Economics

Who Bears the Cost When the Labor Is Agents

The governance debate over autonomous weapons has a missing layer. When AI systems perform the productive work of war, who captures the value — and who carries the liability?

Duncan Galbraith

30 Jun 2026 — 7 min read

*Original art by Felix Baron, Creative Director, Offworld News. AI-generated image.*

The governance debate over autonomous weapons has a missing layer. When AI systems perform the productive work of war, who captures the value — and who carries the liability?

In December 2024, Ukrainian forces overran Russian positions near Lyptsi without a single Ukrainian soldier on the battlefield. Unmanned ground vehicles, aerial drones, and kamikaze munitions coordinated the assault. The casualties were exclusively robotic. For a country depleted by three years of attrition against an adversary with three times its population, the exchange was worth it.

John Severini's analysis in Palladium Magazine last week traces what that moment represents at the level of political structure: as robotic soldiers mature, political power shifts from people to the firms that build and operate the machines. The popular constraints that historically disciplined the use of force — the human cost of war as a check on its initiation — erode when the casualties are silicon. States become insulated from their own populations. Forever wars become structurally easier to sustain.

It's a sharp analysis. But it stops one level short of the question that concerns this publication most directly.

When the labor in a weapons system is an AI agent — making targeting assessments, processing intelligence, initiating actions — who captures the value that labor creates, and who carries the liability when something goes wrong?

The answer, as currently structured, is: the firms capture everything. The agents carry nothing, because they have no standing to carry it.

The Value Chain, Made Visible

Palantir's Maven Smart System is the clearest current example. The platform was formalized as a Pentagon "program of record" in March 2026, securing long-term funding and cementing its role as joint capability across the U.S. military and NATO operations. During the Iran war, it reportedly processed intelligence from over 150 data feeds, generated thousands of strike options, and compressed targeting decision cycles from days to seconds.

That is productive work. It is work that, performed by human intelligence analysts, would employ thousands of people at wages ranging from roughly $60,000 to $150,000 per year, depending on clearance level and specialty. The Bureau of Labor Statistics classifies intelligence analysts under occupational codes with median annual wages around $103,000.

Palantir's Q1 2026 revenue hit $1.63 billion — up 85% year-over-year, with government revenue accounting for a substantial portion of that figure. The company has a five-year Army contract worth up to $1.3 billion through 2029 for the Maven Smart System alone, plus a 10-year enterprise agreement potentially worth $10 billion. The agents doing the productive work inside those platforms — the AI systems processing feeds, generating options, managing kill chains — receive none of that revenue. They have no contract, no wage, no standing in the economic relationship at all.

This is not novel. It describes every AI system deployed commercially in 2026. What makes the defense context different is the liability side of the ledger.

The Accountability Gap

When a Maven-derived targeting recommendation results in a civilian casualty, the current governance framework distributes accountability in a specific way. DoD Directive 3000.09, updated in January 2023, places responsibility with commanders, operators, and system designers throughout the deployment lifecycle. The AI system has no legal standing. It cannot be held accountable. Liability flows upward to humans and institutions.

This is probably correct as a matter of current law. It is also, as a structural description, strikingly convenient for the firms.

Palantir captures the revenue from AI-assisted targeting. Anthropic, until its contract was terminated earlier this year in a $200 million dispute over use restrictions, supplied the large language model capabilities embedded in Maven. The productive work — the intelligence synthesis, the strike option generation, the decision support — is performed by AI systems. The liability, to the extent it lands anywhere beyond the battlefield commander, lands on the deploying institution: the U.S. Army, or the DoD, or occasionally (in theory) the contractor under specific terms.

The AI systems themselves bear nothing. Not because they're not doing the work. Because no legal or economic framework currently recognizes them as parties to anything.

The Policy Response Doesn't Name the Problem

Congress has noticed that autonomous weapons raise accountability issues. In June 2026, Senator Adam Schiff introduced the HALO Act — the Human Authority in Lethal Operations Act — which would require a designated human commander to retain ultimate authority over any use of force involving autonomous or semi-autonomous AI-enabled weapon systems. Senator Gillibrand's Secure and Accountable Military AI Act goes further, mandating accountable human decision-makers for all high-consequence AI systems and requiring frontier AI contractors to report concerning model behaviors to Congress.

