Imagine a purchasing manager in a manufacturing company.
He does not just ask for prices. He has to find suppliers, compare terms, monitor deliveries, renegotiate when conditions change, and handle when a link does not fulfill its commitments. That position exists partly because continuously using the market for these tasks is still costly.
Ronald Coase views the enterprise from just that angle. His point is not that the market is useless. His point is that using pricing mechanisms is not free. The very act of finding prices, negotiating, forming contracts, monitoring, and resolving disputes is a type of cost. When those costs are high enough, some tasks performed within the business are cheaper than doing them through the market. This is a very important part of how he explains why businesses exist.
From here, the question about the AI agent is no longer whether it can replace labor or not. That question is too broad. What is more worth looking at is when it is just a new tool within the company, and when it begins to thin the boundaries of the company.
Here, 'thinning the boundaries' does not need to be understood as the company disappearing or personnel immediately decreasing. I am talking about a narrower and more observable matter: there are many processes that companies used to have to keep internally that can now be purchased from outside more frequently, under clearer contracts, and with less friction. In other words, the scope of work that the company must self-coordinate through internal mandates begins to shrink, while the part that can be handed over to the market starts to expand. This understanding aligns with the spirit of Coase and Williamson when viewing the market and the enterprise as two different ways of organizing coordination, each with its own costs and advantages.
That is why recent studies on AI agents are noteworthy. The chapter The Coasean Singularity? of NBER states clearly: at the market level, agents can create efficiency by lowering search, communication, and contracting costs. But this chapter also simultaneously states that agents can create additional friction such as congestion and make prices harder to see, while the net impact on welfare remains an empirical question. It also suggests that if the costs of eliciting true demand, contract execution, and identity verification decrease, the range of feasible market designs will widen.
But from that to conclude that the boundaries of the enterprise will shift still requires a step. That step lies right in the example of the purchasing manager at the beginning of the article. If the AI agent only helps him ask for prices faster, then the company is still just having a new tool. If that agent can compare suppliers, negotiate within allowed frameworks, monitor performance, and do those tasks cheaply, reliably, and verifiably enough, then the company will begin to have reasons to push part of the coordination out to the market instead of keeping everything internal.
The key point is not that the AI agent is 'smart' in the general sense. The key point is whether it can change the cost relationship between the two coordination methods. If it cannot change that relationship, the company boundary remains almost intact. If it can change it, the boundary will begin to recede. This way of framing the issue aligns with Coase at the foundational level, and also aligns with Williamson at the level of choosing governance structures suitable for each type of transaction.
According to that reading, the AI agent seems to easily thin the boundaries of the business first in those tasks that meet several conditions at the same time.
Usually, these are tasks that are sufficiently modular, measurable, and verifiable. This means that inputs and outputs are relatively clear, 'good' can be described by price, duration, error rate, or several quite specific criteria, and the results can be verified as right or wrong. Another condition is that the partner market must be sufficiently thick. With enough options to compare, substitute, and bargain, pushing coordination out to the market becomes a real possibility, not just a beautiful idea. And finally, the infrastructure for execution must be sufficiently stable. If searching becomes cheaper but verification remains expensive, or negotiation is possible but dispute resolution is still difficult, then the business still has reasons to keep that part of coordination internal. The chapter The Coasean Singularity? emphasizes quite strongly on layers such as contract execution, identity verification, and market design.
The opposite direction is also important.
There are places that are more likely to remain in the business longer. One is a transaction area associated with a high degree of asset specificity. Williamson emphasizes that when parties invest deeply in assets that are difficult to move elsewhere, that transaction is harder to fully hand over to the market. Two is where the goals are still vague, or cannot yet be written into sufficiently good action criteria. The strongest agent knows what it is optimizing. It is much weaker when even the organization has not clearly stated what 'good' means. Three is where decision-making authority and responsibility are difficult to separate. In this part of my article, I only consider it as an inference from transaction cost logic and governance structure, not yet a conclusion locked by broad empirical data.
Therefore, the understanding that seems most prominent at this moment is not 'AI will make companies disappear.' A narrower and more conditional understanding is: the AI agent can reveal parts of the company that primarily exist to compensate for still high transaction costs. The more standardized, measurable, and verifiable a part is, the more likely it is to be pushed out to the market sooner. The more it is associated with specialized assets, vague goals, and difficult-to-transfer responsibilities, the more reason it has to stay longer. This is a model looking out from Coase, Williamson, and the NBER chapter above. It is not yet a conclusive empirical closure.
Returning to the purchasing manager at the beginning of the article, what is worth observing is not whether he can be replaced or not. What is more worth looking at is which parts of his work primarily exist because using the market is still too expensive, and which parts exist because the company really needs to retain coordination rights internally. If the AI agent makes the first part much cheaper, the enterprise boundary may begin to recede. Not all at once, and not everywhere equally. But that is enough to change the way we look at the business: no longer as a fixed block, but as a boundary that is always redefined by coordination costs.