Finance · Artificial intelligence
Everyone’s asking whether AI can beat the market. The economy needs a different answer.
For 140 years, finance has charged the economy about two cents for every dollar it handles. The test for AI trading bots is whether they can bring that number down.
At a glance
Finance keeps about 2 cents a year for every dollar of savings it places and manages. It kept about 2.3 cents in 1884, and it still keeps about 2 cents today.
Past technology made trades cheaper, but finance as a whole got bigger and better paid instead of cheaper.
The first AI trading agents are built for more trading. Firms should point freed-up talent at lending and advice, and track whether the two cents falls.
- Per dollar, 1884
- 2.3¢
- Per dollar, 2011
- 2.4¢
- Per dollar, first half of 2026 · Theoria estimate
- 1.9–2.3¢
This fall I read story after story about AI bots that trade stocks. Almost every one asked the same question: did the bot beat the market?
That is a fair question if you are the one handing a bot your savings. For everyone else, it is the wrong one. The question that matters for the economy is what happens to the time, and the people, the bots free up. History suggests an answer, and it should give pause to anyone hoping AI will make finance more productive.
The scoreboard that matters
Finance does a job for the rest of the economy. It moves savings to the people and companies who put money to work, and it charges for that service. NYU economist Thomas Philippon measured the charge: what the financial industry earns for each dollar of loans, bonds, stocks and deposits it creates and looks after.
In 1884 the answer was about 2.3 cents. In 2011, the last year of his published data, it was about the same. Using government data, Theoria extended his measure through the first half of 2026. The answer is between 1.9 and 2.3 cents, depending on whether you count the recent stock-market boom.
That is the number to watch. If AI makes finance more productive for the economy, it has to fall. A bot that beats the market does not move it.
Exhibit 1
Finance still charges about two cents per dollar, as it did in 1884
Cost of US financial intermediation, cents per dollar of financial assets per year
Sources: Thomas Philippon, “Has the US Finance Industry Become Less Efficient?” (American Economic Review, 2015), supporting data, level-based unit cost; BEA and Federal Reserve data via FRED; Theoria analysis.
Note: Theoria estimate = finance and insurance value added ÷ (nonfinancial debt + M2), with and without nonfinancial corporate equity, scaled to Philippon on 2005–07; three-year average to 2025; 2026 = first-half average. It tracks Philippon’s series over 1960–2011 (correlation 0.98) but omits his flow and quality adjustments.
The last time technology freed up Wall Street
AI is not the first technology to take over work in finance. In 1971, Nasdaq opened the first electronic stock market. Online brokers arrived in the 1990s, and computer algorithms spread through trading in the 2000s. Each made trading cheaper. Commissions on New York Stock Exchange trades averaged about 0.9% in 1974. By 1991, the full cost of buying or selling a share had fallen below 0.5%, according to Columbia economist Charles Jones.
Yet finance as a whole did not get cheaper. Its cost per dollar actually rose, from 1.8 cents in 1970 to 2.4 cents in 2011. Finance got bigger instead. Its income grew from 4.2% of GDP in 1970 to 8.3% in 2009. Government figures put it at 9% in mid-2026. And finance paid more to recruit. In 1979 its workers earned about what everyone else did. By 2025 they earned 1.7 to 1.8 times as much, depending on how the newer government figures are matched to the older series.
Exhibit 2
Since 1970, finance has grown bigger and better paid, not cheaper
Three measures of US finance, 1970 to latest
Cost per dollar
cents per dollar of financial assets
1.8¢ → 1.9–2.3¢
No lasting decline
Share of GDP
finance income, % of GDP
4.2% → 9.0%
About doubled (to Q2 2026)
Pay premium
finance pay ÷ all private pay
1.06 → 1.72
Premium 6% → 72%
Sources: Philippon (AER 2015 supporting data); Philippon and Reshef, “Wages and Human Capital in the U.S. Finance Industry: 1909–2006” (QJE 2012); BEA via FRED (VAPGDPFI, N4457C0A052NBEA, A4403C0A052NBEA); Theoria analysis.
