The Productivity Paradox: Why Most AI Spend Is Disappearing Into Thin Ai

When I started my career at FedEx, I spent years designing supply chains using lean principles. We mapped value streams, hunted for waste, removed bottlenecks, and optimized the flow of work end to end. It was the discipline of the era, and it worked. Companies that took lean seriously pulled away from the ones that did not, and the capital investment was no different. Same dollars. Different mindset. Different outcome.

That was the early 2000s. The tools we had at the time defined the ceiling of what was possible. We optimized within the limits of what humans, paper, EDI, and early enterprise software could do.

Here is the part most leaders are missing in 2026. A process designed in 2010 should not be how you design a process today. AI and agents do not just give you a faster version of the old workflow. They give you the chance to throw the old workflow out and reinvent the work entirely. Most companies are taking option one. The few who take option two are the ones quietly running away with the value.

That is the productivity paradox in a sentence.

The Numbers Behind the Disappointment

In January 2026, PwC's Global CEO Survey found that 56% of CEOs reported zero measurable ROI from AI in the past twelve months. Not modest gains. Zero. From the technology that has dominated every earnings call, every board agenda, and every capital plan for three years running.

Apollo's Chief Economist Torsten Slok borrowed a line from Robert Solow and put it bluntly: "AI is everywhere except in the incoming macroeconomic data."

The data behind the disappointment is consistent across every credible source.

McKinsey's State of AI found that 78% of firms now use AI in at least one function and 71% regularly use generative models. Adoption is essentially saturated. Yet only about 6% of companies are capturing what McKinsey calls outsized AI value. The rest report task-level wins that fail to convert into EBIT.

Workday's January 2026 productivity report measured the leak directly. For every ten hours of efficiency companies gain through AI tools, roughly four hours get lost fixing AI outputs. The phenomenon now has a name: workslop. The people paying that tax are the most engaged employees, the ones using AI most often. Workday calculated 1.5 weeks per highly engaged employee, per year, just fixing AI work.

Forbes summarized the macro view. Task-level productivity gains of 14% to 55% are real. Total productivity growth is projected at only 0.5% to 0.7% over the next decade. Micro wins. Macro silence.

This is not a vendor problem. The models are getting better every quarter. This is an operating model problem. Most companies are using 2026 technology to patch 2010 workflows, and they are surprised the math does not work.

Optimize vs. Reinvent

Here is the trap I see almost every executive falling into. They take their existing process, the one they built or inherited a decade ago, and they sprinkle AI on top of the steps that look automatable. A copilot here. A summarization tool there. An agent that drafts the email a human used to draft. The work still flows the same way. The handoffs are still in the same places. The org chart still has the same boxes. AI just makes a few of the existing steps faster.

That is optimization. It is what I used to do at FedEx, and within the constraints of that era it was the right answer.

In 2026 it is the wrong answer. The constraints that shaped those workflows do not exist anymore. Knowledge work no longer has to be sequential. Approvals no longer have to wait for a human inbox. Documents no longer have to be drafted, reviewed, and edited in five separate handoffs. Customer service no longer has to be tiered. Reports no longer have to be assembled by analysts pulling from six systems.

The lean principle still holds. Find the waste, remove the constraint, optimize the flow. But the canvas is completely different. Designing an AI-native process and then comparing it to a 2010 process is not apples to apples. It is a different shape of work entirely.

When you optimize a 2010 process with 2026 tools, you get a 5% improvement and a giant verification tax. When you reinvent the work assuming agents and AI exist, you get a fundamentally different operating model. That is where the 6% are playing.

McKinsey's data makes this concrete. Out of 25 attributes they tested for ability to drive EBIT impact from generative AI, the single biggest factor was workflow redesign. Only 21% of generative AI adopters have fundamentally redesigned at least some workflows. The other 79% are running AI as a productivity hack at the edge of unchanged processes. That is why their gains evaporate before they reach the income statement.

The Cost Cutting Trap

There is a second mistake that compounds the first one, and it is even more expensive.

When lean came to American manufacturing, the companies that misunderstood it treated every productivity gain as an excuse to cut headcount. The companies that got it right, Toyota being the cleanest example, did the opposite. Toyota famously refused to lay off workers during downturns, and when a process improved, they retrained and redeployed people into the next problem. The savings funded growth. The talent compounded. The institutional learning compounded. Their lead got bigger every year.

Watch what is happening with AI right now and the parallel is almost too clean.

Klarna replaced roughly 700 customer service agents with AI and projected around $40 million in annual savings. Within fourteen months, customer satisfaction had cratered on complex disputes, repeat contact rates spiked, and the company was quietly rehiring humans and pivoting to a hybrid model. The CEO publicly admitted, "we focused too much on efficiency and cost." The savings turned out to be smaller than projected, the rehiring costs were enormous, and the brand damage is hard to put a number on.

