.” These magical numbers keep coming: supply chain optimization that slashes inventory costs by 20-30% based on real-time analysis of supplier and customer data, quality control automation with accuracy surpassing even the most experienced human inspectors, and dynamic production scheduling capable of responding to demand changes within minutes, not days. These numbers are not just marketing claims, academic research published in the journal Computers & Industrial Engineering(2024) shows that graph reinforcement learning approaches for production scheduling can optimize resource allocation and energy consumption simultaneously, with significant efficiency gains. Meanwhile, a case study in the Romanian forklift industry reported productivity increases of up to 40% thanks to AI-based automation. With data like this, who wouldn’t be tempted?
Why These Promises Are So Easily Believed
Yet there are reasons why these promises feel so convincing, even to the most skeptical executives. First, AI success is often demonstrated in perfectly controlled environments, laboratories with clean data, simple pilot projects with limited scope, and ideal conditions rarely found in the real world. Second, immense competitive pressure: when competitors announce AI adoption for supply chains or predictive maintenance, the fear of being left behind becomes very real. Third, the dominant narrative of success from giant companies like Amazon, Tesla, and Siemens, with their impressive case studies, creates a mass imitation effect. An IDC (2025) survey even notes that 95% of manufacturers are now investing in AI, with more than 40% of digital transformation budgets allocated to predictive and generative initiatives.
“When product demos look so smooth and competitors are already moving, the question ‘why not us?’ sounds much louder than ‘are we ready?'”
There Is Another Side Rarely Discussed
What rarely appears in sales brochures and conference presentations is the bitter reality on the ground: the silent failures experienced by mid-sized companies and legacy companies trying to follow in the footsteps of tech giants. According to S&P Global Market Intelligence data released in May 2025, the failure rate of corporate AI initiatives jumped dramatically from 17% in 2024 to 42% in 2025, with the average company abandoning 46% of proof-of-concept projects before reaching production. An EXL survey of C-suite executives across various industries revealed that about 60% of current AI initiatives are stuck in the pilot phase, a condition analysts call “pilot purgatory”, with the main reasons being lack of talent, incompatible legacy systems, and data trapped in silos. An MIT Media Lab analysis even reported that 95% of corporate AI pilots fail to deliver measurable profit impact, not because of weaknesses in the AI models themselves, but because of integration failures into existing workflows and inadequate data infrastructure.
When the Foundation Wobbles, the AI Building Collapses
On paper, AI is a sophisticated prediction engine. But in the real world, AI is merely a mirror of the data it consumes—if that data is dirty, its predictions will be dirty too. The problem is that AI requires clean, structured, labeled data in large quantities, while day-to-day operations are filled with dirty data, isolated in departmental silos, and lacking consistent standards. A recent survey revealed that 98% of organizations acknowledge poor data quality as a critical barrier hindering AI adoption, while only 46% of executives trust their own organization’s data quality. Even more concerning, 84% of data and analytics leaders believe their data strategy requires a complete overhaul before their AI ambitions can succeed.
“Feeding AI with dirty data is like telling the best chef to cook with rotten ingredients—the result will never be satisfying.”
Manifestations on the Ground: Systematic Imperfections
On the factory floor and in warehouses, this problem manifests in very concrete, everyday forms. Sensors that haven’t been re-calibrated since they were installed five years ago send inaccurate data, while poor manual data entry habits, like technicians rushing to fill out forms at the end of their shift, cause much missing or incorrectly entered data. Even worse, data formats between machines or between factories are often fundamentally different: one machine records temperature in Celsius, another in Fahrenheit, with no automatic conversion. Most dangerous is the lack of data lineage, the traceability of data origin, so when an error occurs, the team can never track where the problem started. A 2025 study noted that 27% of organizations cite compliance constraints as the main barrier to AI data quality, while another 25% struggle with data duplication or conflict, and 21% face cross-system integration issues.
A Story from the Field That Happens Too Often
Imagine a component manufacturing factory wanting to implement predictive maintenance for its hydraulic press. The data science team excitedly builds a sophisticated prediction model, but quickly encounters an unexpected obstacle: the last ten years of machine failure history were never entered into a digital system, all records exist only in the notebook of an old technician, now fragile and filled with hard-to-read handwriting. Or the story of a logistics warehouse wanting to automate inventory management using AI vision, only to discover that thousands of old barcodes in the warehouse have faded or use old standards unrecognizable by modern AI systems. Another real case occurred at a production facility where, after routine maintenance, a manual valve on the pump bearing cooling system was accidentally left closed, the room controls had no visibility into the valve’s status, and the system interlock assumed everything was normal. As a result, bearing temperatures began to rise until an IoT sensor eventually detected the anomaly. This incident nearly caused expensive damage simply due to lack of data visibility.
