AI Is Writing 40% of Corporate Earnings Calls, Can You Tell the Difference?

When you listen to enough earnings calls, a pattern that is initially difficult to identify begins to show. The sentences are finished. There are seamless transitions. The forward guidance language has been carefully calibrated to move a stock without making any legally actionable statements. Executives discuss competitive positioning for ten minutes without making any noteworthy assertions. If you’ve been listening to these calls for a long enough period of time, you may have noticed a slight shift in your thoughts over the past year or two. Perhaps more polished. Or even more hollow. A bit too uniform. Some finance chiefs are now using AI specifically to determine which words are likely to impact investor sentiment based on historical data, essentially reverse-engineering the emotional response of an earnings call before the call even occurs, according to Fortune’s reporting on what CFOs say off the record.

The data supporting AI’s widespread use in business communications is significant. Compared to a 5-year average of 114 and a 10-year average of 72, FactSet discovered that 210 S&P 500 companies used the term “AI” during earnings calls in the first quarter of 2025—the fifth consecutive quarter in which more than 200 companies used the term. AI was mentioned by 94% of businesses in the information technology sector. In addition to AI as a business topic, that figure also illustrates AI as a rhetorical tool, a concept that has been deeply ingrained in the vocabulary of public companies’ investor presentations. The question eventually changes from how many businesses are discussing AI to how many businesses are using AI to communicate.

The practice isn’t exactly secret, but it’s also rarely made public. Sammy Abdullah’s analysis of the earnings calls of forty publicly traded SaaS companies, including Salesforce, ServiceNow, Microsoft, Snowflake, Datadog, Cloudflare, and others, revealed that while direct AI revenue is still in its infancy for the majority of the companies involved, AI deployments at the enterprise level are creating margin pressure rather than margin expansion. This indicates that there is frequently a significant discrepancy between what executives say about AI during calls and what the financial reality actually reveals. AI deployment is costly. The returns are still in their early stages. However, the language used in these same calls regarding AI seems assured, calculated, and progressive. AI drafting tools are particularly adept at creating that kind of framing.

Beyond the somewhat unsettling picture of a CFO examining AI-optimized language before entering an investor call, there is something worthwhile to consider here. Theoretically, earnings calls are the time when publicly traded companies address their owners directly, explaining what went wrong, why, and what they anticipate happening going forward. Regarding material misstatements in these communications, the SEC has stringent regulations. However, the rule was not intended for AI systems that generate technically correct language optimized to create positive sentiment, but rather for humans attempting to deceive. According to a BBC and EBU study released in October 2025, about 45% of AI queries result in incorrect answers, which are frequently given with complete confidence and no discernible uncertainty. The same underlying architecture underpins the technology used to draft investor communications.

As this develops, it seems as though the institution’s earnings call is experiencing a subtle form of inflation. Language inflation, not financial inflation. Compared to ten years ago, the calls are now longer, more seamless, and heavily referenced. Every business has a set of secular tailwinds, a multi-year roadmap, a framework, and a platform approach. The people who are actually succeeding in AI seem to be those who are essentially carrying out the same tasks as before, but with improved vocabulary. It has always been necessary to read the financials rather than listen to the prose in order to distinguish between the two. Instead of making that task easier, AI-generated prose makes it more difficult.

Underlying much of this language is a very complex financial reality. The datacenters being constructed in 2025 at a total cost of about $400 billion will produce an estimated $40 billion in annual depreciation against $15 to $20 billion in current AI revenue, according to a long-form analysis published this past August. The numbers are still out of balance. If the market develops as optimists predict, they might eventually. However, the difference between the current returns and where the money is going is so great that any sincere executive speaking on the spur of the moment would probably admit the uncertainty. Sentiment-optimized AI-drafted language typically ignores that ambiguity. It usually presents it as a chance.

Whether this matters in a way that investors or regulators will eventually make public is still up in the air. Many businesses would contend that using AI to prepare investor communications is no different from hiring a group of communications specialists to craft the language. There is some validity to that argument. Speed, scale, and the elimination of human friction—which occasionally results in unintentional candor—are the pertinent differences. When a CFO makes a small error when responding to a challenging question, that error conveys information. The stumble does not occur when the same response has been pre-optimized by a language model that has been trained to inspire investor confidence. And along with it, the information it would have contained vanishes.