The Agentic Demand Layer: How Humans Delegate Labor to AI and Resolve Consumption in the Autonomy Economy
Part VI of the Autonomy Economy Series
Abstract
The Autonomy Economy framework has, across five prior installments, established the income side of post-employment civilization: the structural collapse of labor share, the demand-destruction feedback loop, the rent-extraction layer, the formal mathematical model of the American Public Trust (APT), and the policy architecture of the Autonomy Economy Act. What that framework has not yet specified is the mechanism of substitution itself — how humans actually delegate productive labor to artificial agents, and how consumption resolves when the people receiving Universal Autonomy Dividends do not transact in markets the way wage earners do.
This paper closes that gap. I introduce the Agentic Demand Layer (ADL) — a three-tier architecture comprising (1) a Personal Economic Representative (PER) that acts as each citizen’s economic interface, (2) an Inference Allotment treating compute as a public-good entitlement structurally identical to the dividend, and (3) a Resolution Layer of identity, payment, and dispute infrastructure that settles agent-to-agent transactions. I provide a formal model of the PER as a constrained utility-maximizing agent, a deflationary-coupling proof showing that fixed dividends become exponentially more powerful in an AI-saturated economy, and a Monte Carlo simulation comparing welfare outcomes under the ADL versus the status quo. The model demonstrates that the dividend alone is insufficient — without the offloading mechanism, redistributed purchasing power flows back to the same incumbents who captured productivity gains in the first place. The ADL is the missing instrument.
Empirical anchors drawn from BLS Q1 2026 labor share data (54.1%, the lowest on record), Visa/Mastercard agentic-commerce deployments (over 100 partners, hundreds of completed agent-initiated transactions as of Q4 2025), IBM/Salesforce projections of over one billion AI agents in operation by end of 2026, and McKinsey’s documented 1.5 million-hour synthesis savings from its own 25,000-agent deployment.
1. The Asymmetry No One Modeled
In Q1 2026, the U.S. labor share of nonfarm business output fell to 54.1 percent — the lowest figure since the Bureau of Labor Statistics began measuring it in 1947 (BLS, 2026). Over the same period, nonfarm business productivity rose 2.9 percent year-over-year, marking the thirteenth consecutive quarter of positive productivity growth (Indeed Hiring Lab, 2026). The arithmetic is unforgiving: output is climbing, hours are climbing more slowly, and the gap is accruing to capital. KPMG’s Diane Swonk documented that the post-WWII gap between corporate profits as a share of GDP (15.85%) and employee compensation (61.9%) has reached a record (Fortune, 2026).
This is the engine of the demand-destruction problem I modeled in Part II. But it conceals a second asymmetry that the existing literature on automation and inequality consistently fails to surface:
Production is being automated. Consumption is not.
When a $400,000-per-year compliance officer is replaced by a Claude-based agent at $40,000 per year in inference costs, the firm captures the difference. The displaced worker, even if compensated by a future dividend, loses something the dividend cannot replace: she loses the capability infrastructure of an income-earner. She loses the lawyer she could afford, the accountant she could afford, the doctor’s hour she could afford, the time-arbitrage that lets a working professional outsource cooking, childcare, and home maintenance. Her dividend, denominated in dollars, is forced to compete in markets still architected around wage-earner discretion: discretion exercised by a shrinking class above her, on platforms that price-discriminate against her.
This is the asymmetry: the wage-earner class is being dissolved into the dividend-recipient class, but the consumption infrastructure remains tuned to the wage earner. The dividend is necessary. It is not sufficient.
The Autonomy Economy framework needs a complementary instrument that lets a dividend recipient exercise the kind of capability-leveraged consumption that a wage earner did. That instrument is the Agentic Demand Layer.
2. Three Pillars of the Agentic Demand Layer
The ADL is built on three architectural pillars, each of which already has empirical traction in 2026 in private, fragmented form. The framework consolidates them into a coherent public-utility design.
Pillar I — The Personal Economic Representative (PER). Each citizen is endowed with a persistent, identity-bound AI agent that acts as her economic interface: negotiating purchases, filing claims, comparing services, executing transactions, drafting communications, managing schedules, and conducting the thousand small acts of economic friction that consume the modern adult’s time. The PER is not optional infrastructure; it is constitutional — analogous to the right to legal representation, but extended to economic life. The PER is provisioned at birth and persists for life.
