AI Made It Possible · Economic Reality

The AI Was Real. The Snake Oil Was the Story.

What if the biggest problem in the AI boom was never that the technology did nothing — but that useful technology was packaged inside promises the operating evidence could not support?

By Gary Whittaker · Founder and Operator, JackRighteous.com

I keep coming back to a simple problem with the public AI conversation.

We spend enormous amounts of time looking at what the models can do. We spend much less time looking at what happens when businesses actually try to run them.

Those are not the same question.

A demo can be spectacular. A benchmark can improve. A model can write, code, reason, summarize, generate images or complete a complex task.

Then a company has to deploy it.

That is where the boring numbers begin.

And the boring numbers may eventually tell us more about the AI boom than another prediction about when superintelligence arrives.

63% is not just another AI statistic

KPMG reported in its Q1 2026 AI Quarterly Pulse that 63% of the large U.S. organizations it surveyed now require human validation of AI-agent outputs.

One year earlier, the figure was 22%.

KPMG described that as a near tripling.

Important limitation: this is not a survey of every business. KPMG's U.S. Q1 sample covered 237 C-suite and business leaders at organizations with at least $1 billion in annual revenue, with more than a third representing companies above $10 billion. It is evidence about large-enterprise operating behavior, not a universal 63% rate for all businesses.

That deserves attention because the dominant story around AI agents has been movement toward greater autonomy. Yet the businesses deploying them are increasingly building human validation into the operating model.

KPMG found the same organizations were also wrestling with the economics and execution: 65% cited difficulty scaling use cases as a barrier to demonstrating ROI and 62% cited skills gaps. KPMG also reported that 57% of leaders expected people to manage AI agents.

Read KPMG's Q1 2026 AI Quarterly Pulse →

That does not mean AI agents are useless.

It means the operating reality is more complicated than the sales language.

These are the numbers I call the “advanced stats”

They are not actually advanced.

Twenty or thirty years ago, this was simply the type of information serious business coverage was expected to examine.

Revenue matters. Growth matters. Adoption matters. But eventually you have to ask what sits underneath them.

Those questions become especially important when headline adoption races ahead of successful enterprise scaling.

McKinsey reported in June 2026 that almost 90% of organizations said they were at least experimenting with AI, but only 7% reported scaling it across the enterprise.

Read McKinsey's operational-excellence research →

Gartner reported in January 2026 that by the end of 2025, at least 50% of generative-AI projects had been abandoned after proof of concept, pointing to poor data quality, inadequate risk controls, escalating costs and unclear business value.

Read Gartner's GenAI project-failure analysis →

Again, none of that proves AI is failing.

It proves something more important:

“We are using AI” and “AI is producing sustainable economic value at scale” are different measurements.

Snake oil can contain real ingredients

This is why I use the snake-oil metaphor carefully.

There is even a historical irony here. The original Chinese water-snake oil brought to the United States by Chinese railroad workers was used as an anti-inflammatory. The phrase became synonymous with fraud later, especially through products such as Clark Stanley's heavily marketed “snake oil” liniment, which federal investigators found did not contain snake oil at all.

That history actually sharpens the analogy for me: something associated with a real useful ingredient can become the basis for a much bigger sales story.

Read the Smithsonian history of the snake-oil metaphor →

I am not claiming that there is nothing valuable in the bottle.

The individual ingredients can be extremely useful.

AI can write. It can summarize. It can create music and images. It can assist programmers. It can search through information. It can automate parts of a workflow. It can help one person operate with capabilities that previously required a larger team.

I know that because I use those capabilities myself.

The issue begins when a useful capability becomes a much larger promise.

At that point, we are no longer discussing only the ingredient.

We are discussing how somebody packaged it, what they promised it would cure and what financial incentives existed to keep the story moving.

The technology can be real while the economic story wrapped around it is exaggerated.

Nobody needs to meet in a secret room

This is also where discussions about incentives get unnecessarily dragged into conspiracy.

