Samaya's AI Beat Wall Street Consensus on 40 of 50 Earnings Calls
A $43 million-backed startup out-forecast Bloomberg's analyst consensus on 40 of 50 companies in a two-week Fortune test, and the reason it won says everything about how those estimates get made.

Samaya, a financial AI startup that raised $43 million in 2025, beat the Wall Street analyst consensus on 40 of 50 companies in a two-week Fortune experiment covering Q3 earnings from October 26 to November 5. The result pressures the consensus-forecasting model that IR teams, analysts, and investors have leaned on for decades, and raises the question of whether AI will expose or inherit the biases baked into it.
Samaya, a financial AI startup that raised $43 million from prominent investors in 2025, beat the Wall Street consensus on 40 of 50 companies in a two-week experiment run by Fortune. The setup was simple and brutal: Fortune picked 50 companies reporting Q3 earnings between October 26 and November 5, spanning ten sectors from health care to energy to IT. For each one, Samaya delivered an AI-generated estimate for the single most important metric, EPS or revenue, at least a day before the release. The benchmark was the analyst consensus supplied by Bloomberg. Then the actual numbers landed, and the scoreboard did not flatter the humans.
The early returns were not a fluke. Samaya forecast Verizon would post $34.52 billion in revenue; the actual figure was $34.31 billion, closer to the AI's call than the $34.77 billion consensus. On UPS, Samaya's adjusted EPS estimate of $1.65 beat the $1.62 consensus, with the actual landing at $1.67. Analysts clawed back two wins on S&P Global and PayPal after dropping the first four, but from there it was a rout: Samaya supplied the better pick on 36 of the remaining 44 companies. It called Palantir's revenue at $2.30 billion against an actual $2.31 billion, while the Street said $2.18 billion. On Permian Resources it nailed the exact EPS figure of $0.69; the consensus was $0.625. The misses matter too. Yum Brands was the worst: Samaya said $1.42, the actual was $1.56, and analysts came closer at $1.49. Samaya also lost on Amazon and Google, specifically on how far each would fall short of revenue expectations.
Samaya CEO Maithra Raghu has a theory for the pattern. Where the Street proved more accurate, she says, it was typically on very large companies that are the subject of enormous volumes of private research, and most of that research never entered Samaya's training material. The startup leaned instead on publicly available data: filings, news reports, and interview transcripts. Its clients include banks and hedge funds, and Raghu says Samaya will likely add analyst notes to its training data for earnings in the future. She also flagged the trap in that move: feeding the model analyst notes could bias it toward the same consensus-seeking instinct that shapes human forecasts.
That instinct has a name and a defender. "I think there is safety in numbers," says Sheraz Mian, Director of Research at Zacks, one of the firms that compiles consensus figures. Mian, himself an analyst, describes a system rooted in human nature and a tacit pressure game run by companies. An analyst may privately believe a quarter will land far from the pack, but the upside of being right with an unorthodox call rarely outweighs the reputational hit of being wrong. So the herd holds. Investor relations teams reinforce it. Mian says IR teams monitor individual analysts closely and are quick to call when they think a forecast is too optimistic. Regulation Fair Disclosure, the SEC rule change from 2000, bars companies from handing material information, including forthcoming earnings, to analysts. It does not stop soft pressure.
Paul Hribar, a professor of accounting at the University of Iowa, describes the mechanics bluntly: "In the long run, they provide optimistic numbers that get walked down to beatable targets at the end, so a higher proportion of companies end up meeting or beating expectations right at the close." That is the walk-down, and it is why a "whisper number" has long circulated on Wall Street, a figure closer to what people actually believe than what is published. The stakes are asymmetric. Beating forecasts typically buys a modest stock bump; missing, even narrowly, invites disproportionate punishment. Fortune made the point in 1997 and it still reads fresh: "The price of missing a quarter has risen sharply, particularly among high-priced growth stocks. In the growth stock fraternity, missing by a penny now implies the height of corporate boneheadedness."
For CFOs and IR chiefs, the experiment is a warning shot aimed at the machinery they operate. The consensus number is not a neutral measurement; it is a negotiated artifact, shaped by analyst incentives, IR nudges, and the walk-down ritual. If independent AI forecasts get good enough, that artifact becomes contestable in public. A model that reads filings and transcripts and lands closer to the truth than the Street does not just embarrass analysts, it hands investors a second opinion that arrives before the print. Boards should expect harder questions about guidance philosophy, about whether the company is managing to a beatable number or actually forecasting, and about what happens when the gap between the whisper and the published consensus becomes visible to everyone.
The open question is whether AI forecasts inherit the bias they expose. Raghu raised it directly, and it is the right worry: train a model on analyst notes and you may teach it to want consensus too. There is also the sample-size caveat. Fifty companies over a single earnings period is a small test, and one quarter is not a track record. But the direction of travel is clear. AI is already reshaping how corporate America reads its own numbers, and earnings season, the most ritualized and most merciless measurement in business, is now in its path.
For now, the practical takeaway for operators is uncomfortable and simple. The consensus game works because everyone agrees to play it. Samaya's 40-for-50 run suggests the game can be beaten by a machine that never got the memo, and that the memo, not the math, was doing much of the work all along.
This story's Key Insights and Take-aways are locked.
Create a free account to unlock Executive Actions for one credit.
Register to UnlockAlways free for Executives Club members. Join the Club
More in Business
S&P 500 Enters Correction, Down 10% From Its High on AI Worries
The index's slide past the 10 percent line is a symbolic verdict on investor mood, and the AI trade that powered the rally is now the thing testing it.
Dow and S&P 500 enter correction as Nasdaq logs worst week since 2020
All 11 S&P 500 sectors fell more than 1% in the same week for the first time since April 2025, a breadth signal that matters more than any single index level.
Chip stocks lose $2.3 trillion in a week, worst run in 25 years
A $2.3 trillion weekly wipeout in semiconductors is forcing every board with an AI story to ask whether it owns a thesis or just a crowded trade.




