A16Z: 80% of AI budgets are wasted; management is the next trillion-dollar opportunity

AI adoption continues to rise, but a significant portion of budgets is effectively sitting idle. Research shows that companies with high AI usage have seen a 10.2% increase in employee headcount. Leading enterprises’ AI spending per employee is projected to rise from $20,000 in 2025 to $225,000 in 2026—far exceeding labor costs. Hebbia’s CEO, an a16z portfolio company, notes that 80% of AI budgets are wasted due to poorly defined tasks and inefficient agent loops. For example, a code migration task with clear instructions costs just $4, but when instructions are vague and loops are inefficient, it costs $310. In the AI era, management capability has become a core scarce resource. Companies must identify AI use cases with 100x leverage, establish evaluation systems (Eval), and design incentive mechanisms for knowledge transfer. The greatest AI opportunity in the future is not foundational models, but companies that enable enterprises to sustainably navigate AI transformation through superior management capabilities.

Article author and source: WeChat public account "Silicon Base Observation Pro"

Stacking tokens疯狂 is not as effective as managing AI well.

Many people believe that artificial intelligence should replace human jobs.

The actual outcome, however, is exactly the opposite.

AI has not eliminated jobs; instead, it has created them. A study by Ramp Economics Lab of 21,000 U.S. companies found that two years after adopting AI, firms with high AI usage saw a 10.2% increase in employee headcount, while companies with low AI spending remained largely stagnant.

The reason is simple: human costs are lower than AI costs.

Ramp data shows that among leading companies, AI spending per employee is projected to rise from approximately $20,000 per employee in early 2025 to $225,000 per employee by the end of 2026, far exceeding the average U.S. employee salary and software engineer compensation.

Why does this phenomenon occur?

Recently, George Sivulka, CEO of Hebbia, an a16z portfolio company, wrote an article offering an unexpected assessment of this phenomenon: 80% of AI budgets are running in idle mode.

And the solution he proposed is very simple: a centuries-old craft—management—that emerged 185 years ago.

This article is compiled from George Sivulka, CEO of Hebbia’s article titled “You Just Hired a Million Bad Employees,” published on July 15, 2026. Below is the compiled content:

When organizations encounter problems, the most common response is not to redefine the problem, but to add more resources.

Previously, it was about adding people; now, it’s about adding tokens.

The model's response is poor; if one inference isn't enough, run multiple rounds. On the surface, this appears to increase intelligence density. In essence, it's no different from the traditional company's "manpower strategy"—both rely on resource input to mask unclear task definitions.

The token itself is never the issue. The problem is that most people don't know which information is worth giving to AI.

True high-quality context is one that clearly explains the process and understands the goal of the task. This is, in fact, a form of management skill.

If a task isn't clearly defined, more tokens won't bring more intelligence—only more expensive chaos. As a result, many companies face a rather ironic situation: AI budgets continue to rise, but the number of actual problems solved doesn't increase proportionally.

They purchased more "digital employees" without increasing corresponding management capabilities.

Many agent systems feature what appears to be a clever design: the model performs the task, evaluates the result, and then makes adjustments based on the feedback, repeating this cycle until the goal is achieved.

This mechanism works well when tasks are clearly defined and evaluation criteria are transparent. However, in real business scenarios, cycles often amount to nothing more than another form of unproductive meetings.

Models often retry repeatedly because we never clearly defined the goal from the start. AI can only continuously generate, reflect, discard, and regenerate, using more computational power to compensate for the lack of clarity in the objective.

How much does this cost? a16z provided a set of estimates: for the same codebase migration task, clear instructions cost $4 to complete; ambiguous, looping instructions cost $310.

During the first-day test of Fable 5 in June 2026, the cost per unit for the same task varied by as much as 17 times. What does $310 mean? It’s close to the total labor cost of an average American employee for an entire day.

And the task it delivered is the same as the one for $4.

In short, it’s no different from a group of people holding consecutive meetings on an ambiguous issue—the first meeting reached no conclusion, so they decided to hold another; the outcome of the second meeting was that more people needed to attend the third.

Humans consume man-hours in meetings; agents consume tokens in loops. The underlying problem is identical: managers have failed to define the problem.

A common phenomenon in large companies is that processes, originally created to solve problems, eventually become problems that need to be maintained themselves.

One approval node generates another approval node; a department proves it needs more staffing; a middle layer continuously creates complexities that only it can coordinate. In the end, the organization consumes vast resources merely to sustain its own operation.

AI systems are also subject to this inflation.

One agent is responsible for breaking down tasks, one for executing them, one for reviewing, and another for evaluating the review results. Each additional layer may make the system appear more complete; however, without rigorous validation of benefits, they might simply be generating work for each other.

People create more people; tokens create more tokens.

This means that, in the future, the core metric for managing AI will be how much verifiable business value each unit of Token generates.

In the future, excellent AI managers will also excel at eliminating unproductive reasoning, loops, and context. Token efficiency will become the new organizational efficiency.

The one true irreplaceable advantage of AI is scalability.

Replicating the abilities of an outstanding employee was nearly impossible in the past—you had to hire, train, empower, and accept the inevitable differences between individuals. But a proven AI workflow can be replicated ten thousand times in an instant.

