Even if 70% of AI Projects Fail, AI Will Not

AI projects often fail not because the technology doesn’t work, but because the incentives inside the institution are misaligned.

And yet, despite all that, failure might not even matter. Most information technology initiatives (up to 70 percent) seem to fail, and none of that diminishes the ultimate value.

And there are lots of institutional behaviors that might help explain risk aversion, not only IT project failure.

When risk is localized and visible while benefits are diffuse and hard to attribute, rational managers avoid experimentation even when the enterprise would benefit.

If an AI pilot fails, the sponsoring department (IT, operations, marketing) is clearly identifiable.

If it succeeds, gains often show up in ways that do not proportionally reward taking the risk:

• company-wide cost reductions

• improved customer experience

• better decision quality

• long-term competitiveness. 

Also, what is hard to measure is hard to reward. But failure or investment arguably often is measurable:

• budget overruns

• model bias incidents

• downtime

• compliance issues

• higher departmental spending

Successes, on the other hand, often are ambiguous or show benefit only at the “whole enterprise” level:

• “better decisions”

• “higher productivity”

• “reduced churn risk.”

Also, employee and manager rewards tend to be short term oriented, while AI improvements tend to be long-term.

Innovation	Who Bears the Risk	Who Receives the Benefit	How Rewards Are Structured	Likely Behavior
AI customer service chatbot	Customer support leadership	Entire company (lower costs, better customer experience)	Support measured on call resolution, customer satisfaction	Avoid deployment if early errors hurt scores
Predictive maintenance AI	Plant operations	Enterprise (reduced downtime, capex deferral)	Ops measured on uptime, safety	Resist pilot that might trigger false alarms
AI sales forecasting	Sales ops team	Finance, supply chain, executive planning	Sales measured on quota attainment	Ignore tool that might challenge forecasts
AI-assisted coding tools	IT budget owner	All product teams	IT measured on cost control & system stability	Delay rollout due to security concerns
AI fraud detection	Risk/compliance team	Entire firm (loss reduction)	Compliance measured on zero failures	Avoid model with false positives risk
AI hiring screening	HR	All departments (better hires)	HR measured on time-to-fill and legal risk	Reject tool due to bias concerns
AI pricing optimization	Pricing team	Company-wide margin gains	Team measured on revenue stability	Avoid experimentation that could reduce short-term revenue
AI knowledge management	IT, knowledge management  team	Entire workforce	KM measured on system uptime and adoption	Avoid radical change to workflows

The contrast between “fear” and “greed” explains quite a lot of human behavior, in personal life or in business. We might say the dichotomy is between expectations about “downside” and “upside” as well as the consequences of either happening. 

If the impact or results of a negative realized outcome happen right away, and are quantifiable, while the results of a positive outcome are delayed, intangible in the moment or hard to quantify, then risks might generally exceed rewards in many instances. 

All of which might help explain why artificial intelligence adoption seems to produce fewer immediate gains than we might expect. 

For individual decision-makers in business, downside risk from a model-driven decision, such as credit losses, customer complaints, or operational disruptions, is immediate and personally attributable. We know quickly what we have lost. 

Upside value, by contrast, is probabilistic, delayed, and diffused across the enterprise.

Rational actors therefore discount model recommendations. 

To increase the perceived margin of safety, adoption thresholds rise, discretionary review layers proliferate, and precedent asserts itself when change is called for. 

This behavior is often misinterpreted as cultural resistance or distrust in AI, but it reflects rational responses to incentive structures that punish mistakes more visibly and directly than rewarding successes.

In fact, where AI is able to better quantify risk, it might actually encourage inaction, especially if it is possible to surface specific entities, departments or functions where losses could occur, while benefits might accrue more broadly, and be harder to identify. 

This is analogous to the problem of cutting any government spending: the losers are easily identifiable (people who lose jobs) while the winners (taxpayers) are diffuse and anonymous. 

