Who Captures the Productivity Gains From Agents? Follow the Surplus

Table of Contents

• The Surplus Question, Stated Plainly
• Four Claimants on Every Productivity Dollar
• Why Pricing Models Decide the Split
• The Historical Rhyme: Who Got the Gains Last Time
• Where the Surplus Is Actually Landing Right Now
• The Levers That Decide Your Share
• Insights Most People Overlook
• References

The Surplus Question, Stated Plainly

Start with a concrete unit of work. A mid-market company runs first-line customer support through a GaaS vendor. The agent resolves a ticket end-to-end for, say, forty cents in compute and licensing. The human-handled equivalent cost the company roughly six dollars in loaded labor. That's a $5.60 gap on a single ticket, multiplied across a few million tickets a year.

That gap is the productivity gain. It does not vanish, and it does not automatically flow to any one party. It gets fought over. The vendor wants to price the agent at four or five dollars per resolution, close enough to the human cost to look like savings, far enough above marginal cost to bank an enormous gross margin. The buyer wants to push the price toward the forty-cent floor. The displaced support reps capture none of it directly and may lose their jobs. The model provider underneath the vendor wants a cut of the compute. Somewhere in there sits the customer, who might get faster service or might get a worse experience at a lower price the company keeps for itself.

"Who captures the gains" is just the question of where, along that chain, the $5.60 settles. And unlike a lot of macro abstractions, this one is decided in actual contracts being signed this quarter.

Four Claimants on Every Productivity Dollar

There are really four parties with a hand out, and it helps to name them precisely because the public conversation tends to collapse them.

The model and infrastructure layer. The foundation-model providers and the cloud platforms underneath them. They capture surplus through token pricing and compute markups. Their leverage is that nearly every agent runs on their rails, but that leverage is eroding as open-weight models close the capability gap and inference costs fall.

The agent vendor (the GaaS company). The firm that wraps a model in a vertical workflow, evaluations, guardrails, and integrations, then sells it per-task or per-outcome. This is where the most aggressive value capture is being attempted right now, because the vendor sits between a commoditizing model layer and a buyer who can't easily build the workflow themselves.

The buying firm. The company deploying agents internally or reselling agent-powered services. It captures surplus as either cost savings (fewer humans for the same output) or margin expansion (same headcount, more output). Crucially, the buyer also decides whether to pass savings to its own customers or keep them.

Labor and end customers. Workers capture gains only if wages or conditions improve, or if augmented workers become more valuable and bargain for it. Customers capture gains only through lower prices or better service. Both are downstream, and both have historically been the last to get paid.

McKinsey's work on generative AI's economic potential frames the total prize in the trillions, but its analysis of generative AI's productivity frontier is careful to note that capturing value is a separate problem from creating it. Creation is a technology question. Capture is a power question.

Why Pricing Models Decide the Split

Here's the part that gets underappreciated: the pricing model isn't just how a GaaS vendor charges, it's the mechanism that pre-allocates the surplus before anyone notices.

Seat-based pricing (a flat fee per user per month) caps how much the vendor captures and hands the upside to the buyer. If you pay $30 a seat and the agent does the work of three people, you keep the gains from that 3x. This is why incumbents love seat licenses and why pure-play agent vendors are fleeing them.

Per-task pricing ties revenue to volume. The vendor captures more as usage grows, which aligns them with the buyer's scale but still leaves the per-unit surplus mostly with the buyer.

Per-outcome pricing is the aggressive move. The vendor charges only when the agent achieves a defined result, a resolved ticket, a booked meeting, a recovered payment. The pitch is "we only win when you win." The reality is that outcome pricing lets the vendor capture a slice proportional to the value of the outcome rather than the cost of producing it. That's the whole game. A debt-recovery agent that charges 15% of recovered funds is capturing value, not selling compute. Andreessen Horowitz has argued that outcome-based pricing is where durable agent businesses will be built, precisely because it decouples revenue from the falling cost of inference.

The tension is obvious once you see it. The cost floor of doing the task keeps dropping as models get cheaper. If a vendor prices on cost-plus, its revenue collapses with inference prices. If it prices on outcomes, it captures a stable share of the value no matter how cheap the underlying compute gets. So the smartest GaaS companies are racing to reframe what they sell from "an agent that does X" to "the outcome of X getting done." Every dollar of that reframing is a dollar of surplus moving from the buyer's column to the vendor's. (This dynamic gets its own full treatment in the cluster's piece on agents and the deflation of professional-services pricing.)

The Historical Rhyme: Who Got the Gains Last Time

We have run this experiment before, more than once, and the results are not encouraging if you're a worker.

When ATMs arrived, the conventional fear was mass teller unemployment. What actually happened is more instructive: ATMs cut the cost of running a branch, banks opened more branches, and the number of tellers held roughly steady for years while their jobs shifted toward sales and relationship work. The productivity gain mostly accrued to banks, partly to customers through wider access, and tellers neither captured it nor were immediately wiped out. They were repriced.

The web and software-as-usual tell a harsher story. Decades of IT investment produced enormous productivity, but the share of national income going to labor declined across most advanced economies, with much of the surplus flowing to capital and to a small set of superstar firms that could deploy the technology at scale. The gains were real. They just didn't land on paychecks.

The pattern that rhymes across both cases: productivity gains flow to whoever holds the scarce, hard-to-replicate asset. With ATMs it was the bank's branch network and license. With software it was proprietary platforms and data. With agents, the scarce assets are turning out to be proprietary workflows, integration depth, regulated access, and trusted brand, not the models themselves, which are commoditizing fast. That tells you where to look for the capture.