Frequent Consequences: A Repeating Cycle of Frustration
When this data gap is not addressed from the start, the consequences are costly and repeat across many organizations. First, AI projects spend up to 80% of development time cleaning data, not building models or generating insights, a massive waste of resources confirmed by various industry studies. Second, AI models built from dirty data produce consistently inaccurate predictions, causing operational teams to lose trust in the system and prefer relying on intuition or old methods. Third, the peak of this frustration is the silent abandonment of the system: the AI is turned off, dashboards are no longer opened, and the company returns to old ways of working with the bitter feeling that “AI doesn’t work for us.” A Salesforce (2025) survey found that 89% of data and analytics leaders who have implemented AI in production have experienced inaccurate or misleading outputs due to poor data, and 55% reported significant resource waste because of low-quality data.
Quote or Analogy: A Wise Reminder
“Feeding AI with dirty data is like telling the best chef to cook with rotten ingredients.” This analogy may sound simple, but it contains a profound truth often forgotten by executives lulled by smooth product demos. No matter how sophisticated the algorithm or how large the budget spent, if the input data is low quality, the results will always disappoint. As Craig Gravina, CTO of Semarchy, put it: “Think of AI as a high-performance sports car driving across a dilapidated bridge, no matter how powerful the car, it cannot function if the underlying structure is broken.” A lesson to hold onto before starting every AI initiative.
When Automation Only Accelerates Chaos
Many business leaders mistakenly believe that AI is a miracle cure that can heal all operational ailments when in fact, if the underlying processes are already chaotic, AI will only automate that chaos at much higher speed. A study published by the World Economic Forum in August 2025 revealed that 55% of companies acknowledge outdated systems and processes as their biggest barrier to AI implementation, yet ironically, most remain focused on the technology itself rather than fixing their operational foundation. As Peter Drucker sharply reminded us, “There is nothing so useless as doing efficiently that which should not be done at all”, a warning highly relevant when companies rush to deploy AI on top of flawed processes.
Manifestations on the Ground: When Documents Only Decorate Shelves
In the real world, this process gap manifests in very familiar yet often overlooked forms. Standard Operating Procedures (SOP’s) that should guide work are either poorly documented or even worse the documents exist but are never followed by workers on the production floor. An industry analysis shows that SOP’s do not fail on paper but fail in practice: when procedures are not tied to real job roles, not mapped to how the facility actually operates, or not supported by execution tools like logs and checklists, those documents are mere “shelf fillers” with no real value. Furthermore, workflows that keep changing due to the absence of process discipline, coupled with a lack of clear ownership over each operational step, create an environment where AI can never learn consistent patterns.
“Don’t automate a broken process. You’ll just lock it in place. Instead, redesign that process first.” Joe Blanchett, business transformation practitioner
Concrete Examples: Two Stories That Happen Too Often
Imagine a national logistics company that invests heavily in an AI system for delivery route optimization, hoping to cut fuel costs and delivery times by 30%. But the project quickly hits a dead end when the team realizes that customer address input is still done manually by customer service, and the address entry error rate reaches 15%, ranging from misspelled street names to incorrect postal codes. Recent research in address standardization shows that manually entered addresses often contain irregularities such as missing information, misspellings, colloquial descriptions, and confusing directional offsets. In India, a logistics company even had to develop a custom AI system called SF Maps to overcome a similar problem, because addresses there are often unstructured and highly prone to input errors, a challenge that could actually be reduced if the address input process were standardized from the start.
The second story comes from a manufacturing factory wanting to implement AI vision for product defect detection. The technology team confidently trained an AI model using thousands of labeled images of “defective” and “good” products. But when the system was tested on the production line, its accuracy plummeted. Upon investigation, a shocking underlying problem was found: quality standards between the morning and night shifts were significantly different. The morning shift supervisor considered minor scratches as defects, while the night shift supervisor considered the same condition acceptable. As a result, the AI training data contained label inconsistencies reflecting the lack of uniform process standards. A study from the Kaizen Institute affirms that before implementing AI for quality control, organizations must first map the end-to-end value stream and identify where value is truly created and where bottlenecks occur.
