Who Receives the Gain?

A company that runs customer support for other businesses brings in an AI system, and its cost per ticket falls from $10 to $6.

Ask four people where that $4 goes and you'll get four confident answers. The executive books it as margin. The support agents assume it comes out of their jobs. The AI vendor prices its product as if the saving were its own. The economist says customers will get most of it. Each of them is right in some markets and wrong in others, and which one is right is predictable.

A productivity gain gets divided by competition and bargaining power; who did the work has little say. Where rivals can match the saving, most of it flows to customers. Where a firm is alone or protected, owners keep it. Workers get a share when they're hard to replace. A vendor gets what the next-best alternative lets it charge. None of this is special to AI, which helps: there's a long record of how earlier technologies' gains were divided, and AI's early evidence fits it.

One saving, four destinations

Here is the company in round, illustrative numbers. It handles a million tickets a year at $10 each: $8 of agent labor (100 agents at a fully loaded $80,000) and $2 of everything else. The AI resolves about half the tickets without an agent and cuts agent time on the rest by an eighth. Labor per ticket falls to $3.50, other costs to $1.50. That's a gross saving of $5 a ticket. The AI itself costs $1 a ticket. Net cost: $6, a 40 percent cut.

Now put the identical company, with the identical tool and the identical people, into four different markets.

Scroll sideways to see the whole figure

A. A crowded market. The company is one of nine comparable providers, and all of them adopt similar tools within a year or two. Contracts come up for renewal, a rival bids lower, prices fall.

About 40 cents. Customers end up with $3.60 of the $4, the vendor keeps its $1, and the company is leaner but earns almost nothing extra.

B. Alone in adopting. Four providers, and this one is the only adopter, perhaps because its clients are in a regulated industry and it alone did the security and compliance work. Its rivals' costs haven't moved, so the market price barely does. Customers get 80 cents. Owners get $3.20 a ticket, $3.2 million a year. This is the scenario every business case quietly assumes.

C. Workers with bargaining power. As in B, but agents are hard to hire and the company needs its experienced people for the escalations the AI can't handle. It reduces headcount only as people leave and raises pay for those who stay. Assume that costs 30 percent of the net saving, $1.20 a ticket. Owners still take $2.24. (The 30 percent is an assumption for the example; the evidence below suggests workers' real share so far has been far smaller.)

D. A vendor with pricing power. As in B, but the vendor knows what its product saves and the company has no comparable alternative. At renewal it moves to outcome-based pricing worth $3 a ticket. The company's net saving shrinks from 40 percent to 20, and the vendor takes $3 of the $5.

Two things in the figure deserve a second look. In A, B and D the workers' share is zero, and in all three the company needs about 44 agents instead of 100. The cost of that adjustment falls on the people who leave, and it isn't in the bars at all. And the company's productivity is identical in all four. Same technology, same workflow, same people. Only the market around them changed.

The same saving, with the same tool and the same people, ends up in four different places depending on the market around the firm.

The mechanism: how much of a cost cut reaches the price

The split in A and B rests on one idea, pass-through: the share of a change in a firm's costs that shows up in the price it charges.

The intuition is worth having. When every firm's cost falls, each one can profitably undercut the others, and they keep doing so until most of the saving has moved into the price. When only one firm's cost falls, the market price is still set largely by everyone else's costs, so the lone adopter can keep its price near the old level and pocket the difference.

In the textbook version (identical services, a fixed number of firms, straight-line demand), the numbers are clean. With N comparable firms, if all of them get the saving, the price falls by N/(N+1) of it. If only one does, the price falls by 1/(N+1). Nine firms all adopting: 90 percent to customers. One adopter among four: 20 percent.

Scroll sideways to see the whole figure

The gap between the two curves is the advantage of being first, and it closes as rivals catch up. Real markets bend the exact numbers, so read the chart for its shape.