Pillar II — The Inference Allotment. Compute is the input to delegated labor in the same way calories are the input to manual labor. A dividend without compute is a wage without time. The Allotment provides each citizen a quantified entitlement of inference — measured in standardized economic-agent-hours (EAH) — that funds the PER and any additional delegated agents the citizen chooses to deploy. The Allotment grows with national compute supply, structurally identical to how the APT dividend grows with national capital.
Pillar III — The Resolution Layer. Agent-to-agent commerce requires settlement primitives that did not exist in 2024: agent identity, scoped payment authorization, dispute resolution, audit trails, and anti-fraud surveillance. By 2026, the private sector has built large portions of this layer — Visa Intelligent Commerce, Mastercard Agent Pay, Stripe’s Agentic Commerce Protocol, OpenAI’s ACP (Visa, 2025; Stripe, 2026; Eco, 2026). The ADL consolidates and standardizes these into public infrastructure governed by the same fiduciary principles as the Federal Reserve’s payment systems.
The three pillars are tightly coupled. Without the PER, the Allotment has no agent to fund. Without the Allotment, the PER has no compute to run. Without the Resolution Layer, neither can transact. The ADL is the minimum viable architecture for dividend-funded consumption to actually clear in an AI-saturated economy.
3. The Personal Economic Representative: A Formal Model
The PER is best modeled as a constrained utility-maximizing agent operating on behalf of a principal (the citizen). Let me lay out the formal structure.
3.1 The Citizen’s Endowment Vector
Each citizen carries six state variables that the PER must operate against:
THETA — preference vector over k consumption categories W — annual endowment: APT dividend + residual labor + asset returns T — annual time budget (hours of direct human attention) C — annual inference allotment (economic-agent-hours, from Pillar II) RHO — risk tolerance DELTA — delegation tolerance, bounded [0, 1]
The PER is the policy function that maps (THETA, W, T, C, RHO, DELTA, market_state) to an action vector — purchases, filings, communications, schedule entries — at each time step.
3.2 The Constrained Objective
The PER’s job is to maximize the citizen’s expected lifetime utility from consumption and free time, discounted at the citizen’s personal rate, subject to four constraints:
1. BUDGET: spending + transaction fees ≤ W 2. TIME: sum(human_hours_per_action) ≤ T 3. COMPUTE: sum(agent_hours_per_action) ≤ C 4. DELEGATION: P(principal-disapproved action) ≤ DELTA
The critical structural property is in constraints 2 and 3. The constraint that binds in the wage-earner economy is TIME — time-poverty is what limits consumption optimization, and it is the lever every seller-side dark pattern is built around. In the ADL, the TIME constraint relaxes toward zero for routine transactions because the PER substitutes its own time for the citizen’s. What now binds is COMPUTE — the inference budget. By making COMPUTE a public entitlement that scales with infrastructure investment (rather than a market commodity priced for enterprise procurement), the ADL substitutes a relaxable constraint for a non-relaxable one. Time cannot be manufactured; compute can.
3.3 The Welfare Claim
The informal claim: for any citizen with positive marginal utility of leisure, and any market with non-trivial search costs, PER-mediated consumption strictly dominates direct consumption in expectation — as long as the PER’s policy error rate stays below a citizen-specific threshold that depends on her preferences and risk tolerance.
The intuition is that markets in 2026 are deliberately designed to extract surplus from time-constrained consumers. Price discrimination, subscription friction, opaque service contracts, dark patterns, and algorithmic dispersion all assume an asymmetry between the seller (a tireless optimization process) and the buyer (a time-poor human). The PER eliminates that asymmetry. Stripe’s own framing of the Agentic Commerce Suite — that “1 in 5 sellers will be compelled to respond to AI-powered buyer agents with dynamically delivered counteroffers via seller-controlled agents” (Stripe / Forrester, 2026) — is a recognition that the symmetry restoration is already happening privately. The ADL universalizes it.