This is also where the judgment that "humans are cheaper than software" is most easily misunderstood.

From a single-task perspective, an experienced human may be cheaper than an AI that relies on trial and error. However, at scale, once a company identifies truly effective contexts, workflows, and evaluation systems, the marginal cost of high-quality tokens will still rapidly fall below that of human labor.

The key lies in identifying AI use cases with 100x leverage.

The previous generation of tech companies competed for "10x engineers"; the next generation will compete for business contexts and management systems that can enhance AI efficiency by 100x.

Another often underestimated issue with managing AI is that a company’s most valuable knowledge is usually in its employees’ minds.

An experienced salesperson knows when to make a concession; a customer service manager can tell from a single sentence whether a customer is likely to complain; a supply chain manager knows which supplier promises the fastest but delivers the least reliably.

These insights are difficult to codify into standard processes, yet they determine a company’s true competitive advantage.

The AI transformation requires employees to organize their experience into context that models can understand and utilize. This is not merely a technical task—it directly impacts the internal power structures within the organization.

For centuries, possessing knowledge that others didn’t have has been a form of job security. Medieval guilds protected trade secrets, and modern employees safeguarded their “experience.” Now, companies expect employees to fully hand over their unique methods to AI, while simultaneously telling them that AI is merely a tool to boost efficiency.

Employees, of course, know what this means.

No one will unconditionally nurture a successor who might replace them. Therefore, corporate AI transformation ultimately enters the deep waters of organizational politics: who should own the knowledge, how to measure contributions, how to redistribute benefits after efficiency gains, and why employees should be willing to cooperate.

Therefore, designing an effective incentive mechanism for knowledge transfer will be crucial for AI implementation.

Why has AI made the fastest progress in programming? An important reason is that code naturally comes with evaluation criteria.

But the vast majority of work in the real world doesn't have such standards.

Therefore, the most important task in implementing AI is establishing an evaluation system, or Eval. The extent of AI adoption largely depends on the measurability of the work being tracked.

Eval is like AI's OKR, but it's more specific, breaking down vague business judgments into rules that machines can understand. The model determines the upper limit of AI's capabilities, while Eval determines whether a company can consistently turn those capabilities into results.

Over the past few years, the primary value in the AI industry has been concentrated in foundational models, computing power, and the application layer. Everyone is selling shovels and discussing which services will be reinvented by AI.

But these discussions obscure a larger reality: there are already enough models and applications—the real scarcity lies in the ability to reliably run them within enterprise core processes.

There’s now a popular view in Silicon Valley: traditional companies have complex organizations, intense internal politics, and rigid processes, making it impossible for them to truly undergo an AI transformation. The future belongs to new companies that rebuild their businesses with AI from day one.

That's only half correct.

AI-native companies are indeed leaner and more agile in adopting new technologies. However, traditional enterprises still hold the most critical assets: real customers, distribution channels, industry licenses, historical data, and business expertise that hasn’t yet been documented.

These assets won't automatically lose value just because a new model is released. Instead, those who can transform them into workflows accessible to AI will unlock tremendous productivity.

Therefore, the largest AI company in the future may not be another foundational model company, nor necessarily an “AI-native service provider” aiming to replace all traditional enterprises, but rather a company that helps businesses continuously undergo AI transformation.

At their core, they sell a continuous management capability: streamlining processes, extracting knowledge, designing context, setting up evaluations, controlling token costs, defining human-machine boundaries, and constantly rewriting business operations as model capabilities evolve.

This is also what makes Palantir most noteworthy. On the surface, it sells software. In reality, it’s more like selling the ability to transform complex organizations into computable systems. The software is merely the vehicle; the transformation is the product.

The AI era will amplify this demand tenfold.

AI transformation is not a one-time project. For every use case implemented, businesses discover ten new ones; with each improvement in model capability, existing processes warrant redesign. The more AI is used, the more AI needs to be managed.

This is a classic Jevons paradox: the more efficient the technology, the greater the overall demand for the related resources and services. In other words, managing AI will be the concluding phase following the boom of large models—and it itself will be the most important industry of the next stage.

In the past, every technological revolution ultimately became a management issue.

In the 1830s, railroads expanded rapidly, with U.S. railroad mileage increasing by approximately 120 times over the decade. Technology pushed transportation capacity to an unprecedented level, while also pushing existing management practices to their limits.

In 1841, two trains collided in Massachusetts due to a scheduling error. The accident revealed that it was not the steam engines that were insufficiently advanced, but rather that the system had become so complex that relying solely on human experience could no longer ensure safe operation.

Railroad companies subsequently began dividing regions, appointing managers, defining responsibilities in writing, and establishing clear reporting relationships. The modern corporate management practices we take for granted today gradually took shape with the expansion of the railroad network.

Railways first created capacity, and later developed methods to manage that capacity.

AI is repeating the same process.

Large models have turned "intelligent supply" into a resource that can be scaled almost instantaneously. In the past, adding ten people required hiring, training, and integration. Now, adding ten thousand agents requires only a single adjustment to scaling.

But the easier the supply is to expand, the greater the cost of poor management. In other words, in the AI era, management is becoming increasingly important.

This is a new management discipline in the making—and the next trillion-dollar opportunity.