Why AI’s Productivity Gains Are Showing Up In Earnings And Not Others

NEW YORK, NEW YORK – AUGUST 23: Traders work on the floor of the New York Stock Exchange during morning trading on August 23, 2024 in New York City. Stocks opened up on the rise ahead of Federal Reserve Chairman Jerome Powell’s remarks at the 2024 Jackson Hole Economic Symposium. (Photo by Michael M. Santiago/Getty Images)

Getty Images

The AI productivity thesis has been circulating in equity markets since early 2023: companies that adopt AI tools will see margin expansion as AI substitutes for headcount and accelerates output per worker. Two years into the experiment, the evidence is fragmentary but directionally consistent—some companies are posting meaningful margin gains attributable to AI adoption, and others are absorbing enormous AI-related capital expenditure with no corresponding improvement in operating leverage. The divergence is not random, and understanding what separates the early winners from the laggards is increasingly important for stock selection.

The clearest examples of margin accretion from AI adoption come from companies where the AI application is direct and the output is measurable. Salesforce and ServiceNow have both cited AI copilot features driving sales productivity metrics—more pipeline per rep, faster deal cycles, higher contract values—that show up in revenue per employee metrics. Meta’s AI-driven improvements to its ad auction and targeting systems are well-documented in its financials: revenue per user has been rising while headcount has been held flat or reduced, a combination that drives operating leverage directly. These aren’t abstract claims; they’re verifiable through comparing unit economics across reporting periods.

The contrast with companies in the adoption curve’s earlier stages is sharp. Industrial and manufacturing businesses that have implemented AI tools for predictive maintenance or supply chain optimization often report project timelines of 18 to 36 months before measurable cost reduction shows up in the financials. The implementation phase involves data infrastructure investment, integration with legacy systems, and organizational change management—all of which consume capital before any savings materialize. Healthcare systems pursuing AI-assisted diagnostics are in an even longer cycle, facing regulatory approval processes that extend the timeline further.

Earnings call transcripts from Q1 2025 show a meaningful pattern when sorted by sector and AI capex density. Technology companies that have been integrating AI into software products—productivity tools, development environments, customer service platforms—are increasingly citing AI as a driver of gross margin expansion. The marginal cost of delivering AI-augmented software features is low once the model is built, which creates operating leverage as the feature scales. Professional services firms, by contrast, are spending heavily on AI tools for their consultants and analysts but are finding that client billing rates haven’t moved in proportion—the productivity gain is being competed away in pricing rather than retained as margin.

The capex disclosures deserve specific attention. Companies that are building their own large language model capacity—either through training proprietary models or maintaining substantial inference infrastructure—are incurring capital costs that won’t appear in operating expenses but will eventually flow through depreciation. The accounting treatment can obscure the cash economics: a company spending $2 billion annually on AI infrastructure through capitalized expenses looks more profitable in the near term than its cash generation supports. Tracking free cash flow rather than operating income becomes more important as AI capex scales.

The make-versus-buy decision is one of the more consequential strategic choices playing out in real time. Companies that are building proprietary AI capabilities are making a bet that the competitive moat from owning the model outweighs the capital cost of building it. Companies using third-party models through API access are betting that the model providers will commoditize quickly enough that proprietary ownership is unnecessary—and they’re spending far less capital in the interim. The historical pattern in technology is that the infrastructure layer tends to commoditize over time, which would favor the buyers over the builders, but the current period of rapid model improvement may reward the builders who move fastest.

For investors, the analytical task is to distinguish between companies with genuine AI-driven operating leverage and those that are capitalizing on AI enthusiasm in their narrative without the financial results to back it. The former will show productivity metrics—revenue per employee, gross margin trends, customer acquisition cost efficiency—improving in a consistent direction. The latter will show AI mentions in earnings calls and investor presentations without corresponding financial improvement. That gap between narrative and metrics is where most of the mispricing in the current AI cycle will eventually resolve.