When Microsoft or JPMorgan or any other public company announces earnings, you will frequently read about how it beat the consensus estimates put forth by financial analysts. But how exactly did that consensus arise? And why do the estimates—pooled by a financial-data provider—seem to so often underestimate companies’ performance? The answer may surprise you.
I dove into this topic while reporting on an AI startup that claimed it could beat the analysts’ targets, and came across an eye-opening 1997 Fortune story titled “Learn to play the Earnings Game (and Wall Street will love you).” Even then, it turns out, CEOs had figured out how to work analysts by priming them with pessimistic expectations of their companies’ results, in order to massage the target number downward—and make it easy to “beat” expectations when the actual number came out. Check out this fun nugget from the 1997 story:
“After a typically grim presentation by CEO Bill Gates and sales chief Steve Ballmer at an analysts’ meeting two years ago, Goldman Sachs analyst Rick Sherlund ran into the pair outside and said, ‘Congratulations. You guys scared the hell out of people.’ Their response? ‘They gave each other a high five,’ Sherlund recalls.”
The story, which is worth checking out for its vintage Gates and Ballmer photo alone, went on to point out how the companies atop Fortune’s 1997 Most Admired list—Coca-Cola, Merck, Microsoft, Johnson & Johnson, Intel, Pfizer, and Procter & Gamble—had each missed the consensus target fewer than five quarters in the preceding five years.
Fast forward to today, and not much has changed. The head of research at Zacks Investment Research, one of the firms that compiles consensus estimates, told me candidly that “there is safety in numbers” when it comes to forecasting. He explained that an analyst might have a strong conviction on the quarterly numbers a company will post, but that they will typically adjust it to more closely align with what others are predicting. This account underscores how what seems to be a data-driven exercise is also subject to the human tendency to herd our views, rather than go out on a limb.
Companies, meanwhile, have largely taken a page from Microsoft’s 1990s playbook, and become adept at deploying their investor relations teams to nudge analysts towards lowball target numbers that they can beat.
That may come to an end, however, in the AI era, with companies like Samaya compiling estimates with AI instead of human judgment.
In response to the startup’s claim it could provide a more accurate earnings forecast, I invited the CEO, Maithra Raghu, to prove it. She accepted my challenge and agreed to provide predictions for 50 companies, including Amazon and Palantir, before they reported quarterly earnings. The result? The startup beat the collective wisdom of the analysts for 40 out of the 50 companies.
This is not surprising since, unlike analysts, an AI model is not susceptible to peer pressure or being influenced by friendly phone calls or doom-filled briefings from company executives. This is good news for investors trying to get a clear view on how a given firm will actually perform. The news is less good for companies, which may soon have to beat target earnings numbers that haven’t been massaged downwards.

