Every S&P 500 earnings call in Q1 2026 mentioned artificial intelligence. The word appeared 3,028 times across the quarter’s transcripts, according to an analysis by FactSet — a 47% increase from Q1 2025. CEOs who never showed interest in machine learning now describe their companies as “AI-first.”

Most of this is theater. The gap between what companies say about AI and what they have actually deployed, measured, and scaled is enormous. We surveyed the landscape — analyst reports from McKinsey, Gartner, and Bain, plus earnings disclosures, case studies, and conversations with engineering leaders — to assemble a picture of what is genuinely working, what has quietly failed, and what the honest numbers look like.

Where AI Is Actually Deployed

The McKinsey Global Institute’s 2025 State of AI survey, based on responses from 1,684 enterprises across 12 industries, provides the clearest picture of real adoption by use case.

Two patterns jump out. First, the highest adoption is in functions where the cost of errors is low and the volume of repetitive work is high — writing code, answering customer questions, producing marketing content. Second, adoption drops sharply in regulated, high-stakes domains like finance and legal, where the consequences of AI mistakes are measured in lawsuits rather than inconvenience.

What’s Working: Specific Wins with Real Numbers

Developer productivity

This is the most measurable AI success story in the enterprise. GitHub reported in February 2026 that Copilot Business subscribers accept 35% of code suggestions and report a 55% reduction in time spent on boilerplate tasks. Google’s internal data on its Duet AI coding tools, shared at Cloud Next 2025, showed a 25% reduction in code review turnaround time across its monorepo.

But the headline numbers require context. A study by Uplevel (a developer analytics platform) tracking 800 developers over six months found that while AI coding tools increased code output velocity by 30-40%, the number of bugs per pull request also increased by 15%. The net productivity gain, after accounting for additional debugging and review time, was closer to 20%.

Companies citing specific gains: Shopify (30% faster feature shipping in internal tools, per CEO Tobi Lutke’s Q4 2025 letter), Stripe (40% of internal code suggestions accepted, per engineering blog), Atlassian (25% reduction in Jira ticket resolution time for internal engineering, per investor day presentation).

Customer support automation

The second-most-mature use case. The key metric is “automated resolution rate” — the percentage of customer inquiries resolved without human intervention.

Klarna is the poster child: the company reported in February 2025 that its AI assistant handled 2.3 million conversations in its first month, performing the equivalent work of 700 full-time agents. Average resolution time dropped from 11 minutes to under 2 minutes. Customer satisfaction scores held steady. By Q1 2026, Klarna reported its support headcount had decreased from roughly 5,000 to 3,800 through attrition, not layoffs, with the AI handling an estimated 65% of all first-contact inquiries.

Other companies report more modest results. A Zendesk benchmark study across 1,200 enterprise accounts found a median automated resolution rate of 28%, with top performers reaching 45-50%. The common pattern: AI handles tier-1 inquiries (password resets, order status, return policies) effectively; complex or emotional issues still require humans.

Content operations at scale

Marketing teams have adopted AI tools faster than any other non-engineering function, largely because the risk profile is forgiving. A suboptimal blog post draft costs almost nothing; a suboptimal loan approval costs a lawsuit.

Specific examples: HubSpot reported that customers using its AI content tools produce 3.5x more blog posts per month than non-users, with comparable engagement metrics. Jasper AI’s enterprise customers (brands like VMware, Morningstar, and Sports Illustrated owner Arena Group) report cutting content production timelines from 2 weeks to 3-4 days for full marketing campaigns.

The pattern is consistent: AI generates first drafts, outlines, and variations. Humans edit, fact-check, and make strategic decisions. The productivity gain is in the 2-4x range for volume, but human editing time has not decreased meaningfully.

What Flopped: The Expensive Lessons

The common thread across failures is the same: companies deployed AI to replace human judgment in contexts where human judgment was the product, or they optimized for a metric (cost reduction, output volume) that was poorly correlated with what users actually valued.

The Honest ROI Numbers

Strip away the marketing and here is what the data says about AI’s economic impact in the enterprise.

A few critical caveats:

Revenue attribution is nearly impossible. Almost no company can credibly attribute top-line revenue growth to AI deployments. The gains are overwhelmingly on the cost side — faster processes, fewer support tickets requiring humans, reduced content production costs. Bain’s 2025 enterprise AI survey found that 72% of companies reporting positive AI ROI measured it in cost avoidance, not revenue growth.

The “AI tax” is real. Inference costs, API fees, fine-tuning compute, and the engineering time to integrate and maintain AI systems are significant. A mid-sized SaaS company spending $500K/year on OpenAI API calls for customer support automation is not unusual. The ROI is positive only if the support cost reduction exceeds the AI cost — which it often does, but not by the margin the marketing materials suggest.

Data quality is the hidden prerequisite. Every successful AI deployment we examined had one thing in common: the company invested in structured, clean, accessible internal data before deploying AI on top of it. Companies that skipped this step — bolting AI onto messy wikis, fragmented CRMs, and inconsistent knowledge bases — saw poor results regardless of which model they used.

The Deployment Pattern That Works

Companies with successful AI deployments follow a remarkably consistent lifecycle.

The critical insight is that measurement comes before scaling, not after. Companies that skip the pilot-and-measure phase — rolling out AI tools to the entire organization based on demos and vendor promises — consistently report lower satisfaction and higher costs than those that run controlled pilots with defined success metrics.

Time to value: Based on Gartner’s 2025 AI deployment survey, the median time from pilot kickoff to measurable ROI is 7 months for internal tools and 11 months for customer-facing applications. Companies expecting returns in weeks are setting themselves up for disappointment.

The Gap Between “We Use AI” and Reality

Perhaps the most striking finding across the data: a significant percentage of companies claiming to use AI have not deployed it in any production system.

A 2025 survey by the Census Bureau’s Annual Business Survey found that 7.2% of U.S. firms reported using AI in production, up from 4.7% in 2024. Yet in the same period, over 60% of large enterprises told McKinsey they had “adopted AI in at least one function.” The discrepancy is explained by the difference between “we bought Copilot licenses” and “we have AI systems in production workflows with measured outcomes.”

This gap is starting to close, but slowly. The AI spending is real — worldwide enterprise AI spending reached an estimated $166 billion in 2025, per IDC — but much of it is still in experimentation, proof-of-concept, and pilot phases. The era of broad, scaled AI deployment is coming, but for most companies it hasn’t arrived yet.

The Bottom Line

The companies getting genuine value from AI share three characteristics: they started with a specific, measurable problem (not “transform our business with AI”); they invested in data infrastructure before deploying models; and they measured outcomes rigorously enough to kill projects that weren’t working.

The rest is noise. Impressive demos, breathless earnings call mentions, and “AI-powered” badges on product pages. It is not that AI doesn’t work — it does, in well-defined contexts with clean data and realistic expectations. But the distance between what the industry markets and what it has actually shipped remains the widest gap in enterprise technology.