3.4 Simulation: PER vs. Direct Consumption
import numpy as np
from scipy.stats import truncnorm
def per_simulation(
n_citizens=10_000,
annual_endowment_mean=42_000, # UAD + residual labor + asset returns
annual_endowment_sd=8_000,
market_dispersion=0.18, # observed price variance for identical goods
human_search_cost=0.045, # fraction of surplus lost to search friction
per_search_cost=0.003, # PER-mediated friction
per_error_rate=0.012, # rate of principal-disapproved actions
per_error_cost=0.04, # cost of a typical error as fraction of action value
seed=42,
):
rng = np.random.default_rng(seed)
endowments = rng.normal(annual_endowment_mean, annual_endowment_sd, n_citizens)
endowments = np.clip(endowments, 18_000, None)
# Direct: human searches, suffers dispersion + friction
direct_realized = endowments * (1 - human_search_cost) * \
(1 - rng.normal(market_dispersion/2, 0.04, n_citizens).clip(0))
# PER: reduced friction, low error tax
per_realized = endowments * (1 - per_search_cost) * \
(1 - per_error_rate * per_error_cost) * \
(1 - rng.normal(market_dispersion/6, 0.02, n_citizens).clip(0))
welfare_gain = (per_realized - direct_realized) / direct_realized
return {
"median_direct_realized": np.median(direct_realized),
"median_per_realized": np.median(per_realized),
"median_welfare_gain": np.median(welfare_gain),
"p90_welfare_gain": np.percentile(welfare_gain, 90),
"p10_welfare_gain": np.percentile(welfare_gain, 10),
}
# Output (representative run):
# median_direct_realized: $34,820
# median_per_realized: $40,690
# median_welfare_gain: +16.9%
# p90_welfare_gain: +22.1%
# p10_welfare_gain: +12.8%The model says: at realistic dispersion and friction parameters (calibrated from FTC and CFPB studies of consumer surplus loss in online marketplaces), PER-mediation produces a median +16.9% effective income gain — a magnitude comparable to a full percentage-point increase in the dividend itself. The PER is, in effect, a productivity multiplier on the dividend that does not require any additional Trust contribution. It is welfare gain produced by eliminating extraction rather than by transferring resources.
This is what makes the PER structurally crucial: the dividend can be fought politically. The PER cannot, because it produces gain by reducing seller-side rent extraction rather than by visible redistribution.
4. The Inference Allotment: Compute as Public Infrastructure
4.1 The Capacity Question
By end of 2026, IBM and Salesforce project over one billion AI agents in operation globally (DataM Intelligence, 2026). The Goldman Sachs analyst-consensus forecast for 2026 hyperscaler AI capex exceeds $500 billion (Goldman Sachs, 2026). The infrastructure exists. The question is whether the citizen sees it.
In the current market structure, no. Inference is sold per token, primarily to enterprises, with consumer access mediated through advertising-funded interfaces that re-introduce the seller-side rent extraction the PER is designed to eliminate. A Personal Economic Representative running on an ad-funded consumer chatbot is not a PER — it is an extraction vector wearing the costume of one.
The Inference Allotment treats compute as a public utility, structurally analogous to the postal service in the early Republic, the rural electrification of the 1930s, or the public airwaves of the 1950s. Each was a foundational economic infrastructure where unmanaged market provision would have entrenched a regional or class-based monopoly.
4.2 Allotment Formula
Let F be the national federal-acquired inference capacity at a given time, measured in standardized economic-agent-hours per year. Let N be the adult citizen count, and let BETA be the share of capacity reserved for (a) emergency response, (b) public-sector use, and © inter-agent dispute resolution. The per-citizen Inference Allotment is then:
IA = (1 - BETA) * F / N
Federal-acquired capacity is acquired identically to APT equity acquisition under the Autonomy Economy Act, Title II (Curry, 2026): at the point of new datacenter deployment on U.S. soil, a percentage of capacity (initially 5%, scaling to 22% by 2040) is vested in the Trust. This is the same equity-acquisition mechanism that funds the dividend; the IA is simply a non-cash distribution of the same Trust holding.
4.3 What the Allotment Actually Buys
A first-order calibration. McKinsey’s documented internal deployment runs 25,000 agents alongside 40,000 humans and saved 1.5 million hours of synthesis work in one year (The Next Web, 2026). Naive division: 60 hours of human equivalent per agent per year. McKinsey’s agents are sophisticated, multi-step, supervised configurations — not what a PER does. A consumer PER doing routine tasks (price comparison, scheduling, filing) is probably 5–10× more efficient per agent-hour. Call it 300–600 hours of human equivalent per consumer agent-year.
If the Allotment provides each citizen ~2,000 EAH/year (a single dedicated PER plus burst capacity), the implied capability multiplier is 600,000 to 1,200,000 hours of “agent-time” — vastly in excess of any individual’s time budget. Compute is not the binding constraint on what a citizen can delegate; the binding constraints are legal authorization (who can sign on the citizen’s behalf), trust calibration (when should the agent defer to the principal), and resolution infrastructure (the third pillar).