Customers: the saving travels as far as competition pushes it

History leans hard toward customers. William Nordhaus estimated that between 1948 and 2001, innovators in the US nonfarm business sector kept about 2 percent of the total value their innovations created. The rest went to buyers, because competitors copied fast. That covers all technology, not AI, but it sets the default expectation: over decades, a technology's gains end up mostly with the people who buy things.

The counterweight is concentration. Where firms are few, pass-through is low. One much-cited estimate has aggregate markups of US public firms rising from about 21 percent above cost in 1980 to about 61 percent by the mid-2010s, driven by a minority of high-markup firms. The size is disputed, but in industries like that, scenario B describes the saving better than A.

Customers also get gains that never appear in a price. In a large field study of AI in customer support, customers of AI-assisted agents wrote noticeably friendlier messages and asked for a manager almost 25 percent less often. A company that tracks only cost per ticket sees its own gain and misses the customers' share entirely.

Workers: whose expertise the tool copies

That same study, of 5,172 support agents, is one of the best pieces of evidence on what AI does inside a job. Issues resolved per hour rose 15 percent on average and 36 percent for the least-skilled fifth of agents. The most skilled saw small gains and small declines in quality. New agents with the tool performed as well as agents with more than six months' experience without it.

The mechanism is the interesting part. The assistant learned from conversations to which top performers contributed many of the examples. It took what the best agents knew and handed it to everyone in real time, compressing the distance between the best and the rest, the reverse of earlier computing waves.

What that does to pay depends on bargaining. If two months with the tool matches six months without it, experience is worth less in the market. David Autor and Neil Thompson found the general pattern over four decades of US data: automation that removes the easy tasks from a job raises wages and shrinks employment, because the remaining work needs more expertise. Automation that removes the expert tasks does the reverse. Support work with an AI that supplies the diagnosis looks like the second case. Whether that's bad news depends on whether the job grows into harder work or shrinks to what the machine left behind (A Job Is a Bundle of Tasks).

So far, measured pay has barely moved. A Danish study linking chatbot use to administrative records rules out effects on earnings or hours larger than 2 percent two years after ChatGPT's launch. The adjustment that has shown up is in hiring. US payroll data shows employment of 22- to 25-year-olds in the most AI-exposed occupations about 19 percent below where it would be had it kept pace with less-exposed peers, mostly through fewer hires rather than layoffs, with no comparable gap for experienced workers. The authors present this as an early descriptive signal, not proof of cause. It does fit how businesses are deploying these systems: in Anthropic's data from September 2025, 77 percent of business API traffic showed the model doing the task, not helping a person do it.

For incumbent workers, then, the share so far is close to zero. For the people who would have been hired, it's negative. Workers capture more, as in scenario C, when the labor market is tight, when AI makes their remaining judgment more valuable, or when the firm needs them to grow. None of those is a property of the technology.

Vendors: paid up to the next-best alternative

Software has long been priced per seat. An AI product that works reduces the number of people, so a vendor charging per agent seat would lose 56 percent of its revenue from our company as headcount went from 100 to 44. That's why vendors are moving to charging per action or per result. As of October 2026, Intercom lists its Fin agent at $0.99 per outcome and Salesforce lists Agentforce at $2 per conversation.

Outcome-based pricing makes the vendor's share explicit: a price per resolved ticket is a claim on part of each ticket's saving. How large a part depends almost entirely on what the customer could do instead.

Run the numbers for scenario D. The vendor charges $3 a ticket. Suppose a rival product, or a system built on an open-weight model, would deliver the same saving for $1.50, and switching (integration, retraining, a security review, a few weeks of worse service) costs $1.5 million once. Spread over a three-year contract of three million tickets, that's 50 cents a ticket. The vendor can charge up to about $2 before switching pays. Its $3 price survives only if no alternative exists below $2.50.

What the AI is worth sets the ceiling on a vendor's price. What the next-best option costs sets the price.

And the next-best option keeps getting cheaper. The price of reaching a fixed level of model capability has been falling somewhere between 9 and 900 times a year, depending on the task. Open-weight models put a floor under it (Open Weights and Closed Labs). A vendor's durable share rests on what's hard to replace: integration into the customer's systems, data from running the workflow, accountability for the result. A thin layer over someone else's model is priced against a cost that falls every year.

