On earnings calls and in filings, S&P 500 leaders sound enthralled with artificial intelligence. Behind the hype, many still struggle to quantify benefits, define use cases, or prove returns.

NEW YORK — The corporate love affair with artificial intelligence is now entering its awkward teen years. On paper, the revolution appears unstoppable: record investment, surging data-center spending, and an arms race among cloud giants to deploy ever‑larger models. But sift through second‑quarter earnings calls and regulatory filings from the largest U.S.-listed companies and a more complicated picture emerges. Leaders talk about AI constantly—often with evangelical confidence—yet concrete explanations of where and how value will be created remain scarce outside a small circle of frontrunners.

That disconnect is visible in the numbers. FactSet tallies a record wave of AI name‑checks on S&P 500 calls this summer, while McKinsey and Stanford report that enterprise AI adoption and investment have never been higher. Still, few management teams can quantify productivity gains, margin lift, or incremental revenue attributable to generative AI initiatives. In many transcripts, ‘AI’ has become a stand‑in for ‘we’re modernizing’—a vibe, not a plan.

The market doesn’t seem to mind. Mega‑cap leaders—Microsoft, Alphabet, Meta, Amazon, and Nvidia—have turned AI into tangible products, platform economics, or infrastructure sales. For them, describing benefits is straightforward: higher cloud consumption, premium software tiers with AI add‑ons, and surging demand for GPUs. Outside the club, AI shows up as aspiration: better customer service someday, faster code eventually, smarter supply chains at some point. Many are still running pilots, bolting copilots onto workflows, or waiting for vendor road maps to mature.

If the last decade’s cloud migration taught executives anything, it’s that the bill arrives before the payoff. AI is the same, but more so. Companies face rising spend for model access, vector databases, guardrails, observability tools, and prompt‑safe data pipelines—plus a new wave of security, compliance, and legal exposures. Boards sign off because no one wants to be the company that missed the platform shift. Call it corporate FOMO: the fear of missing optimization.

Earnings‑call language gives the game away. When CEOs can describe AI benefits, they tend to be narrow and operational: a few percentage points off contact‑center handle times; faster product‑content cleanup; a coding assistant that reduces toil for developers. These are real improvements, but they’re also incremental and—crucially—hard to scale without rewiring processes. The majority of executives still talk in broad strokes about ‘unlocking productivity’ or ‘reimagining the experience’ without specifying the P&L line that moves.

Listen closely to the leaders who do quantify. Cloud platforms report accelerating AI workloads and premium pricing for enterprise copilots. Chip suppliers sell out next‑gen accelerators years in advance. Industrials adjacent to the build‑out—power and cooling, specialized construction, and heavy equipment—can point to order books tied to data‑center demand. But for consumer brands, retailers, and many service businesses, the translation from AI rhetoric to ROI is still pending.

Part of the problem is measurement. Generative models blur the line between automation and augmentation: does a marketer with an AI assistant produce more campaigns, or just ship similar work faster? Is a sales team closing bigger deals because of AI summaries or because macro headwinds eased? Attribution is messy; control groups are fragile; and early wins often reflect novelty effects—usage spikes because the tool is new. When pilots end and procurement asks for an annualized business case, the math gets uncomfortable.

Governance and risk are another brake. After the SEC brought cases against firms accused of ‘AI‑washing,’ legal teams are pouring cold water on exuberant claims. Compliance chiefs want model inventories, bias testing, and incident playbooks before projects scale. Security leaders, burned by prompt‑injection demos, demand tight access controls and data‑loss prevention. None of this is bad; it’s just slow.

Meanwhile, the talent equation is shifting. Yes, there’s a bidding war for staff who can tune, evaluate, and ship model‑centric systems. But the real bottleneck is leadership: companies that treat AI like a bolt‑on tool rarely see step‑change outcomes. The leaders that do—often the same set dominating cloud and chips—treat AI as an operating‑model change: re‑platforming data, retraining teams, and refactoring processes so copilots aren’t just a toolbar but the default way work gets done.

The most credible near‑term enterprise benefits fall into four buckets:

• Customer contact: call‑deflection and response drafting reduce agent time per case; quality‑assurance models flag compliance issues earlier. • Product content and search: cleaner product data, richer descriptions, and AI‑native search improve conversion for large e‑commerce catalogs. • Software delivery: code assistants accelerate boilerplate and documentation, especially when connected to internal APIs and runbooks. • Knowledge workflows: summarization and retrieval shrink the cost of finding the right page in sprawling policy and research libraries.

Notice what’s missing: dramatic revenue stories outside the platforms and chipmakers. Some consumer‑app experiments have found traction—AI photo editing here, AI travel planning there—but these tend to be feature‑level wins, not business‑model shifts. Enterprise sellers are learning that customers will pay modest seat uplifts for copilots, but large‑scale price‑mix expansion requires hard proof of time savings or error reduction.

Investors are trying to separate signal from noise. On the one hand, record AI mentions coincide with record capex plans from hyperscalers and record private investment across the ecosystem. On the other hand, valuation premia for many would‑be adopters rest on confidence intervals, not cash flows. That gap is sustainable—until rates rise again, budgets tighten, or a high‑profile AI project publicly face‑plants.

What would make AI talk more than talk? Three shifts:

First, move from use‑case shopping to process redesign. The companies that publish credible numbers map AI to a broken process and re‑engineer the workflow end‑to‑end—data readiness, policy, training, and metrics included. An assistant that drafts emails is nice; a re‑platformed claims process with straight‑through automation is measurable.

Second, standardize measurement. Set baselines for time‑to‑complete, first‑contact resolution, defect rates, or revenue per visit before pilots begin. Run true A/B tests, not anecdotes. Tie procurement to signed‑off savings and commit to killing projects that miss thresholds.

Third, get real about costs. Model‑inference expenses, safety layers, and vendor markups can turn a productivity play into a margin drag at scale. That doesn’t mean ‘don’t do AI’; it means architecting for cost control—caching, smaller task‑specific models, and hybrid search‑plus‑generation patterns that avoid wasting tokens.

Some sectors won’t wait. Defense primes are embedding computer vision and autonomy into programs with multi‑year funding visibility. Healthcare revenue‑cycle vendors are capturing denials faster with pattern‑recognition. Heavy‑equipment makers and miners benefit indirectly as data‑center construction drives demand for power gear, cooling, copper, and steel. These are the places where executives can already put numbers on a slide.

For everyone else, 2025 is the year of narrowing the promise. The fear of missing out has done its job: boards are engaged, budgets are funded, and AI programs have senior sponsors. Now comes the hard part—less demo theater, more delivery. The companies that will stop talking and start compounding value will sound boring: fewer buzzwords, more before‑and‑after charts, and a finance leader who can point to a line item and say, ‘AI did that.’