I do not need a secret meeting to understand how a system can create overly optimistic narratives.

A startup founder wants investment. A venture capitalist needs outsized winners. A public company needs growth. AI laboratories need enormous amounts of capital and compute. Cloud providers sell infrastructure. Chip companies sell the machinery. Consultants sell transformation. Governments want competitiveness. Media organizations compete for attention.

None of those incentives automatically proves misconduct.

That is precisely why the system deserves scrutiny.

When almost everyone in an ecosystem benefits from a bigger, faster and more inevitable future, optimism itself becomes economically useful.

The problem does not have to begin with criminality.

It can begin with something completely normal: people doing what their roles reward them for doing.

The question is what happens when the incentives keep pushing after the evidence starts becoming more complicated.

One symptom does not establish the diagnosis

One failed project proves very little.

One hallucination proves very little.

One company discovering that a deployment costs more than expected proves very little.

One KPMG survey certainly does not prove an AI bubble is bursting.

The useful question is whether the symptoms become persistent, elevated and connected.

Human validation requirements rise sharply.

Enterprise scaling remains far behind experimentation.

Projects die after proof of concept.

Cost visibility remains incomplete.

Approval and governance layers expand.

Organizations discover that the “autonomous” system still needs integration, security, monitoring, exception handling and people who understand when it is wrong.

Those are not individually a verdict.

Together, they are exactly the kind of underlying conditions that should be followed if we want to understand whether expectations and operating reality are diverging.

The spreadsheet eventually gets a vote

Imagine the original business case.

Ten people perform a process.

A system appears capable of automating most of it.

The first spreadsheet looks incredible.

Then deployment begins.

You need integration. Clean data. Security. Monitoring. Testing. Escalation rules. Exception handling. Human validation. Compliance. People who understand the tool well enough to recognize when a confident answer is wrong.

You may still save an enormous amount of money.

That possibility matters.

But you may also discover that the cost model is very different from the one implied by the original pitch.

KPMG's Q2 2026 numbers make this visible. 66% of surveyed organizations had monitoring dashboards and 61% had approval processes, while only 26% had access to real-time AI cost insights.

Read KPMG's Q2 2026 AI Quarterly Pulse →

Think about what that means.

Organizations are building the control layer faster than they are achieving complete visibility into the economics.

That is not an anti-AI statistic.

It is a business statistic.

The business problem AI cannot explain away

There is another part of this conversation that matters: explainability.

An organization can use AI to produce an answer. That does not automatically mean the organization can explain why a consequential decision happened.

Generative systems can produce explanations. But a fluent explanation is not the same thing as a faithful causal record of the process that produced the underlying output.

Businesses still have to answer human questions.

Why was this customer rejected?

Why did the agent authorize this?

Why did the system choose this supplier?

Who approved the recommendation?

What evidence was used?

Who takes responsibility when it is wrong?

So another layer grows around the system: logging, provenance, monitoring, permissions, validation, escalation and accountability.

This is why my broader argument in AI Is Not the Wizard matters here. The machine can be powerful without becoming the accountable actor.

The most dangerous number may be the denominator

I am becoming increasingly suspicious of adoption statistics presented without their denominator.

“Most companies are using AI.”

Using it for what?

“AI agents are being deployed.”

How many complete an entire workflow without intervention?

“Employees are more productive.”

How much more productive after verification and correction are included?

“The model costs less.”

What does the complete operating system cost?

“The company reduced headcount.”

Did operating performance improve?

The numerator gets the headline.

The denominator is often where reality lives.

And now we are talking about superintelligence

This is where the public discussion becomes almost surreal.

At the same time businesses are learning they need better monitoring, more governance, human validation and clearer cost controls around today's systems, we are also being asked to contemplate superintelligence and human extinction.

Those future-risk questions deserve serious examination.

But they are not the same question as whether today's systems are producing the economics already priced into them.

Future capability can become a kind of fog around present performance.