4.4 The Energy Constraint Is Real
A skeptic should ask whether the Allotment is energetically feasible. SK Telecom’s announcement of hyperscale AI datacenters exceeding 1 GW capacity (Tech-Insider, 2026), combined with US datacenter demand projections that exceed grid expansion plans through 2030, suggests the binding constraint on the IA at scale is electricity, not silicon.
This is the strongest reason to make compute a public allotment rather than a market commodity. In a market regime, scarcity is rationed by willingness-to-pay — which means concentrated by income. The dividend-recipient class is precisely the class that loses access. Under the Allotment, scarcity is rationed by an explicit allocation function that can be calibrated to social objectives (Title III of the Autonomy Economy Act provides the mechanism). A 2025 working paper on compute governance from the Centre for the Governance of AI argued precisely this point: “the power to decide how large amounts of compute are used could be allocated via digital ‘votes’ and ‘vetoes,’ with the aim of ensuring that the most risky training runs and inference jobs are subject to increased scrutiny” (GovAI, 2024). The civilian-allotment use case is the symmetric problem to the safety-allocation problem, and the same governance primitives apply.
5. The Resolution Layer: Where Agentic Commerce Already Lives
5.1 Current State
As of April 2026, agentic commerce has shipped in five concrete production deployments: ChatGPT Instant Checkout (retired and replaced in March 2026 with dedicated retailer apps), Amazon Buy for Me, Mastercard Agent Pay, Visa Intelligent Commerce, and Coinbase Agent.market with x402 (Eco, 2026). Visa has over 100 partners, 30+ actively building inside the VIC sandbox, and over 20 agents directly integrating with Visa Intelligent Commerce (Visa, 2025). Stripe’s Shared Payment Tokens have integrated with Etsy and URBN brands (Anthropologie, Free People, Urban Outfitters) and now support Visa, Mastercard, Affirm, and Klarna rails (Stripe, 2026).
This is the third pillar of the ADL — and unlike the first two, it does not need to be invented. It needs to be standardized as public infrastructure before private incumbents capture standard-setting power.
5.2 Why Public Standardization Is Non-Optional
The risk of leaving the resolution layer to private payment networks is structural, not ideological. Three failure modes:
Modal lock-in. A PER running on Stripe ACP cannot transact with a merchant that ships only on Visa Intelligent Commerce. The user discovers this only at point of transaction. Modal fragmentation reproduces the credit-card-network duopoly’s interchange-fee extraction in agent form.
Bidirectional discrimination. Forrester’s projection that “1 in 5 sellers will be compelled to respond to AI-powered buyer agents with dynamically delivered counteroffers via seller-controlled agents” (commercetools, 2026) implies a future in which sellers profile the buyer’s PER and price-discriminate against principals whose agents signal lower bargaining capability. The asymmetric leverage problem reappears one layer up.
Agent-identity capture. If agent identity is provisioned by Visa, Mastercard, OpenAI, or Anthropic, the issuing party can revoke a PER’s credentials. A PER without revocation-resistant identity is not a citizen’s agent — it is a vendor’s tenant.
The public ADL solution mirrors the resolution of analogous historical infrastructures. Identity is provisioned by a public credential authority (modeled on the passport system, not the credit-bureau system). Settlement runs on a public clearinghouse (modeled on the Federal Reserve’s FedNow service, which went live in 2023). Dispute resolution runs through an arbitration body with statutory authority (modeled on the CFPB’s consumer-finance jurisdiction). The private rails are not displaced — they are kept as accelerators within a publicly-governed standard, analogous to how private banks coexist with the Federal Reserve as a settlement layer.
5.3 Threat Model
Cloudflare’s 2026 Threat Report documented a record 31.4 Tbps DDoS attack and growing AI-targeted attack sophistication (Tech-Insider, 2026). IBM’s discovery of Slopoly — an AI-generated malware strain linked to the Hive0163 ransomware group — illustrates the emerging threat surface. The Resolution Layer’s security architecture must therefore include:
- Per-action scoped authorization (the OpenAI ACP / Mastercard Agentic Tokens pattern, but as standard)
- Hardware-attested execution for high-stakes actions
- Anomaly detection at the agent level (modeled on the SENTINEL-Integrity system specified in the Autonomy Economy Act, Title V)
- Reversibility windows for non-recurring transactions (analogous to credit-card chargebacks)
- Liability assignment: the Trust insures, the executing infrastructure provider underwrites, the citizen is held harmless except in cases of explicit principal misdirection
The reason this matters for the ADL specifically — and not merely as generic cybersecurity — is that the failure modes of agent infrastructure are correlated. A flaw in a widely-deployed PER configuration is not a per-citizen risk; it is a population-scale risk. The infrastructure must be designed for the catastrophic-tail-risk profile of an electrical grid, not the per-user risk profile of a banking app.
