AI transfers information control from agents to systems, making the trust premium contestable. Bundled roles get repriced once observation shifts to a system of record. What becomes measurable gets commoditized. What remains illegible becomes scarcer. The result is structural unbundling, beginning with the capital allocator.

I. The Mechanism

The entities essay showed that four roles recur across capital markets — providers, allocators, middlemen, mercenaries — not by convention but because they are the only contracts that hold together when the people doing the work control what the people paying for it can see.

Two problems follow from this structure: completion-based pay rewards action regardless of fit, and time-based pay rewards activity regardless of outcome. Both persist because writing better contracts requires information the principal does not have.

Software did not fix this. It created systems of record without transferring information control. The agent still decided what to enter, what to report, and how to frame it. The principal got a dashboard built on agent-curated data.

AI is different — and the distinction is architectural, not incremental. Software required agent-mediated input: the agent decided what to type into the field, how to categorize it, what context to include. AI processes the raw artifact — the email thread, the lease PDF, the recorded call — without requiring the agent to structure it. The agent is no longer the bottleneck between raw reality and the system of record.

This is what it means to say AI transfers information control from agents to systems: the system can read unstructured data and make sense of it without the agent deciding what to include. The agent loses control of the narrative. The principal can see what happened, whether decisions made sense, and who caused what.

An early consequence: you can replace trust with rules. Instead of “trust the agent to make good decisions,” define the criteria, monitor compliance, enforce consequences.

A concrete example. A tenant representing 12 percent of NOI calls about a broken HVAC in July. The lease expires in eight months. A co-tenancy clause would trigger a 5 percent rent haircut if the tenant leaves. The PM is paid to close tickets. The $24,000 repair exceeds the approval threshold. The PM does not have access to the lease admin system; the context lives across scattered systems, emails, and a filing cabinet — agent-controlled.

An AI layer pulls the lease terms, reconciles notes, links prior repairs to renewal outcomes — system-controlled. The PM’s decision becomes auditable. Pay shifts from time toward outcome. Risk moves from none to partial. Information rights move from agent-controlled to system-controlled.

When this kind of transfer happens at scale, the contract frontier — the set of contracts enforceable cheaply enough to be worth using — expands. Previously impractical contracts become viable: outcome pay for task-level workers, partial risk for bounded executors, post-close accountability for intermediaries. These contracts were not impossible in theory. They were impossible to enforce because the information needed to verify them did not exist in any system the principal controlled.

Legibility creates comparability, comparability creates commodity competition, and commodity competition compresses margins.

This is what happens every time information moves from the agent to the system. When Bloomberg transferred price information from dealer-controlled to system-controlled, dealer spreads compressed — but Bloomberg captured $25B+ in terminal economics. The same pattern repeated in electronic trading, programmatic advertising, legal tech. In every case, the entity that built the information infrastructure captured disproportionate value. The measured parties got competed and compressed.

II. The GP and Securitization 2.0

The mechanism is general. The clearest place to see it play out is the general partner in real estate private equity.

The GP exists as a trust aggregator. They bundle deal sourcing, underwriting, asset management, disposition, and LP reporting into a single package. The LP pays 2 and 20 for the whole bundle because they have no way to see which component generated alpha. Each component is opaque — the GP controls the information.

AI transfers information control on each component. You can specify, monitor, and benchmark asset management decisions. Contribution attribution becomes feasible at the component level. LP oversight can be continuous rather than quarterly.

Once the principal can observe per-component contribution, the bundled fee becomes contestable. The role does not disappear. It gets repriced and narrowed to what remains genuinely hard to measure: sourcing relationships that depend on personal trust, judgment on novel situations, the ability to structure deals that do not fit templates.

The Securitization Parallel

This has happened before.

Securitization 1.0 asked: Can we standardize the cash flows? Securitization 2.0 asks: Can we standardize the decision-making?

Securitization 1.0 took the bank’s bundled role and disaggregated it by transferring information control from the bank to the structure. The servicer became a commodity — interchangeable, replaceable, paid on basis points. The market expanded because every conforming loan had a guaranteed buyer.

The exception proves the structure: in the first European NPL securitization post-crisis (ERLS NPL1, Lone Star, 2017), the servicer was the central question — Lone Star had to acquire specialist servicers, present their recovery methodology directly to bond investors, and convince the rating agency that their workout expertise justified the structure. The role was not a line item. It was the deal. That distinction — commodity servicer for simple assets, differentiated servicer for complex ones — is the key to the analogy.

If AI can transfer information control on management decisions from the GP to a system, then what is being securitized is not just a payment stream but an operating protocol. The ERLS NPL1 structure already moved in this direction — hard-coded sale floors, governance constraints, reserve triggers — because post-crisis investors refused “trust me” servicing discretion.

Securitization 2.0 generalizes that move: specification replacing discretion, enforced by system rather than structure. The investor buys exposure to “this building operated according to this specified standard, with continuous system verification, and automatic replacement mechanisms if standards are not met.”