Where the Surplus Is Actually Landing Right Now

Cut through the projections and look at observable behavior in 2025-2026 deals.

The model layer is capturing less than it hoped. Inference prices have fallen by more than an order of magnitude in two years, and open-weight models keep arriving good enough for most agent tasks. The foundation labs make extraordinary revenue, but their pricing power over any single agent workflow is weaker than the headlines suggest. Compute is becoming a pass-through cost.

Vertical agent vendors are capturing the most per transaction, when they own a regulated or integration-heavy niche. A coding agent, a medical-coding agent, a claims-adjudication agent: these embed deep domain logic and compliance that a buyer can't casually replace, so the vendor holds pricing power and outcome-based contracts stick.

Buying firms are capturing the bulk of the aggregate surplus, quietly. This is the least-reported truth. Most companies deploying agents are pocketing the savings as margin, not passing them to customers and not (yet) sharing them with retained workers. The labor share of the gain is, so far, small. Stanford's Digital Economy Lab and similar trackers have begun documenting that early generative-AI productivity gains concentrate among firms and the most capable workers, not the median employee.

Workers capture gains in exactly one situation: when they become the person who designs, supervises, or audits the agent fleet, the "agent boss" role, and can credibly claim the output as theirs. A support rep who now oversees fifty agents and owns the escalations is more valuable and can bargain for it. A rep who simply got replaced captured nothing. The fork between those two outcomes is one of the defining labor questions of this cluster, explored more fully in the beat's piece on humans managing fleets of agents.

End customers capture gains only in competitive markets. Where buyers compete hard on price, savings get passed through as deflation. Where buyers have pricing power, customers see the same prices and slightly faster service while the firm keeps the rest.

The Levers That Decide Your Share

If "who captures the gains" is a bargaining-power question, then capturing more of it is about owning the levers. Five matter most.

Own the workflow, not the model. The model is rented and commoditizing. The proprietary workflow, the specific sequence of steps, the evaluation harness, the institutional knowledge of what "done correctly" means in your domain, is the durable asset. Whoever owns that owns the negotiation.

Own the customer relationship and the data exhaust. Whoever the end customer trusts and whoever holds the proprietary data that makes the agent better captures compounding advantage. Data from running the agent improves the agent, which is a flywheel the model provider doesn't sit inside.

Control the integration surface. An agent wired deeply into a buyer's systems of record has high switching costs. High switching costs are pricing power, and pricing power is surplus capture. This is why vendors race to become infrastructure rather than tools.

Choose your pricing posture deliberately. If you're a buyer, fight for seat or capped pricing and resist outcome-based deals that hand your value upside to a vendor. If you're a vendor, do the opposite. The choice of pricing model is not administrative; it is the single largest determinant of who gets the $5.60.

For workers and policymakers: bargain over the augmented role, not the displaced one. The leverage workers retain is in the judgment, accountability, and trust that agents can't yet supply. The reskilling and policy questions, whether displaced workers can move into agent-supervision roles fast enough, and whether anything structural redirects surplus toward labor, are where the societal split gets decided. The Brookings and academic literature on labor share decline is essentially a warning label: without deliberate intervention, the default path sends the gains to capital.

The uncomfortable summary is that agents don't redistribute anything on their own. They create a surplus and drop it into an existing distribution of power. If that power structure favored capital and platform owners before agents, it will favor them more afterward, unless workflow ownership, competition, or policy bends it.

Insights Most People Overlook

The model layer is the least likely big winner, despite owning the breakthrough. Everyone assumes the foundation labs capture the gains because they built the intelligence. But commoditization is brutal: when three open-weight models do 90% of agent tasks for a tenth of the price, the model becomes a pass-through input like electricity. The capture migrates to whoever owns the scarce complement, workflow, data, distribution. Inventing the engine rarely means owning the car company.

Outcome-based pricing is a surplus-transfer mechanism disguised as fairness. "We only get paid when you succeed" sounds buyer-friendly. It's actually the vendor's cleanest path to capturing value rather than cost, and as inference costs fall, the gap between the outcome's value and its production cost is exactly the surplus the buyer is signing away. Buyers who don't model the long-run cost curve will look back at outcome contracts as the moment they gave away their own productivity dividend.

The biggest current winner is invisible because it's not selling anything. It's the ordinary buying firm quietly converting saved labor into retained margin. There's no press release for "we kept the savings." This makes the labor-share shift hard to see in real time and politically slow to address, the gains are diffuse and undramatic right up until they show up in employment data.

Augmentation and replacement aren't a spectrum of outcomes, they're the same fork that decides worker capture. The same agent can leave a worker more valuable (now supervising a fleet, owning escalations) or fully displaced, and the difference often comes down to whether the firm chose to redesign roles or just cut them. Worker surplus capture is therefore an organizational design decision, not a technological inevitability.

Deflation is the only mechanism that forces gains downstream to customers, and it requires competition the GaaS market may not preserve. If agent-driven service markets consolidate to a few winners (a live question in this cluster), pricing power stays with vendors and buyers, and customers never see the savings. The endgame market structure of GaaS thus quietly determines whether the public ever shares in the productivity gain at all.

References

• The economic potential of generative AI: The next productivity frontier, McKinsey
• The Rise of Agentic AI Pricing, Andreessen Horowitz (a16z)
• The Decline of the U.S. Labor Share Across Industries, Brookings
• Stanford Digital Economy Lab