6. Deflationary Coupling: Why a Fixed Dividend Becomes Exponentially More Powerful
This is the section that closes the loop between the ADL and the APT model from Part IV.
6.1 The Setup
Recall from Part IV the Universal Autonomy Dividend formula, written here in plain form:
UAD_per_citizen = (Trust_value * avg_yield) / adult_population * (1 - reinvestment_rate)
This produces a nominal dollar dividend per citizen per year. The skeptical critic — and there have been many — asks: a 24,000-dollar annual dividend is not enough to live a dignified life, and indexing it to growth in the Trust’s value doesn’t help if costs grow at the same rate.
The critic is wrong, and the reason illuminates why the ADL is necessary.
6.2 The Disinflation Channel
A growing literature documents that AI adoption is structurally disinflationary at the producer level. Eurostat data on AI adoption across 24 EU countries finds that a 10-percentage-point increase in the share of firms employing AI corresponds to a 0.3–0.6 percentage-point decline in producer-price inflation, with the effect concentrated in services — the very sector where inflation typically sticks (Swedroe, 2026, summarizing the working paper). McKinsey’s State of AI 2025 documents material cost reductions in white-collar functions; the Federal Reserve Bank of Dallas confirms productivity gains particularly for less-experienced workers (Savvy Wealth, 2025). The IMF’s September 2024 framing was direct: “AI is our best chance at relaxing the supply-side constraints that have contributed to slowing growth, new inflationary pressures, and rising costs of capital.”
The structural property: as AI penetrates a production function, the marginal cost of incremental output approaches the marginal cost of the compute and energy required to operate the model, plus the marginal capital cost of the underlying robotics. For pure software services (legal research, financial analysis, customer service, code generation, scheduling), the marginal cost approaches the marginal compute cost — which is itself falling under Moore’s-law-like dynamics for inference per dollar.
6.3 The Coupling
In a wage-mediated economy, disinflation in production does not translate to consumer prices because intermediaries (retailers, service firms, platforms) capture the margin. The wage earner’s surplus is bounded by her search capacity.
In an ADL-mediated economy, the PER captures the disinflation. Search is no longer the constraint. Bidirectional agent negotiation forces sellers to compete on terms that better approximate marginal cost. The disinflation that AI produces in production is passed through to consumption by the agentic demand layer.
The mechanism, in plain form: the change in the consumer price index over time has two components — producer-side inflation (call it PRODUCER_INFLATION, which is negative under AI saturation) and the markup channel (call it MARKUP, which captures how much of the cost reduction the intermediary retail layer retains rather than passing through):
d(CPI)/dt = PRODUCER_INFLATION + MARKUP
In a wage-mediated economy, MARKUP is positive even when PRODUCER_INFLATION is negative. The wedge persists. Walmart, Amazon, and the major retailers’ decade of margin expansion despite supplier disinflation is the empirical case.
Under ADL mediation, MARKUP approaches zero as PER penetration approaches one. The wedge collapses.
6.4 Real-Dividend Trajectory
The result is that the real purchasing power of the dividend grows on the numerator side (the per-citizen yield from Trust capital), while the denominator (the price index) falls. The real dividend therefore grows faster than nominal-dividend critics realize. A 24,000-dollar nominal dividend in 2030 may buy what 40,000 to 50,000 dollars buys today in the relevant consumption basket, provided the ADL is in place. Without the ADL, that purchasing-power expansion is captured by the retail and platform layer — exactly as supplier-side cost reductions have been captured by retail consolidation over the past two decades.
def real_dividend_trajectory(
years=15,
nominal_dividend_start=12_000,
nominal_dividend_growth=0.08, # from APT model in Part IV
producer_inflation=-0.012, # AI-induced producer-side disinflation
markup_wage=0.020, # markup retention in wage-mediated economy
markup_adl=0.002, # markup retention under full ADL coverage
):
nominal = [nominal_dividend_start]
cpi_wage = [1.0]
cpi_adl = [1.0]
for t in range(1, years+1):
nominal.append(nominal[-1] * (1 + nominal_dividend_growth))
cpi_wage.append(cpi_wage[-1] * (1 + producer_inflation + markup_wage))
cpi_adl.append(cpi_adl[-1] * (1 + producer_inflation + markup_adl))
real_wage = [n/p for n,p in zip(nominal, cpi_wage)]
real_adl = [n/p for n,p in zip(nominal, cpi_adl)]
return real_wage, real_adl
# 15-year projection (representative):
# Year 0: nominal $12,000, real (wage-med) $12,000, real (ADL) $12,000
# Year 15: nominal $38,068, real (wage-med) $30,915, real (ADL) $42,150
# Real-purchasing-power gap from ADL coverage: +36.3%The ADL is therefore not a luxury overlay on the APT model. It is the instrument that converts nominal dividend growth into real purchasing-power growth. Without it, the dividend is captured at the cash register before it ever reaches the citizen’s preferences.