Key Principle: Automation Will Not Fix Chaos
From the two examples above, one key principle must be held firmly by every business leader: “Automate a broken process, get broken automation”, a simple truth but often overlooked in the frenzy of AI excitement. Grant Wild, CEO of Wild Tech, states firmly that AI does not magically fix broken processes; AI will actually accelerate existing chaos. If your operations are not well understood, layering AI on top will only create more confusion. The World Economic Forum recommends an “improvement-first approach” that eliminates waste and optimizes processes before deploying AI, creating a solid foundation where technology can truly thrive. In other words, AI innovation must be preceded or at least accompanied by disciplined and thorough Business Process Re-engineering (BPR).
Brief Solution: A Roadmap to AI-Ready Processes
So, what must be done? The first non-negotiable step is: map and simplify your underlying processes before adding AI. This means conducting end-to-end value stream mapping to identify where waste occurs and where bottlenecks disrupt workflow. The equally important second step is to involve operational teams in process redesign, not just the IT team. As expressed in industry analysis on AI implementation, “SOPs don’t protect you, compliance systems do,” and those systems must be built together with the people who will use them every day. Wild adds that running workshops with teams, getting their input on what works and what doesn’t, and ensuring everyone is aligned are critical steps before introducing any automation. Finally, ensure you are very clear on the specific problem AI is meant to solve, not just “we want to be more efficient,” but “we want to predict supply chain delays” or “we want to automate repetitive reporting tasks.”
Core Problem: When Technology Collides with the Human Spirit
The scenario most often overlooked in any AI project is not data or process problems, but people themselves. AI, despite all its sophistication, fundamentally threatens the sense of security, autonomy, and even identity of operational workers, because for many people, their job is not just a way to make a living but also a source of self-worth and meaning. An academic study published in 2025 involving 112 employees from various manufacturing companies found that the relational tension between humans and machines is a major source of resistance, where workers grapple with shifting expectations around trust, control, and human identity. Without mature and empathetic change management, AI will be resisted, either openly or covertly, and even sophisticated projects will end up as expensive ornaments never touched.
“People don’t resist change. They resist being changed” Peter Senge.
Manifestations on the Ground: Resistance in a Thousand Faces
On the factory floor and in warehouses, resistance to AI takes various clever forms that are hard to detect early. There is the technician who suddenly becomes “constantly busy” so has no time to install new sensors, when in fact he fears that the sensor will prove his 20 years of experience could be replaced by an algorithm. There is the supervisor who “deliberately neglects” to enter wrong data into the system, hoping AI will produce false predictions so management loses trust in the technology. The most common and hardest to overcome are employees who silently continue using old methods even though the AI system is available right in front of them, a form of passive resistance almost impossible for any dashboard to detect. A KPMG 2025 survey confirms that employee resistance to change is among the top three barriers to AI adoption, with 47% of organizations admitting to experiencing it. Meanwhile, a global Expereo-IDC survey found that 35% of technology leaders identify employee resistance as a primary barrier to AI adoption in their organizations.
Stories from Two Different Worlds
Imagine an automotive manufacturing factory that invests billions of rupiah in a sophisticated AI production scheduling system. The algorithm carefully calculates the optimal production sequence based on order data, material availability, and machine capacity. But the production line foreman, who has worked for 25 years and takes pride in his instincts, firmly ignores every system recommendation. “Experience knows better than some algorithm cooked up by youngsters,” he says while sticking to his own version of the schedule, which, ironically, causes the very recurring bottlenecks the AI was meant to solve. In another warehouse, an automated stock prediction system trained on highly accurate historical data recommends just-in-time reordering. But warehouse staff, accustomed to “just-in-case” ordering by buying extra for fear of stock-outs, continue to place manual orders twice the system’s recommendation. As a result, carrying costs balloon and working capital is tied up uselessly. Matt Kropp of Boston Consulting Group explains that the deepest resistance stems from threats to identity: “Someone might think, ‘I’m an engineer, I write code, that’s what makes me special. If AI can write code, why am I still special? Who am I?'” This is the core issue that cannot be solved by ordinary technical training.
