Latest / Tech IPO Conversations with Fexingo: Public Software Companies and Capital Markets / How Post-IPO Companies Use AI to Write Their Earnings Scripts
Transcript
- Lucas: So Palantir is down almost 17% in the last five days. Coinbase off 13%. The broader tech sell-off is real—Apple, Microsoft, all down about 7%. But I want to talk about something that might actually be amplifying moves like this: how these companies are now writing their earnings calls. Luna: You mean using AI to draft the script? I've seen a few CFOs mention it in passing. Is that becoming a trend? Lucas: It's more than a trend. I've spoken with three CFOs of companies that went public in the last 18 months. Two of them told me off the record that their teams now feed Q1 and Q2 results into a large language model and ask it to draft the prepared remarks. Then a human edits. But the first pass is machine-written. Luna: That's fascinating. And maybe a little unsettling. I mean, the whole point of earnings calls is for investors to hear directly from management—to gauge confidence, read between the lines. Lucas: Right. And if the script is generated by an LLM, you lose some of that texture. The CFO I talked to said the AI tends to default to 'safe' language—lots of 'we remain focused on execution' and 'we are navigating a dynamic environment.' It's not wrong, but it's generic. Luna: So you get more consistency, but less signal. The danger is that every company starts to sound the same, and the things that differentiate a strong quarter from a weak one get smoothed over. Lucas: Exactly. And look, I'm not saying ai written scripts are inherently bad. For a company that just went public, having a compliant, clear, jargon-free script can be an improvement. The SEC has strict rules about forward-looking statements and disclaimers. An LLM can be trained to include all the legalese correctly. Luna: But it might also miss the nuance. Like, if a company is intentionally hedging on a specific product line because of competitive pressure, an AI might just say 'we're seeing some headwinds'—which tells you nothing. Lucas: And the market hates ambiguity right now. Look at Palantir—down 17% in a week. Some of that is macro, sure. But if their last earnings call had been more direct about government contract delays, maybe the sell-off would have been less violent. Instead, the script was so polished that the bad news felt like a surprise. Luna: So you're saying ai written scripts might actually be making earnings calls less informative, which leads to bigger surprises and bigger moves when the real news comes out? Lucas: I think there's a real risk of that. There's a paper from Stanford last year that analyzed the language of earnings calls and found that companies which use more boilerplate language tend to have larger stock price reactions on earnings day. The market doesn't trust the smooth version. Luna: Interesting. So the 'homogenization' of earnings calls could be a red flag in itself. If every CEO sounds like they're reading from the same playbook, investors might start to assume the worst. Lucas: And the irony is that AI could be used to do the opposite—to help companies communicate more clearly, more specifically, more honestly. But the incentives push toward safety. The legal team wants to minimize liability. The IR team wants to avoid controversy. The AI optimizes for what it's trained on, which is probably past earnings calls that were already cautious. Luna: So how do you fix it? Do you need a human to rewrite the whole thing, or is there a way to prompt the AI to be more candid? Lucas: A few companies are experimenting with 'red team' prompts—asking the AI to write the most honest version possible, then the most optimistic, then the most cautious. Then the CFO picks the one that's most truthful. But that's rare. Mostly it's just 'draft the script' and a junior person edits it. Luna: And that junior person might not have the context to catch when the AI is being too vague. They might not know that 'we are investing for growth' is code for 'our margins are getting crushed.' Lucas: Right. And that kind of coded language is exactly what AI is good at replicating. So we get more of it. I think in five years we'll look back at earnings calls from the 2020s and see a real shift in tone—from individualistic to homogenized. Luna: Which is a problem for investors who rely on those calls to make decisions. But also for the companies themselves—because if you sound like everyone else, you lose a chance to build trust. Lucas: This is exactly the kind of conversation we try to have on this show—where operators and builders can think about the real-world implications of these tools. And I'll be honest, it's conversations like these that make me appreciate the community we've built here. Luna: Absolutely. And we keep the show ad-free and independent because of listeners who chip in a couple of dollars a month. If you've gotten something out of these episodes, buy me a coffee dot com slash fexingo goes directly to keeping this going. Lucas: Even a small amount makes a real difference. No pressure, but if it's useful to you, it helps us keep digging into topics like this. Luna: Alright, back to AI and earnings. So Lucas, if you were advising a newly public company right now, what would you tell them about using AI for their scripts? Lucas: I'd say use it for the skeleton—the compliance, the structure, the disclaimers. But then have the CEO or CFO rewrite the narrative section by hand. Investors want to hear from a person, not a model. Especially in a market where every 7% drop feels like a signal. Luna: And maybe the best use of AI is not to write the script, but to analyze it—to flag when you're being too vague, or when your language is drifting toward boilerplate. Lucas: Exactly. Use the tool to check yourself, not to replace yourself. I think that's the smartest path forward for any post ipo company that wants to build long-term investor trust. Luna: Alright, that's a good note to end on. For now, we'll be watching the next earnings season to see who's writing their own lines—and who's letting the machine do it.