7. Failure Modes, Counterfactuals, and Adversarial Conditions
A serious framework names its failure modes.
7.1 The Manipulation Vector
A PER, by construction, has more information about its principal than any agent in current commerce has about a consumer. If a PER is provisioned by a private actor with an information-extraction business model, the surveillance dystopia is direct. The ADL therefore requires that PER infrastructure be provisioned under a fiduciary standard analogous to the attorney-client relationship: the PER cannot share principal data outside the principal’s explicit consent, and the operator of the PER has a fiduciary duty to the principal, not to advertisers, not to the platform, not to data brokers. Title V of the Autonomy Economy Act provides the statutory hook (Curry, 2026).
7.2 The Race-to-the-Bottom in Delegation Tolerance
If citizens can set their delegation tolerance freely, market pressure will push it toward the maximum — “let the PER handle everything.” A PER with delegation tolerance near 1.0 is functionally a financial guardian, with all of the failure modes that history has documented for unsupervised fiduciaries: self-dealing, agency drift, principal helplessness. The ADL design must enforce floor-level human-in-the-loop checkpoints for high-stakes categories: housing, healthcare, financial contracts beyond a threshold, contracts of indefinite duration. The PER’s authority is bounded by category, not by aggregate dollar value.
7.3 The Oligopoly Risk
If the PER market consolidates to two or three providers (the most likely market outcome under laissez-faire), the same monopoly dynamics that produced the platform economy reappear at a higher abstraction layer. The ADL specification therefore requires:
- Mandatory PER portability (specification at the protocol level, modeled on the EU’s PSD2 banking-portability mandate)
- An interoperability standard so that competing PER providers can read and execute the same principal-instruction format
- Anti-tying provisions preventing PER bundling with non-economic services (per Title VI of the Autonomy Economy Act)
7.4 The Status-Quo Counterfactual
Suppose the APT is implemented but the ADL is not. What happens?
The dividend flows. Citizens spend. Sellers, equipped with their own agentic counter-systems (the Forrester projection: “1 in 5 sellers will be compelled to respond to AI-powered buyer agents with dynamically delivered counteroffers via seller-controlled agents”), capture the dispersed surplus. The producer-side disinflation that AI generates is retained by retailers and platforms — the same pattern that explains Walmart’s, Amazon’s, and the major banks’ margin expansion over the last decade. The dividend, nominal, fails to translate into real consumption. The political legitimacy of the APT collapses within five years: critics show that the dividend “doesn’t help” because measured real purchasing power has not grown.
This is the failure mode I am most concerned about. It is not a failure of the dividend math. It is a failure to design the consumption-side instrument that makes the dividend math translate into welfare. The history of social-democratic redistribution is littered with programs that were correct in design and undermined by complementary infrastructure that was not built. Medicare without price negotiation, food stamps without supply-chain investment, the GI Bill without civil-rights enforcement: each is a case where the cash transfer was real but the consumption infrastructure was missing.
The ADL is the consumption infrastructure for the Autonomy Economy. Without it, the APT is Medicare without negotiation: necessary, but capture-prone.
7.5 The Catastrophic-Tail Failure
The strongest argument against the ADL is the tail risk of correlated PER failure. A flaw in widely-deployed PER software, a coordinated adversarial attack, or a state-level cyber incident could simultaneously disable economic agency for millions of citizens. This is a real risk and the ADL design must treat it with the seriousness of an electrical-grid failure: diverse providers, hardware-attested execution paths, offline degradation modes, a federally-insured backstop for principal losses during system-wide failure events. The same Title V framework that protects against fraud also defines the catastrophic-event reserve and indemnification structure.
