We often talk about enterprises as a block. But in the story of AI and blockchain, that perspective obscures exactly where the change is happening. To see more clearly, I want to separate it into two layers.
One layer is responsible for coordinating work through transactions: finding partners, comparing terms, assigning tasks, monitoring, correcting errors, switching parties when necessary. The other layer handles the organizational part of the transaction: who is named, who holds the assets, who has the authority to act within what scope, and where the responsibility lies in case of issues. I am not saying Coase has separated it in exactly this way. I am just separating it like this to better see where AI and blockchain intersect.
My previous article mainly focused on the first layer. Coase explains the existence of enterprises from the cost of using price mechanisms. Using the market is not free. One must find prices, negotiate, draft contracts, and then deal with discrepancies in transactions. The chapter “The Coasean Singularity?” from NBER carries this idea into the context of AI agents: agents can reduce some search, communication, and transaction costs in the digital market. But at the same time, that chapter also retains an important barrier: agents can create additional friction, make pricing less clear, and the net impact remains an empirical question. Moving from “AI is better” to “enterprises will dissolve into the market” is moving faster than the data allows.
The missing part lies in the latter layer. Reducing coordination costs does not automatically create a new entity that can hold assets, act within a clear scope of authority, and leave a transaction history sufficient for the other party to rely on. In simpler terms, AI can thin out the coordination layer of the enterprise. But by itself, it is not enough to replace the organizational layer of the enterprise.
From here blockchain starts to make sense, and only in a much narrower sense compared to the familiar narrative. Its main value here does not lie in being “smarter.” It lies in the fact that a smart contract account can hold digital assets and act according to encoded logic. The official documentation of Ethereum describes smart contracts as a type of account with a balance, which can be the target of transactions and self-execute rules through code. The documentation on account abstraction goes even further to say that smart contracts can be used to hold assets and authorize transactions based on the specific logic of that account.
From this angle, blockchain does not turn software into a fully-fledged enterprise. But it can give software pieces of a minimal organizational layer in the digital domain. “Minimal” here should be understood very narrowly: holding digital assets, enforcing rules, authorizing actions, and keeping a transaction history on a shared infrastructure. It does not itself solve broad legal status, off-chain disputes, or responsibilities tied to physical assets. Therefore, the connection from technical capability to “pieces of organization” only stands when I keep it within that proper scope.
Saying this also helps to separate AI and blockchain from each other without going overboard. AI mainly makes certain coordination steps cheaper: finding, comparing, responding, and trading under sufficiently modular and measurable conditions. Blockchain is only worth mentioning when it adds a part of the infrastructure to the organizational layer: assets, rights, commitments, and verifications. These two do not do the same thing, even though they may meet in the same process.
From this framework, EIP-8004 is noteworthy not because it proves that agents will become enterprises. It is noteworthy because it shows that the infrastructure is attempting to solve a very specific problem: how to give agents identity, reputation, and verification when interacting across organizational boundaries without needing pre-existing trust. This proposal self-describes its goal as helping parties “discover, choose, and interact with agents across organizational boundaries without pre-existing trust.” It is still just a proposal, not yet a finalized standard. But even the way it frames the problem shows that the missing part is not only the capability to act. The missing part is a way for others to know who the agent is, how to evaluate it, and to what extent it can be verified.
From here, the argument of “software as firm” becomes clearer but also narrower. It fits better in digital, modular, and verifiable activities. It weakens when transactions require high specialized assets, have many exceptions, or when responsibilities are difficult to succinctly write into verifiable rules. This is where Williamson remains useful. If the value lies largely in governance, in incomplete contracts, and in decisions needing to be handled contextually, then keeping that activity within the enterprise still makes more sense.
Therefore, what is emerging right now may not be that enterprises are about to disappear. It is also not that blockchain is about to replace companies. What is worth retaining is a different perspective: enterprises are not necessarily a unified block. They are more like a bundle of functions. There are functions of AI that are eroding at the coordination layer. There are functions of blockchain that are attempting to recreate in the digital domain at the organizational layer. And there are also functions that still lack any replaceable infrastructure.
Looking at it this way, the question worth asking is no longer whether AI will replace people or whether blockchain will replace companies. The more pertinent question is: among the functions being grouped under the term enterprise, which functions are becoming cheaper at the coordination layer, which functions have the opportunity to be encoded at the organizational layer, and which functions still require governance in the traditional sense. Until those layers are separated, it is very easy to overstate both AI and blockchain.