Both securitizations also restructure time horizons: 1.0 turned a 30-year bank relationship into a tradeable security; 2.0 turns a 10-year blind pool GP commitment into a rolling mandate with continuous monitoring.

The binding constraint on the best GPs is not capital — it is cognitive throughput. If AI removes that constraint, the parallel to Fannie Mae is direct: when a capacity constraint is removed, the market expands. Deals previously too small to bother with become viable. More structured fund products drive more capital into the asset class.

But securitization 1.0 also demonstrated the moral hazard risk: when origination separates from ownership, quality can collapse. Securitization 2.0 must make sure someone still has skin in the game on the hard calls.

III. The New Landscape

The GP does not wake up one morning as a servicer. The transition has phases.

First, monitoring: LPs get system-controlled visibility; the GP bundle is intact but the trust premium starts to erode.

Then, repricing: observable components are benchmarked; fee pressure begins.

Then, disaggregation: commodity functions migrate to specialized providers; the GP narrows to judgment.

Phase 1 is already happening. Phase 2 is beginning. Phase 3 is the structural prediction.

The transition produces specific new contract types. The strongest is the Accountable Executor — outcome pay, partial risk, policy-bounded, system-controlled info. The GP-as-servicer: executing a defined mandate, monitored continuously, compensated on system-verified outcomes, replaceable for underperformance. Previously unstable because you could not pay task-level workers on outcomes without granular attribution. System-controlled information makes attribution feasible.

Other new types include the Direct Principal (the LP who bypasses the GP in monitored domains — already exists at sovereign wealth funds; AI extends it downmarket) and the Accountable Intermediary (the middleman with system-enforced post-close risk — viable for high-volume transactions).

A structural constraint applies to all: you cannot make someone bear risk they have no power to control. Partial risk with system attribution works. Push higher without granting discretion and the contract falls apart.

Three parties compete for the value created:

Principals get better alignment and lower costs, but become dependent on whoever owns the information infrastructure.

The infrastructure owner captures disproportionate value — not just as a toll booth but by determining which contract types can exist. The Accountable Executor is only stable if the system provides attribution; turn off the infrastructure and those contracts collapse back to the old four. The long-term analog is Fannie Mae: setting the standard, taking economics on every transaction.

Judgment holders — the small number of people whose contribution remains permanently illegible — command higher premiums because everything around them got cheaper.

This cuts two ways. Top GPs with genuine judgment see margin compression on commodity components but volume expansion as cognitive throughput ceases to be the bottleneck — they do more deals across a larger portfolio, capturing premium on the judgment residual.

The median GP, whose value is mostly trust aggregation, compresses in both margin and volume — commodity components reprice to market rates and volume does not compensate because the median GP was never judgment-constrained.

The infrastructure owner captures the expansion surplus. This is the Bloomberg pattern: total market activity grew, but dealer economics shrank while infrastructure economics grew.

IV. Where This Might Be Wrong

The 2008 lesson. When specification replaces discretion, you remove a certain kind of accountability. If the model breaks, nobody exercises judgment — because the structure assumed specification was sufficient. A biased monitoring system makes the same bad call in every building at once. Securitization 2.0 needs a clear answer to “what happens when the system fails?”

Goodhart’s law is recursive. Any monitoring system becomes the target of optimization. The Accountable Executor optimizes for measured outcomes, not actual outcomes. The thesis holds only if the monitoring side iterates faster — plausible in high-volume workflows, uncertain in complex ones.

And the monitoring infrastructure itself is not immune: if GPs are the paying customers, the system may optimize for what GPs want to see rather than what LPs need to know. Who the infrastructure serves determines whether the Goodhart problem is solved or just moved up a level.

Incumbent absorption. Large GPs may internalize the information tooling, preserving the bundled structure while adopting the technology internally. If so, the change happens inside firms, not between them — the GP role narrows without being disaggregated.

The judgment question. The framework predicts that illegible judgment becomes more valuable and that the GP’s bundled fee compresses. Both can be true if most of the current fee is trust-aggregation rent. But if the judgment component is larger than the thesis assumes, the compression is modest and the GP survives largely intact.

If these constraints dominate, the thesis becomes narrower and slower. But the direction holds.

V. The Question the Framework Cannot Answer

The structural logic holds. The mechanism checks out across the full space. The direction is clear.

What the framework cannot answer is the question that determines the magnitude of everything above:

How much of the capital allocator’s current value is trust aggregation on functions where information control is transferable, versus genuine judgment that no system can replace?

If the answer is 80/20, the repricing is dramatic — the median GP’s fee structure collapses, new fund structures emerge rapidly, and the market expands significantly.

If the answer is 40/60, the repricing is real but moderate — fees compress on the observable fraction, the judgment premium increases, and the GP role narrows without fundamentally changing.

The answer varies by strategy, by asset class, by market cycle. It will only become visible as system-controlled information is actually built and adopted — as information control actually transfers and the market reveals how much of the current structure was trust aggregation and how much was genuine judgment.

That boundary — between what is transferable and what is not — is the whole game.

Research Map: Sources and claim-by-claim citations for this essay