The Verifiable Enterprise

From Job Replacement to Organizational Compression

Much of the debate about AI displacement asks which jobs agents will replace. The more important question may be which layers of the organizational chart they will compress.

Agents’ Last Exam, a benchmark led by Berkeley RDI and cited in OpenAI’s GPT-5.6 release, evaluates agents on more than 1,500 long-horizon professional tasks across 55 fields. These are not multiple-choice questions. Agents must use professional tools to produce finished work, and performance is evaluated against verifiable outcomes. OpenAI reports that GPT-5.6 Sol achieved a benchmark score of 53.6, although full end-to-end reliability remains considerably lower and the hardest tier remains far from saturated.

The useful conclusion is not simply that “agents are not ready.” Their capabilities are unevenly distributed across workflows. Agents are increasingly effective in the broad middle, such as collecting information, conducting first-pass analysis, preparing documents, updating systems and coordinating routine communication. They remain less reliable at the expert tail, where work shifts from processing what is already known to exercising judgment under ambiguity in real life.

This suggests that agents may compress teams before they replace complete jobs. A professional’s role typically combines execution, coordination, review and judgment. Agents may absorb much of the execution and coordination while leaving final judgment and accountability with a smaller number of experienced people.

For operators, the relevant metric is therefore not the number of AI subscriptions. It is the share of labor hours spent on tasks with machine-checkable acceptance criteria. The largest productivity gains come from the company’s ability to translate expert judgment into explicit evaluation systems.

For investors, the diligence question is not simply, “How good is the company’s AI strategy?” It is: “What percentage of its labor cost is attached to tasks that can be reliably verified?” Reconciliations, ticket resolution and code are relatively inexpensive to transform because outcomes can often be checked automatically. Client-facing judgment work is more expensive because the acceptance criteria remain inside experts’ heads.

There is also a second-order problem. The bottom of the organizational pyramid has historically served as the industry’s apprenticeship system. Junior employees perform repetitive work, observe senior judgment and gradually become decision-makers themselves. Compressing that layer raises an uncomfortable question: where will the next generation of judgment come from?

Instrumented work environments could become part of the answer. Juniors can practice on real or simulated cases, receive immediate feedback on verifiable components and escalate ambiguous decisions to experts. Apprenticeship does not disappear, but it becomes more deliberate, measurable and compressed. Firms that build this capability may own not only a better automation system, but also a stronger talent pipeline.

From Static Data to Experience Loops

In real-world deployment, the scarce resource is not access to models or more scraped data. It is the ability to define what “done correctly” means.

That knowledge typically lives inside domain experts. They understand which steps matter, which errors are acceptable, when an exception should be escalated and what distinguishes a merely complete output from a good one. The challenge is to convert that tacit judgment into verifiable acceptance criteria that a system can apply consistently.

Static datasets depreciate as models improve and every AI lab distill from one another. Live operations are different. Every acceptance, rejection, correction and escalation can generate a proprietary signal. When those signals are captured and used to improve evaluation, routing, workflows and, where appropriate, model training, the company develops an experience loop that compounds over time.

The organizational design follows from the loop. Senior experts define standards and handle ambiguous cases; agents perform much of the execution; and a smaller operating layer monitors quality and manages exceptions. In some workflows, an organization that once required 500 people may eventually require only 50, not because every employee adopts an AI assistant, but because the organization itself is redesigned around machine execution and human judgment.

The objective is therefore not simply AI adoption. It is to build a closed feedback loop in which live work continuously improves the system while preserving a high density of useful signals.

A company that captures thousands of weak or ambiguous interactions may learn less than one that captures a smaller number of clearly verified outcomes. Signal quality matters more than raw data volume. The asset is not merely the company’s historical archive, but its ability to generate new, decision-relevant feedback through everyday operations.

This is why experience loops, in another term reinforcement learning environment, can become more durable than static data. The archive records what the company knew in the past. The loop determines how quickly and cheaply it improves in the future.

From Reasoning to Action

For most of the past decade, training and inference were separate. A model was trained through an expensive, occasional process and then deployed into production, where inference generated outputs but rarely contributed directly to further learning. That separation is beginning to narrow.

Agentic reinforcement learning connects the two through a tighter feedback loop. An agent performs real tasks in an executable environment, its outcomes are verified, and those verification signals inform subsequent improvements. As the system becomes more capable, it can take on increasingly difficult work. The competitive question is therefore not only who has the best model, but who can run this loop fastest and most economically.

The o1 era demonstrated that reinforcement learning could teach models to reason for longer on static, verifiable problems. The next phase is about reasoning in order to act. Its infrastructure is considerably more difficult. Agents must interact with browsers, codebases, databases and enterprise applications, often waiting for tools or maintaining state across long-range tasks.

Training and inference therefore need to be integrated at the learning level but decoupled at the systems level. Otherwise, an agent waiting for a test to finish, a page to load or an external application to respond can leave expensive training infrastructure idle. The system should keep training resources fully utilized while thousands of agents operate asynchronously across stateful environments.

Two assets become particularly important.

First, proprietary task streams with verifiable outcomes. A reconciliation balances or it does not. A claim is approved, rejected or escalated. A customer accepts the deliverable or requests revisions. Unlike a static archive, these tasks recur through the normal operation of the business, creating a continuously refreshed stream of feedback.

Second, executable environments. Agents need stable and realistic places in which to perform work using the same tools, information and constraints that human professionals face. The quality of these environments depends on their realism, stability, difficulty distribution, feedback richness and ability to scale.

By comparison, the harness, the wrapper around the model, may provide a less durable advantage. Better prompting, routing and tool orchestration can improve performance, but these techniques are relatively easy to replicate. Proprietary workflows, verification signals and execution environments are much harder to reproduce because they are embedded in a company’s operations and real-world interactions.

The emerging AI moat may therefore belong to companies that do more than deploy models. They own the work, control the environment, observe the outcomes and continuously turn those outcomes into better performance.

If agents compress teams before they replace jobs, then the near-term unit of disruption is not the worker—it is the org chart. And org charts, unlike jobs, are something investors can own, redesign and reprice.