Don’t Let AI Automate Your Hidden Factory
In my last article, I wrote about systems thinking and the iceberg — the idea that visible problems are rarely the whole problem. This article moves from the iceberg to the work itself.
When leaders look below the surface, they often find the same things: rework, handoffs, workarounds, duplicate reviews, unclear ownership, poor data, delayed decisions, and exception handling that has quietly become part of how the organization operates.
In Lean Six Sigma, we often call this the hidden factory.
It is the work people do because the process does not work right the first time. And in the age of AI, that hidden factory matters more than ever.
AI can help eliminate hidden work. But if leaders are not careful, it can also automate it, accelerate it, and make it harder to see.
Speed Is Not the Same as Value
Many organizations are asking how AI can improve productivity. That is understandable. Productivity matters, and AI clearly has the potential to help people work faster.
But productivity is not just about speed. And automation does not automatically create value.
If AI makes a broken workflow faster, the organization may see more activity without better outcomes. The work may move more quickly, but the rework, handoffs, unclear ownership, poor data, extra reviews, and delayed decisions remain.
A faster draft that still requires heavy correction is not transformation. A faster report built on unreliable data is not better decision-making. A faster customer response based on unclear policy is not better service. A faster exception process does not necessarily reduce the number of exceptions.
This is one of the risks I see in the current AI conversation. Organizations are rightly excited about the ability to generate, summarize, analyze, predict, and automate. But speed alone does not create value if the work itself has not been redesigned.
AI should accelerate flow, not friction.
Existing+ Is a Starting Point, Not a Destination
The MIT Center for Information Systems Research has described several business models emerging in the AI era. One of them is called Existing+ — where a company augments its existing business model with AI capabilities. In this model, AI assists customers through established products, services, and processes.
That is where many companies will start, and often it is a sensible place to begin. Most organizations do not leap immediately into fully autonomous, AI-enabled operating models. They start by adding AI to the work they already know.
But Existing+ also carries a risk.
If AI is simply layered onto the current process, the organization may get more activity without real transformation. The tools may be more powerful. The work may look more modern. Employees may save time on certain tasks. But the underlying operating model may remain unchanged.
If the existing process is strong, AI can improve speed, quality, decision support, employee experience, and customer outcomes. If the existing process is weak, AI may scale the weakness.
That is why Existing+ should be treated as a starting point, not the destination.
Like any meaningful transformation, AI requires internal capability building, not just tool implementation. Organizations need people who can understand the work, test new ways of operating, learn quickly, and adjust as the technology and the work continue to evolve. This is where Lean and agile disciplines remain very relevant — not as slogans or ceremonies, but as practical ways to learn from the work, respond quickly, and improve the system.
The goal should not be to preserve every existing process and make it slightly faster. The goal should be to understand what work creates value, what work creates friction, and where AI can help redesign the flow.
Productivity Gains Are Not the Same as Enterprise Value
This distinction is now showing up across recent research.
MIT Sloan / MIT CISR recently summarized five things leaders still get wrong about AI. One of the central points is that organizations often mistake individual productivity gains for strategic business value. Quick uses of generative AI — writing emails, summarizing documents, drafting presentations — may help individuals save time, but they should not be confused with broader AI solutions that create enterprise value.
McKinsey’s 2025 State of AI research points in a similar direction. The organizations getting more value from AI are more likely to redesign workflows, define where human validation is required, and use broader operating-model practices across strategy, talent, technology, data, adoption, and scaling.
That matters.
The organizations getting more value from AI are not simply adding tools to work. They are redesigning the work around AI.
That is the difference between task productivity and business transformation.
A person may save 20 minutes preparing a summary. That is useful. But if the downstream meeting still happens the same way, decisions are still delayed, ownership is still unclear, and no one acts on the insight, the enterprise value may be limited.
The value is not in the AI output alone.
The value is in the action that changes because of it.
The Value Is in the Next Action
One reason I like the MIT CISR framing is that it pushes leaders beyond AI output and toward business outcomes.
In the MIT discussion of AI value, the logic is not simply “use AI to generate insight.” The sequence is more disciplined: collect the right data, generate insights, take action, create value, and then connect that value to economic outcomes.
That is a critical distinction.
AI may predict that a customer is likely to churn. Value is created only if the organization changes the next action in a way that improves retention.
AI may identify patients at risk of a fall. Value is created only if the care team changes the intervention and reduces preventable harm.
AI may flag an invoice exception. Value is created only if the workflow reduces delay, improves accuracy, or prevents the exception from recurring.
AI may recommend the next best action for a sales team. Value is created only if that recommendation is trusted, timely, embedded in the workflow, and tied to better customer outcomes.
In other words, AI does not create value because it produces an answer. It creates value when the organization uses that answer to change the work.
When Workflow Reality Is Underestimated
Not every process is ready for AI just because it looks repetitive.
McDonald’s experience with AI-powered drive-thru ordering is a useful cautionary example. The company tested automated order taking with IBM across more than 100 restaurants, but ended the test in 2024 after mixed results and customer complaints about order accuracy. McDonald’s has not walked away from AI broadly, but the drive-thru example is a reminder that real-world work contains variability that does not always show up in a clean technology demo.
A drive-thru order may look like a simple transaction. But anyone who has spent time around operations knows it is more complicated than that.
There is background noise. Customers change their minds. People speak with different accents. Orders are modified. Children shout from the back seat. Two cars may be speaking at once. Employees must recover quickly when something goes wrong. The experience matters because McDonald’s has built its brand around speed, consistency, and customer convenience.
The lesson is not that AI has no role in quick-service restaurants. The lesson is that automation has to be designed around the real conditions of the work.