Klarna is not an outlier. Forrester's Predictions 2026: Future of Work report found that 55% of employers regret their AI-attributed layoffs, and roughly half of those layoffs are expected to be reversed within eighteen months. Google has already disclosed that one in five of its AI software engineer hires in 2025 were former employees coming back. The pattern is repeating at Microsoft, IBM, Salesforce, and Meta.

Treating AI as a cost cutting program is the most expensive way to use it. You get the perceived benefit once. Then it is gone. You laid off the people who carried the institutional knowledge, you damaged customer relationships you cannot easily rebuild, and you handed your competitors a recruiting pipeline. The savings are a sugar high. The damage is structural.

The companies that win do the opposite. They take the productivity gains and reinvest them. New products. New customer experiences. New markets. Higher value work for the same people. The savings get recycled into growth, and the talent gets redeployed into harder problems that AI cannot do alone yet.

The Compounding Advantage

This is where the math turns brutal for the cost cutters.

When you treat AI as a value creation engine instead of a budget item, something happens that the spreadsheet does not show on day one. The flywheel starts to turn.Better data leads to more trustworthy outputs. More trustworthy outputs lead to greater adoption. More adoption leads to deeper integration. Deeper integration produces richer proprietary data. The loop accelerates.

World Wide Technology framed it by tracking what AI leaders do with their gains. The top reinvestment priorities are expanding existing AI capabilities (47%), developing new ones (42%), and strengthening cybersecurity and data foundations (41%). They are not banking the savings. They are channeling them right back into the system that produced them. Each cycle compounds.

IBM's internal AI program saved an estimated 3.9 million employee hours in 2024 and is on track for $4.5 billion in annual productivity savings by the end of 2025. They did not get there by buying licenses or cutting jobs. They got there by industrializing AI inside their core processes and putting the savings back to work.

Novo Nordisk's enterprise platform supports more than 25,000 employees who have built over 2,500 use cases. One regulatory document workflow that used to take a 50-person team fifteen weeks now takes minutes. The company has been explicit that the program is about elevating employees into higher-value work, not eliminating them. The same people now ship more, faster, into harder problems.

The cost cutter takes the win once. The value creator turns the win into the next win, and the one after that. Five years from now, the gap between the two will not be 5% or 10%. It will be the kind of gap that shows up on org charts, market caps, and obituaries.

That is what separates the companies that will win from the companies that will lose.

What Leaders Should Do This Quarter

If you are in the 56% reporting zero ROI, here is where I would start.

• Pick three workflows that matter and reinvent them, do not just optimize them.Order to cash, claims processing, content production, whatever drives real economic value in your business. Ask the harder question. If we were designing this process from scratch today, with AI and agents available, would it look anything like what we have? Then build that.
• Treat productivity gains as growth fuel, not severance. When AI frees up capacity, redeploy the people. Move them to the next bottleneck, the next product, the next customer problem. The companies that recycle the gain into growth pull away. The companies that bank it as a layoff pay for it twice.
• Build the flywheel deliberately. Capture every output, every override, every outcome as a data product. Feed it back into the next model, the next agent, the next workflow. AI value compounds when you design for compounding. It evaporates when you do not.
• Stop counting seats. Start counting outcomes. Build an AI value scorecard that tracks cycle time, error rates, revenue per FTE, and direct cost takeout. If you cannot put a dollar sign on it within two quarters, kill it and reallocate.
• Name an executive owner for AI value capture. Not the CIO alone. A cross-functional owner with the authority to redesign processes, reassign people, and shut down work that no longer makes sense. Without that role, you will keep funding pilots that never reach production.

The Bottom Line

The lean principles I cut my teeth on at FedEx still matter. Map the value stream, remove the waste, own the outcome. Those ideas are not going anywhere. What has changed is the canvas. We used to optimize within a fixed set of constraints. AI takes those constraints away and lets you reinvent the work itself.

Two things separate the companies that will pull away from the ones that will fall behind. First, they reinvent processes instead of optimizing the old ones. Second, they treat AI as a way to create value, not a way to cut costs.

The cost cutter wins the quarter and loses the decade. The value creator turns every productivity gain into the next one, and the next one, and the next one. The advantage compounds until it is not a gap anymore. It is a moat.

Solow's paradox eventually resolved. So will this one. The question is whether your organization is going to be in the 6% capturing the value, or in the 56% writing checks for nothing.

The technology is not the bottleneck. It never was.

Patrick Phillips is CIO at Vasion and Adjunct Professor at Westminster College. He writes The AI Catalyst to help leaders navigate the real complexities of enterprise AI, not the hype cycle.