AI AGENTS ARE CHANGING WORK

AI agents that can plan, use tools, and iterate across multiple steps are moving from prototypes into production environments. The change is not uniform across the economy, but it is measurable in specific workflows inside large organizations.

The practical effect is a re-pricing of certain types of labor and a shift in what kinds of process improvements deliver the highest returns — a thread that runs through much of our AI coverage.

From tasks to workflows

Previous generations of automation, including early generative AI tools, primarily accelerated discrete tasks: drafting text, writing code snippets, summarizing documents. Agents differ because they can chain actions, call external systems, and maintain state across steps.

When a system can take a high-level goal, break it into sub-tasks, execute them using available tools, and adjust based on intermediate results, the unit of automation expands from the individual task to the end-to-end workflow. This changes the economics for the roles that previously owned those workflows.

Early deployments in customer operations, software engineering support, and financial analysis show agents handling sequences that previously required multiple people or significant human coordination. The same dynamic is visible in specialized domains like credit underwriting. The gains appear largest where the process is digital, the data is structured or semi-structured, and the cost of occasional errors can be contained.

Where do agents still fall short?

Agents struggle in domains that require deep judgment under ambiguity, negotiation with incomplete information, or coordination across groups with conflicting incentives. Stanford's AI Index has documented the underlying pattern for several years: benchmark capability rising quickly while reliability on long-horizon, open-ended tasks lags behind. These limitations are not primarily about model scale. They stem from the difficulty of encoding tacit knowledge and organizational context into reliable tool use and decision rules.

As a result, the roles least affected so far are those that combine technical skill with responsibility for outcomes that are hard to specify in advance. Senior engineering judgment, complex sales, and certain types of strategic work continue to command premiums even as agent capabilities advance.

The pattern is consistent with prior automation waves: the middle of the skill distribution in knowledge work is seeing the most direct pressure, while the tails — routine execution and high-judgment work — are less immediately displaced. The wage-arbitrage version of this story is already playing out in outsourcing economics.

What does this mean for capital allocation?

For builders and operators, the relevant question is no longer whether agents can perform a given task in isolation. It is whether the surrounding systems, data quality, and exception-handling processes allow the agent to deliver reliable throughput at acceptable risk.

Firms that treat agent deployment as a process redesign problem rather than a model selection problem are capturing larger gains. Those that simply layer agents on top of existing workflows are seeing more modest and uneven results.

For investors, the implication is that software companies whose moats rest on labor cost advantages in knowledge work face a different competitive dynamic. The value of owning the workflow and the data that trains and constrains the agent is rising relative to simply providing access to a general model.

Companies that control proprietary data and well-instrumented processes have a structural advantage in making agents effective. Pure model providers or low-context application layers have less leverage as the technology matures.

The Bottom Line

AI agents are compressing the cost and time of certain categories of knowledge work faster than many organizations anticipated. The effect is not uniform replacement of jobs but a reallocation of which activities remain scarce and therefore valuable inside organizations.

Builders who redesign processes around reliable agent loops, rather than bolting agents onto legacy workflows, are positioned to capture the productivity gains. Investors evaluating software and services businesses need to assess whether the company owns the data and process context that turns general models into durable, high-margin automation.