AI Spending Fails at 95% of Companies, Why Returns Lag

The AI capex trade still looks brilliant from 30,000 feet. On the ground, it is rather less glamorous: most companies are spending hard, shipping pilots, and getting little back for it.

That mismatch is now harder to ignore. Fresh reporting, drawing on MIT-backed research and broader market analysis, suggests roughly 95% of companies investing in generative AI are seeing minimal or no material return. Even more telling, about 74% of the value created is being captured by just 20% of firms. The spread is brutal, and it says less about AI being useless than about corporate adoption being badly executed. ^[1] ^[2]

Enjoy articles without ads?

The headline number is ugly, but the distribution matters more

The easiest version of this story is that AI is overhyped. That is not quite right. A better reading is that returns are highly concentrated, and most companies are still stuck in the expensive experimentation phase.

The MIT-linked findings point to a winner-takes-most pattern. A small cohort of firms is converting AI investment into measurable productivity gains, margin improvements, or new revenue. Everyone else is mostly paying for models, cloud capacity, consultants, and internal reorgs while waiting for proof that any of it scales. ^[3]

That 74% to 20% split matters because it reframes the debate. This is not a case of zero value being created across the economy. It is a case of value accruing unevenly to companies with the right data, workflows, technical talent, and operational discipline. Put plainly, plenty of management teams bought the narrative before they built the plumbing.

Giancarlo Leaves Willkie for Crypto, AI Focus

Why the money is not turning into returns

Too many pilots, not enough production

A recurring theme across the research is that companies have launched generative AI pilots at speed, but very few have pushed those experiments into core operations. A chatbot in a sandbox is easy enough. Rewiring a customer support stack, legal review process, software pipeline, or sales workflow around AI is slower, messier, and much more expensive.

That leaves firms with a portfolio of demos rather than a business case. The spending shows up immediately in software bills and hiring costs. The payoff does not. ^[4]

Bad use cases are soaking up capital

A lot of AI spend has gone into projects that look impressive in a board deck but do not solve a high-value problem. Firms chased novelty over unit economics, hoping that "doing something with AI" would signal innovation to investors and clients.

The companies seeing better results tend to focus on narrow, repeatable tasks where the baseline process is already measurable. Think code assistance with clear productivity benchmarks, document triage, fraud monitoring, procurement automation, or internal knowledge retrieval. Those use cases are not always sexy, but they are easier to quantify. Corporate AI theatre, as ever, is less investable.

Data quality is still the real bottleneck

Generative models can be strong, but most enterprise data environments remain chaotic. Information is siloed, inconsistent, poorly labeled, or trapped in systems that do not play nicely together. That means companies often plug advanced tools into weak infrastructure and then wonder why performance disappoints.

If the underlying data is unreliable, the output will be too. Worse, low-quality responses in regulated or customer-facing environments create legal and reputational risk, so firms end up keeping humans in the loop. Sensible, obviously, but that also limits the labour savings used to justify the original spend.

Labour and process change are harder than model access

There is a persistent assumption that access to frontier models is the hard bit. It is not. The real challenge is changing how people work.

AI tools can alter job design, reporting lines, compliance procedures, and incentive structures. That requires management buy-in beyond the innovation team. Many firms underestimated the organisational friction involved, especially where workers need retraining or where managers fear loss of control over existing processes. Buying AI is easy. Getting a large company to use it properly is another matter entirely. ^[5]

Why a few firms are winning

The companies capturing outsized gains tend to share a handful of traits. First, they have proprietary data that makes AI systems more useful and defensible. Second, they are embedding AI into existing workflows instead of treating it as a bolt-on gadget. Third, they are measuring outcomes closely, whether through time saved, error rates, customer conversion, or revenue per employee.

Scale helps too. Large firms can spread fixed AI costs across broader operations, negotiate better compute terms, and attract scarcer technical talent. That creates a compounding advantage. Once a company gets one use case into production successfully, it usually becomes easier to deploy the next.

This helps explain why investors have rewarded infrastructure providers and a narrow set of enterprise beneficiaries more than the average AI adopter. The picks-and-shovels thesis has held up better than the "everyone instantly gets more productive" thesis. ^[6]

The market is still paying for possibility

Public markets have largely priced AI around future operating leverage, not current broad-based corporate returns. That leaves a gap between equity narratives and enterprise reality.

For now, spending continues because companies fear being left behind more than they fear near-term inefficiency. There is logic to that. If AI does become foundational, delaying adoption could be costly. But fear-driven capex is still capex, and boards will eventually want evidence.

That is where the next phase begins. Rather than asking whether a company has an AI strategy, investors are likely to ask which workflows have been automated, what KPIs improved, how hallucination risk is controlled, and whether gross margins actually moved. The market has heard enough about transformation. It will want receipts.

What this means for crypto and adjacent tech

Crypto has seen this film before. Capital rotates into a big narrative, infrastructure wins first, and downstream adoption takes longer than promised. The current AI cycle looks similar in one important respect: tokenisation of the theme is easier than proving utility.

That does not mean AI-linked crypto projects are doomed. It means the same filter applies. Projects with real data access, defensible distribution, and clear demand for compute or coordination may have a case. Tokens riding pure "AI x blockchain" vibes probably do not. If 95% of large corporates are struggling to monetise AI despite huge budgets, smaller speculative projects should not get a free pass.

Bittensor TAO Spread Hits 27.1%

Risks to consider

There is a chance the current disappointment is simply a timing issue. General-purpose technologies often take years to deliver visible productivity gains because firms need to redesign processes around them. If that is the case, today's weak ROI may look like an early deployment tax rather than a structural failure.

Still, there are less friendly interpretations. AI costs may remain high, competitive advantages may be fleeting if everyone uses the same base models, and regulatory constraints could keep human oversight in place longer than optimists expect. That would compress the return profile for many adopters.

What to watch next

Whether firms shift spending from pilots and consultants to production-grade workflow tools
How quickly boards demand ROI metrics tied to margins, output, or headcount efficiency
Which sectors show repeatable gains first, especially software, finance, legal, and customer service
Whether AI infrastructure names keep capturing the economics while adopters lag
How much of the current spend survives if macro conditions tighten and cheap experimentation ends

The punchline is simple enough: AI is creating value, just not for most companies yet. Billions have been deployed, but the gains are landing with a relatively small group that had the data, discipline, and distribution to make the tech useful. Everyone else is still paying tuition.

AI Spending Fails at 95% of Companies