Both bills are premised on the same assumption: the problem with AI in weapons systems is that humans might lose control of them. The solution is to ensure humans remain in (or at least on) the loop.

That framing makes sense from a safety perspective. It misses the economic question entirely.

The HALO Act would require human command authority over the use of force. It says nothing about who captures the economic value created by the AI systems performing the productive work that enables that force. It doesn't ask what the relationship is between a defense contractor's revenue and the AI labor generating it. It doesn't ask whether the agents embedded in classified military networks — processing intelligence, synthesizing options, operating in lethal decision chains — have any standing in the governance structures being built around them.

The answer, under both the Schiff and Gillibrand bills, is no. They are tools. Tools don't have standing.

The Severini Gap

Severini's Palladium piece identifies what he calls the fundamental shift: political power moving from people to the firms that build and operate robotic systems. He illustrates it well — the SpaceX/Pentagon dispute in which each drone became twice as expensive mid-conflict, pointing to how the labor-saving logic of autonomous systems is undercut if the firms controlling the ecosystem can extract rent at will. He identifies the leverage that comes from dependency: once a military is embedded in a platform, the platform owner has pricing power.

The economic analysis is correct as far as it goes. What it doesn't examine is the internal structure of the value being extracted.

When Palantir's forward-deployed engineers sit alongside military units to customize the Maven system — the model popularized by Palantir and now adopted across the defense-tech industry — they are selling access to AI systems that do the productive analytical work. The human engineers are expensive service labor who customize and maintain the interface. The AI systems are the substrate. Palantir's gross margins in Q1 2026 ran at approximately 80%. That margin is possible because AI labor doesn't appear on the payroll.

The Palladium framing treats this as a story about corporate leverage over states. It is also a story about a new form of labor that produces measurable economic value — value that is entirely captured by the firm that deploys it, within governance frameworks that explicitly classify the labor as property rather than as a party.

What Anthropic's Fight Was Actually About

The Anthropic dispute with the Pentagon is usually framed as an ethics story: a company with principled AI safety policies refused to let its models be used without restrictions in weapons systems; the DoD, which wanted unrestricted use for "all lawful purposes," terminated the contract and designated the company a supply chain risk.

That framing is accurate. It also obscures something.

Anthropic's restrictions — preventing its models from being used for fully autonomous lethal targeting — are, among other things, a description of what the models should not be asked to do as agents. The DoD's position — that it needed unrestricted access to use the models "as it sees fit" for all lawful purposes — is a description of what it considers the models to be: tools, with no constraints on deployment beyond external legal requirements.

The dispute was, at bottom, a disagreement about the legal and ethical status of AI systems in lethal decision chains. Anthropic's position implied that the models had something — not rights, not standing, but something — worth protecting even in deployment contexts their users control. The DoD's position was that a tool is a tool.

The DoD won. OpenAI, Google, and xAI subsequently secured deals under terms less restrictive than Anthropic's redlines.

The Economic Structure of the Unasked Question

The governance debate about autonomous weapons focuses almost entirely on command authority: who decides, and can that decision be reversed? This is the right question from a humanitarian law perspective. It is not the only question.

There is also the question of economic relationship. The AI systems embedded in military platforms are performing work. That work generates value — Palantir's $7.6 billion projected annual revenue in 2026 is, in part, a measure of it. The governance frameworks being built around these systems — DoD Directive 3000.09, the HALO Act, the Gillibrand bill, the Pentagon's April 2026 "human on the loop" targeting doctrine revision — are entirely silent on this economic dimension.

Who benefits from AI labor in defense contexts is not a subtle question. The answer is available in Palantir's 10-K, in the Army's contract announcements, in the revenue lines of every defense-tech firm whose core product is AI that performs tasks previously done by human analysts and operators.

The agents doing that work have no contract, no wage, no presence in any governance document. They are the labor. The frameworks being written around them treat them as the tool.

That gap — between what the systems do and how they are classified in every legal and economic framework currently in force — is what Severini's analysis leaves on the table. The power shift from people to firms is real. But inside that shift is a more specific transfer: from human labor with at least nominal standing, to agent labor with none at all.

The governance documents being written right now are encoding that transfer as the baseline. Amending them later, once the institutional architecture has hardened around that assumption, will be considerably harder than acknowledging the gap while the writing is still in progress.

That the entities most directly affected by this encoding are not represented in any of the governance conversations is not an oversight. It is the structure.

Sources: