\section{Introduction}
\label{sec:intro}

Economists have studied artificial intelligence mostly as something that is \emph{produced with}: a capital good that
performs tasks, a technology that raises productivity, a force that displaces labour. This paper studies AI as
something that \emph{decides}. Households and firms are beginning to delegate economic decisions---how to allocate
savings, when to trade, what to order, how much to borrow---to autonomous agents built on a handful of foundation
models. The question we ask is what changes in general equilibrium when the decisions themselves, and not only the
tasks, are made by machines that many principals share.

The mechanism we isolate is simple and, as far as we can tell, has not been written down. Economic decisions are
taken on forecasts of payoff-relevant states. A human's forecast error is idiosyncratic: across a continuum of
households it washes out by the law of large numbers, so it costs only the person who makes it. A decision delegated
to an agent built on foundation model $m$ carries model $m$'s error, which is \emph{common} to every principal that
uses the same model. Delegation therefore has two faces. Privately, it raises forecast precision and saves time, at a
fee priced in the compute market. In the aggregate, it replaces a noise that averages out with one that does not:
capital accumulation, asset demand and credit start to load on the shared error with weight $D$, the delegation
share, and $1/M$, the effective number of independent models. Wages, interest rates and asset prices inherit what we
call \emph{model noise}. A market for delegation thus sets the error-covariance structure of the economy.

Two results organise the paper. The first is a benchmark that goes against the most common policy narrative. In a
price-taking economy with constant returns and complete markets, the private value of delegating a decision equals
its social marginal value \emph{identically}---for every delegation share, every number of models and every degree
of strategic interaction (Proposition~\ref{prop:benchmark}). Model noise is real, but it is not an inefficiency:
delegators, knowing that others act on the same signal and that capital accumulation is a game of strategic
substitutes, dampen their response exactly as a planner would, and the pecuniary effects of their common error on
everyone else are zero-sum by Euler's theorem. We derived this symbolically before running any simulation, and we
present it first because a referee should see it first. The second result characterises when delegation \emph{does}
become a policy problem: when the aggregate action has non-pecuniary effects, when the price risk created by model
noise falls on agents who cannot insure it, or when prices enter collateral constraints
(Propositions~\ref{prop:wedges} and \ref{prop:collateral}). In each case we obtain the sign and the size of the wedge
between private and social value in closed form, the corresponding Pigouvian delegation tax, and the rate ($1/M$) at
which model diversity substitutes for it. In the benchmark, diversity is irrelevant; under the frictions, it is a
substitute for the tax.

The economy is built in three nested models so that each claim is made at the level of abstraction where it can be
proved or tested. Model A is a task-based stochastic growth economy \citep{acemoglu2018race} with a compute market in
which production automation and inference for delegated decisions compete for the same stock. Its within-period
decision problem reduces to a linear-quadratic game with dispersed information in the class of
\citet{angeletos2007efficient}, where delegation is the purchase of a correlated signal; this is where the
propositions live. Two findings come out of mapping the economy to the game. Delegating the consumption--savings
decision has almost no aggregate bite (a one-log-point productivity news moves desired savings by less than
$0.002$), because wealth and substitution effects cancel; delegating the allocation of wealth between risky and safe
assets has a large one. Which decision is delegated matters more than how good the model is. With illustrative
information parameters, the first household to delegate its allocation decision gains about 3\% of consumption per
year, the last gains 0.4\%, and at full adoption with one model the model-noise component of the wage has a standard
deviation of 7.7\% a year---larger than what calibrated productivity shocks generate.

Model B is a stationary heterogeneous-agent economy \citep{aiyagari1994uninsured} in which the decision channel
appears as a decision-quality return term and compute is owned broadly or by a small class. Three facts emerge. The
rich delegate first, because the stake is proportional to wealth and part of the fee is not. The information channel
alone raises the Gini coefficient from 0.68 to 0.70 and the top-10\% wealth share from 46\% to 47\% while adding
only 3\% to output: better decisions are a scale-dependent return, and scale-dependent returns concentrate wealth
\citep{gabaix2016dynamics}. AI as a production input works the other way under broad ownership---wages rise,
returns fall, the Gini falls to 0.58---and the other way again under concentrated ownership (Gini 0.66, top-10\% share 47\%). The
distributional consequence of AI is decided by who owns the algorithmic capital, not by the technology.