What Is Needed (Not Just Training): Four Pillars of Human Transformation
Technical training on how to use AI is completely insufficient to bridge the people gap. A much more holistic and deeper approach is required. First, worker involvement from the start in the co-design process, not a top-down approach that imposes solutions from above. AWS research emphasizes the importance of addressing “organizational debt”, the accumulation of outdated processes, rigid hierarchies, and cultural resistance to change, before deploying AI. Second, role redefinition: communicate clearly that AI is an assistant that amplifies human capabilities, not a replacement. The “find the toil and find the joy” approach developed by BCG invites employees to jointly redesign their work processes: eliminating boring tasks (toil) while preserving what they love (joy). Third, incentives for adoption, not punishment for resistance, because threats will only drive underground resistance that is even harder to detect. Fourth, genuine up-skilling programs, not mere formalities. KPMG surveys show that companies are beginning to train employees with new approaches such as prompt skill learning, sandbox environments for practice, and shadowing programs. Academic studies also affirm that organizations need to “embrace paradoxical thinking”, using AI to boost productivity while actively supporting employees in transitioning to evolving roles.
An Eternal Reminder About the Nature of Change
“People don’t resist change. They resist being changed.” This classic quote from Peter Senge contains a profound truth often forgotten by leaders rushing to implement AI. Change forced from above, without involving those who will be affected, will almost certainly meet resistance, both visible and hidden. Matt Kropp adds an important insight: teams whose managers actively use AI tools themselves have adoption rates four times higher than teams whose managers do not. Leading by example, not by command, is key. Furthermore, how top leadership frames the purpose is crucial: saying “we want a 30% headcount reduction through AI” will never earn voluntary compliance, while saying “this is about making us more successful, innovating faster, and serving customers better” opens the door to collaboration. The bottom line: AI must be introduced as an empowering tool, not as a threatening executioner.
An Undeniable Truth
After diving into the three gaps that often trip up operational AI projects, the data gap, the process gap, and the people gap, one conclusion becomes irrefutable: AI in operations is a remarkable tool, but not a magical one. It will not heal a sick organization simply by being installed like a patch. The tempting promises of efficiency, from 50% downtime reduction to 20-30% inventory cost cuts, will only materialize if companies honestly and courageously face all three gaps simultaneously. Neglect just one, whether dirty data, chaotic processes, or threatened people, and even the most sophisticated AI project is certain to fail, or at best become an expensive ornament that never delivers real business value. As various studies remind us, including RAND Corporation’s finding that 84% of AI implementation failures are caused by leadership and organizational factors, not technical algorithm weaknesses.
“Sustainable AI innovation is built on a clean operational foundation, not on dreams.”
Between the Temptation of Demos and On-the-Ground Reality
To business leaders, especially Chief Operating Officers and operations managers, we entrust this message: do not be tempted by smooth vendor demos. Sales presentations are designed to dazzle you, with beautiful interfaces, perfectly accurate predictions, and impressive success stories. But behind the scenes, those demos are typically run on data that has been cleaned for months by the vendor’s engineering team. Start with a readiness audit, not by buying the most expensive software. Ask yourself: Is our data AI-ready? Are our processes mature enough to be automated? Is our team mentally and emotionally ready? Sustainable AI innovation is built on a clean operational foundation, managed data, standardized processes, and people who feel involved, not on dreams sold by marketing brochures. As expressed in various industry studies, AI success is not determined by how sophisticated the model you buy is, but by how ready your organization is to receive it.
Three Concrete Steps for the Next 90 Days
Theory without action is just wishful thinking. Therefore, we invite you, operations managers and COO’s, to start your AI transformation journey with small but impactful steps over the next 90 days. First, within the next 30 days, conduct a data quality assessment in one operational line, not the entire company. Choose one machine, one warehouse, or one production line, then audit the cleanliness, completeness, and consistency of its data. Identify where the most critical data failure points are. Second, identify one simple process that can be “assisted” by AI, not replaced entirely. Do not choose a complex, high-risk process.
Choose a repetitive, boring task that does not require highly critical decisions, for example, automatic document classification or simple anomaly notifications. The goal is to build a quick win that builds trust, not a threatening big revolution. Third, form a small cross-functional team, comprising representatives from IT, operations, HR, and most importantly, representatives of front line workers who will use AI every day for a limited 90-day experiment. Give them autonomy, a small budget, and the authority to make decisions. Let them fail fast and learn cheaply. From these small experiments, true learning is born, and the foundation for future scalability is built. As proven by various successful companies, a gradual and inclusive approach is far more effective than a large project forced from the top down.