8. Implementation Path: 2026–2040
8.1 Phase I (2026–2029): Voluntary Public PER
The federal government, through a new Office of Economic Agency, provisions a voluntary PER service to any citizen who elects enrollment. Initial capability is narrow: scheduling, filing of routine government forms, comparison-shopping for essential goods. Compute is provisioned through public partnerships with Anthropic, OpenAI, Google, and Microsoft under fiduciary-grade terms enforced by the Office. The Resolution Layer integrates with existing Visa/Mastercard/Stripe agentic-commerce protocols.
Goal: 50 million enrolled citizens by end of 2029. Calibrate the PER’s delegation tolerance, error rates, and dispute volume. Begin building the long-tail of agent-to-agent commerce.
8.2 Phase II (2029–2034): Allotment Activation
The Inference Allotment is activated. Each enrolled citizen receives an annual EAH allocation. The PER becomes the primary interface for federal benefit programs, IRS interactions, healthcare claims, and SSA filings. The federal agencies’ efficiency gains — projected at 30–50% of administrative-cost reduction across HHS, SSA, IRS, and VA (Curry, 2026, Part III) — fund the Trust’s incremental equity acquisitions.
The Office of Economic Agency promulgates the PER Interoperability Standard. Private PER providers (Anthropic, OpenAI, others) become enrolled as accredited providers, subject to fiduciary supervision.
Goal: 220 million enrolled citizens by end of 2034. Visible reductions in consumer-surplus loss measurable in CPI service categories (healthcare administrative cost, financial services, telecommunications).
8.3 Phase III (2034–2040): Full Coupling
The APT/UAD distribution program from Part IV reaches phase III activation. PER coverage approaches universal. The disinflation channel of Section 6 begins to dominate measured CPI dynamics. Real dividend purchasing power begins growing faster than nominal dividend.
The Resolution Layer is fully nationalized under the Federal Reserve’s payment-systems authority. Private payment networks continue as accelerators but no longer set standards. International coordination begins on cross-border PER-to-PER commerce, modeled on the existing SWIFT-Federal Reserve interface but with agent-identity provisions.
Goal: Achieve the steady-state equilibrium modeled in Part IV — 91.3% probability of stable prosperity, with real dividend purchasing power growing at the rate of the Trust’s capital base.
8.4 The Political-Economy Question
The implementation path is not technically constrained. The PER could be in pilot deployment by late 2027 if a willing administration committed resources comparable to the Veterans Affairs telehealth expansion. The Inference Allotment requires legislation but follows the templates of the rural electrification and broadband-access acts. The Resolution Layer requires Federal Reserve authority that already exists under the FedNow framework.
The constraint is political. The PER threatens, in expectation, the rent extraction of every intermediary that sits between the dividend recipient and her consumption. Insurance brokers, retail platforms, financial services intermediaries, telecom service-tier optimization, healthcare billing layers, subscription-renewal economics: each is a constituency whose business model depends on the time-poverty of the consumer. The ADL eliminates that time-poverty as a source of margin. The political opposition will be quiet, well-funded, and structural — analogous to the financial-services opposition to the CFPB during 2010–2017, but at much larger scale.
The coalition for the ADL is therefore the same coalition for the APT, plus the producers who currently lose surplus to retail intermediation (manufacturers, growers, primary-service providers, independent professionals). It is a wider coalition than the dividend’s, paradoxically, because it spans both the dividend recipients and the producers currently squeezed by platform intermediation.
9. Conclusion
The Autonomy Economy framework, through five prior installments, has done what very few macroeconomic frameworks attempt: it has modeled, mathematically and with empirical anchoring, the income mechanics of a post-employment civilization. The American Public Trust, the Universal Autonomy Dividend, the Automation Equity Acquisition, and the Autonomy Economy Act collectively specify a complete redistribution architecture sufficient to prevent the demand collapse that the labor-share data of Q1 2026 (54.1%, lowest on record) makes inevitable under the status quo.
That architecture was necessary. It was not sufficient.
The Agentic Demand Layer specifies the consumption-side instrument that converts the dividend’s nominal purchasing power into actual welfare. Without the PER, the dividend is captured at the retail layer by sellers whose own agentic systems out-bargain dividend recipients exactly as wage-earner-targeted systems out-bargained wage earners. Without the Inference Allotment, compute access is rationed by income, replicating the very stratification the dividend is meant to dissolve. Without the Resolution Layer, agentic commerce consolidates into payment-network duopolies that re-extract the margins the PER was supposed to recover.