So AI return on investment results are the predictable outcome of deploying probabilistic technologies into institutions optimized for deterministic control, loss avoidance, and fragmented accountability. 

In other words, AI might create, highlight or accentuate a greater degree of uncertainty; allocate potential losses in clear ways while producing benefits that are more diffuse. 

The example is AI quantifying costs in specific departments, run by specific individuals, while producing enterprise-level profits or revenues that are not clearly attributable to any specific department or person. 

The risks of failure are clear and personal; the benefits non-attributable in the same way. 

So to increase the chances that AI adoption will work, entities can change incentives:

• Share credit across departments

• Tie success metrics to enterprise outcomes, not just local KPIs.

• Create “safe-to-fail” zones (innovation budgets and sandboxes reduce career risk)

• Centralize funding for cross-functional AI to avoid penalizing the initiating department

• Reward experimentation, not just outcomes

• Attribute value explicitly (use internal accounting to show which teams enabled enterprise gains).

Yawning about 6G?

Most people outside the communications industry will be unaffected and largely unconcerned about Mobile World Congress happening. But even many who are in the computing and software industries might not have to be too interested.

Yes, every next-generation mobile platform has featured higher capacity (speeds, bandwidth) and lower latency. Sometimes that really makes a difference, enabling new and compelling features. Text messaging doesn’t take lots of bandwidth, but it arguably was a “killer app” for 2G. Steaming video enabled by 4G networks might be in the same category.

Most observers might agree that proposed 3G apps actually did not emerge until the time of 4G. But most assessments of new killer features or apps since 4G have yet to emerge. 

We might note that the hoped-for advances often happen only in every other generation of networks, as hyped apps for any particular generation take longer to be commercialized than was hoped. 

That has led to some thinking that “every other generation” of mobile platforms is consequential (in terms of killer features and apps). So 2G and 4G were more consequential, 3G and 5G less so, with perhaps some expectation that 6G could be the platform that is more important than 5G. 

Platform	Theoretical Peak	Real-World Speed	Latency	The Pitch	What Actually Mattered	Verdict
2G	0.3 Mbps (EDGE: 384 Kbps)	0.05–0.1 Mbps	300–1000 ms	"Digital" wireless; wireless internet on your phone	SMS texting — arguably the most transformative app in mobile history. Basic WAP browsing (barely usable). MMS.	Genuine Leap
3G	7.2–21 Mbps (HSPA+)	0.5–3 Mbps	100–500 ms	"Mobile broadband" — internet everywhere, video calling	App stores became viable. Google Maps (basic). Email on the go. Social media feeds (early Twitter/Facebook). The original iPhone ran on 2.5G/3G.	Partial Win
4G LTE	150–1000 Mbps	10–50 Mbps	30–70 ms	"True broadband speeds" — replace home internet, HD video everywhere	Streaming video (Netflix, YouTube) became genuinely good. Uber/rideshare apps. Instagram, Snapchat, TikTok. Video calls (FaceTime, Zoom). Hotspot as home broadband backup.	Genuine Leap
5G (sub-6 GHz)	1–10 Gbps	50–300 Mbps	10–30 ms	"Connected everything" — AR, autonomous vehicles, smart cities, remote surgery	Faster hotspot. Marginally better congestion in stadiums/airports. No killer app has emerged for consumers after 5+ years.	Mostly Hype
5G mmWave	4–20 Gbps	1–3 Gbps (indoors: near zero)	1–5 ms	"Gigabit wireless" — fixed wireless broadband replacement	Fixed wireless home broadband in specific markets. Dense venues. Not useful for mobile users — signal doesn't penetrate walls or travel more than a few hundred feet.	Mostly Hype
6G (proposal)	1 Tbps	???	<1 ms (theoretical)	"Holographic communication," digital twins, connected senses, brain-machine interfaces	Unknown. Researchers candidly admit there is no identified 6G killer app.	Mostly Hype

But we might also be at a point where speeds and feeds simply matter less, as the value of the mobile access connection is less driven by bandwidth and latency, and more by device and app capabilities. 