Note: GDP share: the two series use different definitions and are shown separately, not joined; the BEA 2026 point is Q2 2026. Pay: average wage per full-time worker in finance ÷ all private industries; the BEA series is rescaled to match Philippon and Reshef on their 1998–2006 overlap (factor 0.97).
Philippon’s explanation is that the gains from technology went mostly into trading. Trading sets prices and lets people buy and sell easily, which the economy needs. But beyond that point, more of it adds little for anyone else. Stanford economist William Sharpe showed why in 1991. Before costs, active investors as a group can only earn the market’s return, so one trader’s win is another’s loss.
Better trading moves money between pockets. It does not make the pie bigger.
Why AI trading agents fit the old pattern
AI trading agents automate exactly the part of finance that history says helps the economy least. And the first ones are being built for more trading, not less.
Robinhood launched agentic trading in May. Its agents can trade stocks, options and crypto. In September it announced “Loops,” standing instructions that would let an agent trade around the clock, and a store of paid data apps for agents at $5 to $30 a month. Both are listed as coming soon. Robinhood says it does not supervise, monitor or audit the agents.
None of this proves agents are inflating trading. Their footprint is tiny. More than 150,000 customers had opened agentic accounts by late September, and in July the accounts held just over $100 million in total. The stock of America’s nonfinancial companies alone is worth about $83 trillion. But the direction is telling. Most of Robinhood’s revenue comes from transactions, which rose 44% in the second quarter, to $776 million. A platform paid by the trade has every reason to build agents that trade.
Exhibit 3
Most of Robinhood’s revenue comes from trades, the activity its AI agents perform
Robinhood net revenue, second quarter of 2026, $ million
59%
of revenue came from transactions, up 44% from a year earlier
- Options$342M
- Event contracts$156M
- Equities$129M
- Crypto$100M
- Other transaction$49M
- Net interest$389M
- Other revenue$143M
Source: Robinhood, “Robinhood Reports Second Quarter 2026 Results” (July 29, 2026); Theoria analysis.
Note: Total net revenue $1,308 million: transaction-based $776 million, net interest $389 million, other $143 million. “Other transaction” = $776 million minus the four named categories. The share of revenue from AI agents is not disclosed.
The best case that this time is different
There is a serious argument that AI will break the pattern. Index funds and free trading really did make investing cheaper for ordinary savers. The average stock mutual fund’s fees fell 62% between 1996 and 2025, to 0.40% a year, according to the Investment Company Institute. Philippon himself found that the raw cost of finance dipped after the 2008 crisis. Theoria’s update shows a dip too, though most of it disappears if you leave out the stock-market boom.
AI could also reach the expensive, human parts of finance that trading technology never touched: financial advice, loan decisions, compliance paperwork. Advice that costs a percentage of your savings every year could cost a small flat fee. And some gains, like the hours ordinary investors no longer spend managing their money, never show up in GDP at all.
That case could prove right. On the evidence so far, we think it is unlikely, because it depends on where firms point the technology, and so far the money points toward trading.
What should happen next
Finance leaders are about to get back a great deal of their analysts’ time. They should decide now where it goes. The work that grows the pie is unglamorous: lending to small firms that banks find too costly to assess, advice for households too small for a human adviser, faster capital for new companies. AI makes each of those cheaper to do. Trading faster against rivals who own the same tools does not.
They should also change how they score AI. Beating a benchmark is one test. The other is cost per client served, or per dollar of lending, and whether that cost falls.
And someone should keep the economy-wide score. Philippon’s measure is built from public data, yet no government agency we could find publishes it as a regular statistic. The Federal Reserve or the Bureau of Economic Analysis could publish it every year. It is the simplest way to know whether AI is making finance cheaper for everyone else, or just making finance bigger.
The two-cent test
Within a year we will likely have early evidence on whether the bots can beat the market. It will take a decade to know whether they can beat the two-cent toll. That is the scoreboard worth watching.
What would change our mind
- The cost of finance falling below its 1964 low of about 1.6 cents per dollar and staying there, on measures with and without stock values.
