Let’s take a look at Ann, a settlement analyst at a broker-dealer. A broker-dealer is a firm that executes stock trades for clients and makes sure the trades settle, i.e., the stocks and cash change hands. On a typical day, Ann needs to diagnose why some trades fail to settle, which happens with roughly 2% of trades. She has several tabs open: the order management system, the clearing portal, the custodian’s website, multiple email threads and spreadsheets. For each failed trade, she figures out what caused the failure, fixes the records, and notifies the counterparties.
Where does the logic for resolving failed trades live? It’s not in any of the systems; rather it lives in Ann’s head, and in the seams between applications. The logic has never been written down in its entirety; a rudimentary version of it exists in a training manual that was last updated two years ago, but there are a myriad of corner cases that only Ann or a colleague can resolve. The company’s smooth functioning relies on hundreds or thousands of routines of this nature. The company’s day-to-day operations are in effect the execution of these people-mediated routines.
Agents are about to change the status quo. In the companion whitepaper attached to this post (Tractability of Enterprise Semantic Reconstruction), we argue that a properly designed system can, given a sufficiently long trace of Ann’s work, learn the business logic that lives in Ann’s head. And similarly for the thousands of routines the company relies on, using activity traces and data exhaust.
Much of Ann’s tedious repeated work can be replaced by a suitable agent, and if you are reading this blog, you are likely already persuaded as we are that repeated digital work done by humans will be among the first dominoes to fall. The key question is where the logic lives. In the straight line agentic approach, the logic will end up in model weights, in a vendor’s context store, or in a platform skill, none of which the company owns.
The enterprise would gain the labor savings from automation, but in turn would end up losing the knowledge of its own functioning that it may have built up over years or decades and distributed across its people and applications. What lies at the end of this road? The enterprise becomes a tenant in its own home: the system that embodies its operations. While the starter rent is a sweetheart deal, it is practically a law of the universe that rents go up once the lock-in happens. This is a Faustian bargain no firm can refuse, because to not automate using AI is to get left behind.
We propose an alternative scenario: Cognitive Sovereignty.1 In this approach, the enterprise owns the computational model of how it works. Moreover, the model is encoded in an explicit form (e.g., procedural code or declarative rules) such that the enterprise can inspect the model, test it, make changes, and run it on whatever infrastructure it chooses. The provider sells the machinery that learns, validates, and monitors execution, as a continuous learning process. The enterprise keeps what was learned. Nothing enterprise-specific is embedded in model weights outside the enterprise.
The Computable Enterprise
True cognitive sovereignty imposes strict requirements on how the automation model works.
Explicit and Deterministic Logic. Business logic is captured as procedural code or declarative rules over business objects, each procedure or rule annotated with the cases it was validated against. The logic runs deterministically, and AI models are used only to handle new inputs, exceptions, or modifications to the routine itself. Incidentally, this approach also has the benefit of lowering inference costs dramatically over the pure agentic approach, because the model is called upfront to learn the routine and thereafter only on exceptions, rather than on every run.
Modifiable. Modifications happen through a continuous learning process as corner cases are surfaced and resolved by business people. They can also happen explicitly by the enterprise changing the way it does business e.g., changes in pricing or terms or even whole workflows.
Portable. The logic does not require a specific vendor’s runtime infrastructure, and the rule and procedural languages are open. This is a key requirement to make sure the solution doesn’t just push the dependency from one kind of vendor to another.
We call this vision the Computable Enterprise. The day-to-day operations of the business are encoded in explicit logic, while people set its purposes, oversee operations, maintain human relationships, and handle corner cases and situations that the system is unable to handle on its own. Why is the Computable Enterprise vision achievable now, but not earlier? Because of two key enablers, both recent:
Models can reconstruct enterprise semantics, at least for a large set of workflows.
Code has become nearly free to write and test.
So far we have viewed cognitive sovereignty, and the consequent Computable Enterprise, as a defensive move to keep control of the company’s soul. But it is more than defensive -- it can make enterprises staggeringly more efficient. SpaceX, in its IPO filing identified enterprise applications to be a $22.7T opportunity. Anthropic, in its filing mentions that knowledge workers are valued at $40T annually (before the first of the 10 times the filing uses the word “catastrophic”). The routinizable bits that we are looking to automate are only a part of the work that knowledge workers do, but in banking and healthcare about 30% of costs go into middle and back office administrative and operational roles, so reasonably routinizable work still constitutes trillions of dollars of expense that can be reduced.
A still greater payoff is the logical next step: the Superintelligent Enterprise. A computable enterprise can reason about itself. This enables the emergence of superintelligent corporations where humans provide oversight, judgment and direction, like mission control to the AI’s astronauts. We’ll take up the superintelligent enterprise in a future post, where we explore both how it might transform current firms and enable the creation of de novo superintelligent firms. For now, imagine being able to do things like change a business policy and immediately see what impact it might have; or have swarms of agents continually try to find ways to cut costs or improve SLAs.
Let us return to Ann. If the future plays out along the lines we discussed, how does Ann make a living? What about the human cost of automation? These are civilizational questions, and we don’t think anyone, including us, has good answers yet. We hope to address this topic in its own post, and please subscribe to be emailed when we do.
Companion Whitepapers
We provide three companion papers that go into much more detail about the Computable Enterprise. Roughly speaking, each of the three papers tackles a potential objection to the thesis, and collectively they lay out a rigorous intellectual framework for the Computable Enterprise.
How is this different from what’s happening as the “AI transformation” today with agents? Why won’t the current incumbents, especially the Robotic Process Automation companies do this, and can one or more high-scale companies be built in this space? Part 1 (Towards the Superintelligent Enterprise) tackles these questions and the civilizational impact superintelligent enterprises could have.
Can you really recover business rules reliably from activity traces and data exhaust? Part 2, Tractability of Enterprise Semantic Reconstruction, describes how we can produce Business Logic as a Service (BLaaS) for a large set of enterprise workflows.
Where does the training data come from, because enterprises certainly won’t share it with a startup? Part 3, Data Acquisition, lists several possible approaches, and argues that two might be companies in their own right.
If You’re In the Trenches
If you are the kind of person who actually read through and made sense of the whitepapers, or if you’ve done Ann’s or a similar job and would like to share your insights, we’d love to hear from you in the comments. If you are a founder or a prospective one, and find these kinds of topics irresistibly interesting, like we do, write to us at investors@foldingspace.ai or use this link. Tell us what you’re building or thinking of building, and share one or two of the coolest or most technically challenging things you or anyone on your founding team has ever done. We’d especially welcome links to your work, such as GitHub repositories, research papers, or demos. We wrote this essay, but we mostly write checks, have funded great companies, and would love yours to be one of them. And if you like to hear more from us, do subscribe.
Acknowledgments
We would like to thank Uri Kolodny, Elias Bareinboim, Sriram Viswanath, Vishal Misra and Sonu Aggarwal for their comments and discussions.
Strictly speaking, this is enterprise cognitive sovereignty, to distinguish it from the national kind ("sovereign AI"), and from data sovereignty, which is about where the bytes sit. We will use the short form for convenience.