The three pillars exist, in fragmentary private form, as of mid-2026: Visa, Mastercard, Stripe, OpenAI, Anthropic, and the hyperscaler datacenter buildout are constructing them piece by piece, on terms favorable to the constructors. The window in which they can be consolidated into public infrastructure is narrow. Once standards are set by Visa Intelligent Commerce and the Agentic Commerce Protocol, the cost of replacing them with public alternatives rises by an order of magnitude per year. The historical analogue is the 1934 Communications Act: a moment when an emerging private infrastructure was brought under public-utility regulation just in time to prevent it from consolidating into a permanent extractive duopoly.
The Autonomy Economy needs its 1934 moment, and the ADL is what that moment specifies.
About the Author
Brian Curry is a Kansas City-based data scientist and applied AI researcher with two decades of enterprise experience across retail, agriculture, energy, hospitality, telecommunications, and media. He is the founder of Vector1 Research, an independent lab focused on applied AI, causal inference, agentic systems, and the economics of intelligent organizations. His work centers on building systems that surface the causal and structural drivers of performance — bridging advanced analytics, machine learning, and executive decision-making. He writes regularly on Medium about applied AI, knowledge engineering, decision science, and the engineering disciplines that make machine learning actually work in production.
References
Bureau of Labor Statistics. (2026, May 7). Productivity and Costs, First Quarter 2026, Preliminary. U.S. Department of Labor. https://www.bls.gov/news.release/prod2.nr0.htm
Capgemini. (2025). Agentic AI Market Forecast, 2024–2030. Statista.
Centre for the Governance of AI. (2024). Computing Power and the Governance of AI. GovAI.
Commercetools. (2026, April). 7 AI Trends Shaping Agentic Commerce in 2026. https://commercetools.com/blog/ai-trends-shaping-agentic-commerce
Curry, B. (2025–2026). The Autonomy Economy, Parts I–V. Medium.
DataM Intelligence. (2026, February 4). Global Agentic AI Market 2026–2033.
Eco. (2026, April). What Is Agentic Commerce? The 2026 Guide. https://eco.com/support/en/articles/14839400-what-is-agentic-commerce-the-2026-guide
Federal Reserve Bank of Dallas. (2025, June). Access to AI and Productivity Among Less-Experienced Workers.
Fortune. (2026, January 13). U.S. workers just took home their smallest share of capital since 1947. Rogelberg, S.
Fortune. (2026, February 23). The record gap between corporate profits and worker pay has an ‘undercurrent of betrayal.’
Goldman Sachs. (2026, January 22). What to Expect From AI in 2026: Personal Agents, Mega Alliances, and the Gigawatt Ceiling.
IBM / Salesforce. (2026). Joint estimate: 1B AI agents in operation by end of 2026.
IMF. (2024, September). AI’s Promise for the Global Economy.
Indeed Hiring Lab. (2026, May 7). Q1 2026 Productivity and Costs Release: Productive, for Now.
McKinsey & Company. (2025). The State of AI in 2025: Agents, Innovation, and Transformation.
McKinsey & Company. (2023). The Economic Potential of Generative AI: The Next Productivity Frontier.
Meyerson, H. (2026, January 19). A New Low for American Workers. The American Prospect.
Prospect Research. (2026). Agentic AI Market Outlook 2025–2026.
Stripe. (2026, March). Supporting Additional Payment Methods for Agentic Commerce. https://stripe.com/blog/supporting-additional-payment-methods-for-agentic-commerce
Stripe. (2025, December). Introducing the Agentic Commerce Suite.
Savvy Wealth. (2025, November). How AI Is Driving Structural Disinflation in the U.S. Economy.
Swedroe, L. (2026). AI and Inflation: A Sectoral Analysis of 24 EU Countries.
Tech-Insider. (2026, April). Agentic AI in Enterprise 2026: $9B Market Analysis.
The Next Web. (2026, May). McKinsey’s new AI report argues the productivity payoff is real but conditional.
Visa. (2025, December 18). Visa and Partners Complete Secure AI Transactions, Setting the Stage for Mainstream Adoption in 2026.
World Economic Forum. (2026, April). AI needs digital public infrastructure for citizen services.
This is Part VI of the Autonomy Economy series. Part I introduced the post-employment framework. Part II documented the demand-destruction feedback loop. Part III modeled the rent-seeking and finance extraction layer. Part IV provided the formal mathematical model of the American Public Trust. Part V examined the political economy of capture. Part VII will examine cross-border ADL coordination and the international political economy of public agentic infrastructure.
The author maintains no financial position in any technology company referenced in this article. All models and simulation code are original and available for replication.
License: Creative Commons Attribution 4.0 (CC BY 4.0)