- Finance’s share of GDP shrinking while lending to small firms and new companies grows.
- Platforms reporting that customers with AI agents trade less, and earn more after costs, than customers without them.
Sources, method and every number checked
Where each number comes from
Data checked October 10, 2026
| Claim | Kind | Source |
|---|---|---|
| 2.3¢ per dollar in 1884; 2.35¢ in 2011; 1.8¢ in 1970; low of 1.6¢ in 1964 | Published data | Philippon, AER 2015 supporting data, level-based unit cost |
| 1.9–2.3¢ in the first half of 2026 | Theoria estimate | Method below; government data via FRED |
| Finance income 4.2% of GDP (1970), 8.3% (2009) | Published data | Philippon, finance GDP-share composite |
| 9% of GDP in mid-2026 | Official data | BEA via FRED, VAPGDPFI, Q2 2026 = 9.0% |
| Finance pay 1.03× others (1979, series low); 1.7–1.8× (2025): 1.72 rescaled to Philippon & Reshef, 1.77 on raw BEA data | Published + official | Philippon & Reshef, “Wages and Human Capital in the U.S. Finance Industry” (QJE 2012); BEA wages per full-time worker, all private industries |
| NYSE commissions about 0.9% (1974); full trading cost under 0.5% from 1991 | Published research | Jones, “A Century of Stock Market Liquidity and Trading Costs” |
| Before costs, active investors as a group earn the market return | Published theory | Sharpe, “The Arithmetic of Active Management,” Financial Analysts Journal, 1991 |
| Technology gains concentrated in trading | Author’s interpretation | Philippon, Brookings, 2016 |
| Stock mutual fund fees fell 62% from 1996 to 2025, to 0.40% | Industry data | Investment Company Institute, March 25, 2026 |
| Raw cost declined after the crisis | Published research | Philippon, NBER w26330, 2019 |
| Agents trade stocks, options and crypto; Loops and data apps ($5–$30 a month) announced as coming soon; no supervision | Company disclosure | Robinhood HOOD Summit 2026 |
| Over 150,000 agentic accounts opened (Sept 29); over $100M in agentic accounts (July); Q2 revenue $1,308M, of which transactions $776M (+44%), net interest $389M, other $143M | Company disclosure | Robinhood Q2 2026 results |
| Nonfinancial corporate stock worth about $83 trillion | Official data | Federal Reserve Financial Accounts via FRED, Q2 2026 |
How Theoria extended the cost of finance
Philippon’s published series ends in 2011. Theoria rebuilt a simplified version from government data: finance and insurance value added (BEA) divided by nonfinancial debt (Fed) plus money and deposits (M2), with and without the stock of nonfinancial companies (Fed).
Test: built the same way with Philippon’s own income figures, this measure tracks his series closely from 1960 to 2011 (correlation 0.98).
Each version is scaled to Philippon’s level over 2005–2007 and smoothed with a three-year average through 2025. The 2026 value is the average of the first two quarters, unsmoothed. The two versions bound the range shown in Exhibit 1.
Limits: Philippon’s full method also counts new lending and stock issuance and adjusts for the quality of borrowers. This estimate does neither, so it is labeled as Theoria’s, not his.
Which Philippon series: the article uses the unadjusted, level-based series (“finlcostma”) from the supporting data file for his 2015 paper (data vintage March 2012), because Theoria’s extension is also level-based. His papers’ headline figures use other versions, so their ranges differ slightly. On the composite column of the same file, the range is 1.56–2.49¢, with 2.0¢ in 1884.
“No agency publishes it”: a web search on October 10, 2026 covering the Bureau of Economic Analysis, the Federal Reserve (Board and FRED) and the Office of Financial Research found no regular official publication of the measure.
Closest earlier work
Paul J. Davies, Bloomberg, “AI Threatens the Finance Industry’s Perpetual Profit Machine” (February 2026) pairs Philippon’s two-cent finding with AI. Checked against the full text on October 10, 2026: it does not examine AI trading agents, where freed-up finance professionals go, or any data after Philippon’s (it does criticize speculative trading as “economically pointless”).