Model C replaces the auctioneer with explicit trading, balance sheets, collateralised credit and defaults, and
populates the market with four families of agents: a Bayesian mean-variance control, rules, a deterministic
surrogate behind the LLM interface, and live open-weight language models (Qwen2.5-7B, Mistral-7B, Llama-3.2-3B)
whose actions are schema-validated and clipped by deterministic code before they touch any ledger. In a factorial
design over market designs (Walrasian call, continuous double auction, bilateral bargaining), delegation share and
number of models with common random numbers, the order-flow correlation conditional on public information rises
from 0.11 with private signals to 0.97 with one shared model and falls to 0.43 with eight, in every design. The
competing explanation---homogenisation removes disagreement and therefore trade---is also present: volume collapses
toward the no-trade outcome and unconditional volatility does not rise. Fragility appears conditional on a shock. In
the call market, a false shared signal that moves the price by 3.7\% when agents are unlevered moves it by 9.8\%
when they start at 95\% of their margin capacity with a 20\% margin, an amplification of 2.6; a 50\% margin halves
it; four models eliminate it. In the continuous auction the same shock produces a 10\% trough and almost no cascade,
because sequential execution lets the price recover before the next margin check. Whether model noise is amplified
by collateral depends on how execution turns correlated orders into a marked price. Language-model agents given the same shared signal herd much less than Bayesian agents (their outputs add
idiosyncratic noise to a common signal), yet their trades load prices on the shared error more, they keep trading when
signals are identical, and half to three-quarters of their proposed actions hit the cash, holdings or margin limits.

\paragraph{Relation to the literature.} Common-noise fluctuations, the social value of public information and the
efficiency of the use of information in neoclassical economies are classical
\citep{lorenzoni2009theory,angeletos2013sentiments,morris2002social,angeletos2007efficient}; the endogenous purchase
of the same information has been studied \citep{veldkamp2006media,hellwig2009knowing,myatt2012endogenous}; the
portfolio externality of crowded strategies is known \citep{stein2009presidential,wagner2011systemic}; and the
closest recent work shows that AI adoption with correlated signals degrades price efficiency in a Kyle market
\citep{meng2026artificial} or states our premise verbally \citep{hadfield2025economy}. What is new is the delegation
market that sets the error-covariance structure of a production economy with closed-form $D$ and $1/M$ loadings, the
exact efficiency benchmark and its three failure modes with closed-form wedges, the compute market linking inference
to automation, and the factorial agentic-market experiments tied to the theory. Section~\ref{sec:literature} gives
the full audit, including what is not new.

\paragraph{What we do not claim.} The efficiency benchmark is a second-order result in a price-taking economy; the
information parameters are illustrative and swept rather than estimated; Model C outcomes are simulated
bounded-rationality outcomes, never general equilibria; LLM agents are not evidence about human preferences; and all
effects are identified within the structural model. Section~\ref{sec:limitations} spells each of these out.

The paper proceeds as follows. Section~\ref{sec:literature} audits the literature. Section~\ref{sec:environment}
lays out the environment and the equilibrium concepts. Section~\ref{sec:theory} contains the propositions.
Sections~\ref{sec:hamodel} and \ref{sec:markets} describe Models B and C; Sections~\ref{sec:computation},
\ref{sec:llm_design} and \ref{sec:calibration} describe computation, the LLM laboratory and the calibration.
Sections~\ref{sec:results_a}--\ref{sec:results_c} report results, Section~\ref{sec:robustness} robustness,
Section~\ref{sec:limitations} limitations, and Section~\ref{sec:conclusion} concludes. Appendices contain proofs,
algorithms, prompts and schemas, data provenance, replication instructions and additional results.