Perhaps it always is true that the value is driven by “what the platform enables” rather than “bandwidth” the platform supports. That has been true, arguably, for decades, as broadband internet access has gotten better.  

But we might also be at a point where, generally speaking, the networks support “more than enough” bandwidth, and “better than required” latency for most useful consumer or business use cases.

What Ails U.S. Student Math and Reading Skills?

Some might claim educational technology is to blame for declining U.S. student capabilities in math and reading. That probably is not the most-common explanation, however. 

U.S. student performance in reading and mathematics has been tracked primarily through the National Assessment of Educational Progress (NAEP). Overall, scores showed steady improvements from the 1970s through the early 2010s, peaking around 2013. 

Since then, performance has stagnated or declined, with sharp drops during the COVID-19 pandemic (2019–2022) due to school disruptions. Test scores indicate high school scores for the 2024 graduating classes have dropped. 

Many reasons have been advanced, but some might point to a couple of decades of declining scores. 

Post-pandemic recovery has been uneven: math has shown slight rebounds in some grades, while reading continues to decline. 

These trends are more pronounced among lower-performing students, widening achievement gaps. By 2024–2025, 12th-grade scores reached historic lows, with reading 10 points below 1992 levels and math at its lowest since 2005.

Study/Author(s)	Year	Key Proposed Reasons
Scammacca et al	2019	Initial proficiency levels create Matthew effects (higher initial skills lead to faster growth); demographic factors (socioeconomic status, ethnicity) influence starting scores and growth rates; grade-level differences affect trajectories, with lower performers growing slower over time.
Wyckoff (Annenberg Working Paper)	2025	Pre-2013 gains slowed due to policy shifts (reduced accountability post-NCLB); Great Recession funding cuts; Common Core disruptions; rising smartphone/social media use; demographic shifts (more English learners); pandemic accelerated existing declines, especially for low performers.
Malkus (AEI Report)	2025	Similar to Wyckoff: accountability rollback, funding reductions, smartphone proliferation, and out-of-school factors (student well-being) explain stagnation since 2013; pandemic worsened trends but isn't the sole cause; factors outside school (e.g., screens) play a major role.
DeBord (Dissertation)	2026	Post-pandemic: classroom factors (rebuilding routines, teacher adaptability); school collaboration; family/district supports; persistent gaps in foundational knowledge, reduced struggle tolerance, technology habits, and socioemotional needs hinder recovery.
NWEA Brief	2021	COVID-specific: remote learning disruptions led to below-pre-pandemic gains; greater impacts on math than reading; inequities in access (devices, support) affected lower performers more; recovery requires addressing missed foundations.

But there may be lots of other cumulative reasons, including social promotion (advancing students to the next grade despite inadequate performance to maintain age-appropriate grouping) and efforts to avoid stigmatizing learners (through reduced retention or labeling to prevent emotional harm), may contribute to declining academic achievement in U.S. students. 

These practices are often adopted to mitigate short-term social-emotional risks like low self-esteem or dropout likelihood from grade retention, but research suggests they can exacerbate long-term performance issues by allowing students to advance without mastering foundational skills, leading to compounded learning gaps, disengagement, and lower scores on assessments like NAEP.

This is particularly evident among lower-performing students, where declines have been steepest since the early 2010s. It’s complicated.

AI Workloads Might Reach 50-70% of all Data Center Workloads by About 2030

By 2030, AI operations might represent half of all data center workload. But some analysts believe AI operations will account for as much as 70 percent of total data center workloads by 2030. 

source: McKinsey

At the very least, we might note that new data center construction (about 75 percent of new capacity)  now focuses on high-performance computing capabilities. 

High-performance computing sites might never number more than conventional data centers. But capability and “locations” are different metrics, as are “hyperscale” sized facilities and “typical” data centers.