If the process is variable, customer-facing, and time-sensitive, the design has to account for exceptions, recovery paths, human intervention, customer frustration, and quality control. Otherwise, AI may create more friction than it removes.
That is the hidden factory showing up in the customer experience.
What Better Looks Like
The stronger examples are not simply about using AI to produce faster outputs. They are about embedding AI into operational flow and decision-making.
The MIT CISR research offers a useful example from One New Zealand, a telecommunications provider shifting toward an AI-driven business model. In 2024, One New Zealand deployed 15 AI use cases that had reached a margin hurdle, with plans to expand significantly. Some early uses focused on customer support and marketing, including knowledge agents that helped resolve customer queries and marketing agents that accelerated audience segmentation.
But the more interesting example is operational.
During a major weather event, One New Zealand used task-based AI agents to verify power failures and cell status, understand battery capacity, forecast demand, estimate when generator support would be needed, and recommend optimal actions to decision-makers. According to the MIT briefing, work that previously took hours could be completed in minutes, helping the company improve decision-making and customer experience.
That is a different level of AI adoption.
It is not just using AI to write faster or summarize faster. It is embedding AI into the operational flow of the business. It is using AI to support decisions, improve response time, and create better outcomes.
There is still human oversight. MIT uses a helpful distinction between human in the loop and human at the helm. The human-in-the-loop reviews and approves decisions where needed. The human-at-the-helm sets goals, defines constraints, aligns incentives, establishes guardrails, and monitors outcomes.
That distinction is important. The future is not simply humans doing the work or AI doing the work. The real question is how work is designed so people and AI each do what they are best suited to do.
The Lean-First AI Test
This is where a Lean-first approach becomes especially useful. Before applying AI, leaders should pressure-test the work through four lenses: value, flow, people, and governance.
Value and outcomes
What customer or business outcome are we trying to improve?
What work actually creates value?
What decision or action should change?
How will we know whether value was created?
Process and flow
Where is the hidden factory?
What work exists only because the process does not work right the first time?
Where are handoffs slowing flow or decisions being delayed?
What should be eliminated, simplified, standardized, or redesigned before it is automated?
People and adoption
How will this change the work people actually do?
What new skills, behaviors, or judgment will be required?
Where could AI improve employee engagement by reducing friction or low-value work?
Are managers equipped to help teams adopt the new way of working?
Technology, data, and governance
Is the data reliable enough to support the use case?
What decisions can AI support, and what decisions require human approval?
What escalation path exists when the answer is uncertain or wrong?
What governance processes need to change so they guide responsible use without slowing every decision?
Who owns the outcome after AI is deployed?
These questions may sound simple. But in many organizations, they are not asked with enough discipline before technology is introduced.
Too often, the pattern is predictable: organizations buy the tool, search for use cases, push adoption, measure activity, and hope ROI will follow. A better pattern starts earlier: understand the work, expose the hidden factory, redesign the flow, and then apply AI where it improves judgment, speed, quality, employee engagement, customer outcomes, or business results.
That is the shift leaders need to make.
AI Needs an Operating Model, Not Just Use Cases
This is also why AI adoption needs an operating model.
Without one, AI work becomes scattered. Different teams experiment in different ways. Similar use cases are built more than once. Governance reacts after the fact. Risk controls vary. Data access is unclear. Success metrics are inconsistent. Employees may not know which tools to trust or how AI fits into their daily work.
That is how AI activity grows without AI impact.
An operating model does not have to mean bureaucracy. In fact, it should avoid bureaucracy. It should provide enough structure to answer practical questions: which use cases matter most, who owns the outcome, what data can be used, where human validation is required, how risks will be managed, what workflows need to change, what capabilities employees and managers need, and how value will be measured and scaled.
This is where an AI Center of Excellence, or a lighter-weight version of one, can help. Not as a technical team sitting off to the side, but as a way to bring consistency to decision-making, governance, prioritization, enablement, and scaling.
The point is not to centralize every AI decision. The point is to make sure AI is not just being adopted, but being translated into better work and better outcomes.
From Hidden Factory to Customer Outcomes
Lean has always pushed leaders to ask what creates value for the customer and what gets in the way. In the AI era, that question becomes even more important.
MIT CISR’s research on domain-oriented companies found that companies focused on customer outcomes, not just products and services, performed significantly better than product-oriented companies. Domain-oriented companies organize around end-to-end customer needs and measurable outcomes rather than only around internal products, channels, or functions.
That idea is very consistent with Lean thinking.
If AI is only used to make existing tasks faster, the organization may miss the bigger opportunity: redesigning work around the outcome the customer actually cares about. That could mean faster resolution of a service issue, better prevention of a health risk, more accurate financial guidance, more reliable supply chain decisions, more personalized customer support, or better use of employee capacity.
AI can help with all of that. But only if the organization understands the outcome it is trying to improve and redesigns the work accordingly.
Otherwise, AI becomes another tool layered onto the existing system.
And the hidden factory remains.
Final Thought: Accelerate Flow, Not Friction
AI is a powerful accelerator, but acceleration is only valuable if the system is pointed in the right direction.
If the work is unclear, AI may create more confusion. If the data is weak, AI may produce faster but less reliable outputs. If ownership is unclear, AI may add another layer of ambiguity. If workflows are broken, AI may make the broken work move faster.
That is why leaders should resist the temptation to start with the tool.
Start with the work. Find the hidden factory. Understand the flow. Redesign what needs to change. Clarify decision rights. Build trust in the data. Define where human judgment belongs. Then apply AI where it can genuinely improve speed, quality, decision-making, employee engagement, customer experience, or business results.
AI should not be used to automate waste. It should be used to help people and organizations create better outcomes.
That starts with a powerful discipline:
Fix the work before you automate it.