Owners: margins last as long as something stays scarce

Scenario B is where owners want to be, and the pass-through chart shows why it rarely lasts. As rivals adopt the same tools, often from the same vendors, the firm slides from the lower curve to the upper one and the saving migrates to customers. Firms have to keep adopting just to stay competitive, running to stand still, which is why this is sometimes called a Red Queen race. Nicholas Carr made the same argument about information technology in 2003, and it fits best the parts anyone can buy off the shelf.

An AI gain that rivals can buy from the same vendor is a temporary margin and a permanent price cut.

Two kinds of firm keep more. The first changes the flow of work instead of speeding up a step. A rival can buy the same tool in a month. It can't buy new roles, redesigned handoffs and a process rebuilt around what the AI does well, which take a year of organizational effort and are specific to the firm (Where the Saved Time Goes). The second owns the complements the gain depends on: data it has the right to use, the customer relationship, the permissions to operate inside clients' systems. If the AI only works well with ten years of resolved tickets from one industry, whoever owns those tickets keeps a share no vendor can price away (Where the Moat Moves).

How big is the pie, and will wages catch up?

Economists disagree about the size of the gain by more than an order of magnitude. Daron Acemoglu, building from measured savings on tasks AI handles profitably today, gets a gain in total factor productivity (the part of growth that comes from using labor and capital better) of about 0.06 percentage points a year over the next decade, and calls that an upper bound. Anthropic, using observed time savings across the tasks people already bring to Claude and assuming adoption spreads, estimated about 1.1 points, then revised its own figure down. The difference comes from three inputs: how many tasks AI touches, how much it saves on each, and how fast firms reorganize. Nobody knows yet which is closer, and I wouldn't build a plan that depends on either.

The strongest objection to everything gloomy above is that wages have tracked productivity over long periods. Between 1973 and 2016, each extra point of productivity growth was associated with roughly 0.7 to 1 point of extra growth in average and median pay. If that holds, workers get their share eventually.

I think that's right about direction and the long run, and weak about timing and about who in particular. During Britain's industrial revolution, output per worker grew for decades while real wages stagnated. The link held over the long run and didn't help the people who lived through the first decades. And much of the long-run link runs through customers: when competition pushes a gain into lower prices, everyone who buys gets richer in real terms. That's why scenario A, bad for the firm, is the good outcome for most people. It works only where competition actually passes the gain through. For anyone deciding something in the next five years, the timing and the particular are most of the question.

What to do

Decide which market you're in before you build the business case. How many rivals will have the same tool within two years? How scarce are your people? How replaceable is your vendor? Those answers tell you more about your margin than any demo.

If rivals can buy the same AI from the same vendor, plan for customers to get most of the saving. Treat the margin as temporary and price accordingly. What lasts comes from reorganizing the work and owning complements others can't buy.

Negotiate with your vendor against the next-best alternative, not against the saving. Keep a credible alternative alive, keep your data and evaluation cases portable, and know your switching cost in dollars. That number sets your vendor's price more than any value calculation.

Be honest about the workers' share. The early evidence shows adjustment through hiring, concentrated on young people entering exposed jobs. If you want experienced people to stay and keep teaching the system what good work looks like, that has to be paid for.

Check yourself

Why does a lone adopter keep more of a cost saving than one of nine adopters?

When only one firm's cost falls, the market price is still set largely by everyone else's costs, so the price barely moves. When all firms save, they undercut each other until most of it reaches customers.

What sets a vendor's price, according to the essay?

What the AI is worth sets the ceiling, and what the next-best option costs sets the price. A vendor can charge only up to the point where switching starts to pay.

What lets a firm keep more of an AI gain over time?

Changing the flow of work, which a rival can't buy in a month, and owning complements the gain depends on, such as data it has the right to use and the customer relationship.