If people are repeatedly told that vastly superior machine intelligence is inevitable, then today's flaws can be mentally discounted as temporary.

The hallucinations are temporary.

The supervision is temporary.

The high costs are temporary.

The weak scaling is temporary.

The business case will arrive later.

Maybe it will.

But temporary is not an economic measurement.

That is why I keep returning to evidence standards in Superintelligence: How Would We Actually Know? and If AI Is Going to Kill Us, Show Me the Evidence.

Maybe we have been looking for the wrong extinction event

We keep asking whether AI becomes so intelligent that humans lose control.

There is another risk that requires no sentient machine at all.

Humans can make enormous economic, political and social decisions because they misunderstand what AI can actually do.

Companies can reorganize work around assumptions that prove wrong.

Governments can design policy around forecasts that do not arrive on schedule.

Markets can price distant theoretical capability as though it were near-term economic reality.

Organizations can delegate authority before building accountability.

Workers can be displaced before the systems replacing parts of their work are dependable enough to carry the whole responsibility.

None of that requires AGI.

None of it requires superintelligence.

It requires people.

AI may be the fog, not the perpetrator

That is where I currently land.

AI is the spectacular object everyone can see.

Behind it are older forces: capital, competition, power, politics, status, career incentives, shareholder expectations and the human tendency to turn a technological possibility into an inevitability.

AI did not invent those forces.

It may simply be the latest technology through which they operate.

That does not make AI unimportant.

It makes understanding it correctly more important.

What would prove this argument wrong?

If I am going to treat these as underlying indicators instead of collecting statistics that confirm what I already suspect, then the argument needs conditions that could weaken it.

I would expect to revise this thesis if we begin seeing several things happen together:

• enterprise-wide AI scaling rises substantially while failure and abandonment rates fall;
• measurable ROI becomes common rather than concentrated among a small minority;
• total operating costs fall after governance, security, integration and verification are included;
• agent reliability improves enough that human validation requirements decline in comparable workflows;
• companies demonstrate durable productivity gains without simply shifting hidden work into review, correction and exception handling;
• the economic case becomes less dependent on forecasts of future autonomy and more supported by current audited outcomes.

If those indicators move decisively in that direction, then the “gap between story and operating reality” gets smaller.

That matters because the purpose of tracking these numbers should not be to prove a bubble. It should be to find out whether there is one.

Follow the operating reality

So I am going to keep following a different category of AI statistics.

Not just adoption. Successful deployment.

Not just investment. Return on investment.

Not just agent launches. Intervention rates.

Not just benchmark scores. Real-world reliability.

Not just model prices. Total operating costs.

Not just jobs theoretically replaced. Human labour actually removed, changed or added.

Not just capability. Accountability.

The KPMG number does not prove an AI bubble is bursting.

Neither does Gartner.

Neither does McKinsey.

But the same broad operating tensions are now appearing from different directions.

That is not a reason to declare victory for the boosters or the skeptics.

It is a reason to investigate.

AI did not write the contracts.

AI did not set the valuations.

AI did not make the forecasts.

AI did not choose the headlines.

People did.

That is a human story.

And perhaps that is the AI story we should have been investigating all along.

Sources

• KPMG — Q1 2026 AI Quarterly Pulse
• KPMG — Q2 2026 AI Quarterly Pulse
• McKinsey — Putting AI to Work: The Operational Excellence Imperative
• Gartner — Why 50% of GenAI Projects Fail
• Deloitte — AI Agents Are Only the Beginning
• KPMG Global AI Pulse Q2 2026 — ROI, cost visibility and accountability
• Smithsonian — How Snake Oil Became a Symbol of Fraud and Deception

Reporting note: “Snake oil” is the author's editorial metaphor for the way AI capabilities have been packaged and marketed. The cited survey findings do not establish fraud, coordinated deception or motive. They document operating conditions that are relevant to evaluating the economics and claims surrounding AI deployment.

AI made it possible.
What we do with that possibility is still human.