How companies are actually using AI tools in 2026 (not the hype version)

A 2025 McKinsey survey found that 72% of companies reported using AI in at least one business function. That sounds transformative until you read the methodology.

"Using AI" includes one marketing manager with a ChatGPT Plus subscription. A developer with GitHub Copilot autocomplete turned on. The CEO who asked Siri to set a reminder. All of that counts.

Real adoption looks different. Here's what's actually working, and where companies have gotten burned.

Marketing: the first-mover advantage is gone

Two years ago, using AI for content gave you a real edge. Content volume was the bottleneck, AI dissolved it, and teams that figured that out early moved faster than competitors who were still doing everything by hand.

That window closed. Everyone's using AI for first drafts now. The bar has gone up, not down, because there's more content competing for attention.

The teams winning now figured out the second-order move: using AI not to produce content but to test faster. One e-commerce company was generating 50 variants of a product description, A/B testing them across audience segments, feeding results back to refine the next batch. That flywheel, done manually, would require a small army. With Jasper and Writesonic handling generation, one person runs the whole thing.

Where it breaks down: turning off human review to move faster. Factual errors in AI output are often subtle enough that readers don't catch them individually, but they accumulate into a credibility problem. McKinsey found accuracy was the top concern for marketing teams using AI - ahead of cost or compliance. That's not paranoia. It's experience.

Engineering: real gains, but not where you might expect

GitHub's research found Copilot users completed tasks 55% faster. MIT found 40% productivity gains. These numbers get cited constantly.

What's less cited: the gains are skewed toward junior developers on routine tasks. One engineering manager put it plainly: "My juniors got dramatically better. My seniors got marginally better. The gap between them shrank."

The downstream effect is quieter than the headlines. A 4-person team that would have hired becomes a 3-person team that doesn't. Headcounts are staying flatter through natural attrition. It doesn't show up in employment statistics, but it's happening.

The failure mode is predictable: accepting AI-generated code without understanding it. Companies that dropped code review because "the AI wrote it" ended up with security vulnerabilities and architectural debt that took longer to fix than the time saved. The tools work. Skipping oversight doesn't.

Customer support: the one that actually works everywhere

If there's a use case with consistent results across company sizes and industries, it's support. The pattern is almost identical wherever you look. Deploy AI for first-line tickets. Handle 60-80% automatically. Route the rest to humans who now spend their time on things that actually require judgment.

Most support queues are dominated by the same 10-15 questions. "Where's my order?" "How do I reset my password?" "Can I change my plan?" AI handles these well. Complex or emotional cases get to human agents faster because the queue is clear.

Something unexpected: NPS scores have gone up at several companies post-AI implementation. The explanation makes sense once you think about it. Humans who spent 6 hours a day answering "where's my order" were burned out and it showed. Now they spend 6 hours on problems they can actually solve, and customers feel the difference.

The failure case is the confident wrong answer. A customer with a genuinely unusual situation who gets a polite, firm, and incorrect AI response that can't recognize its own limits. Escalation paths need to be obvious. The companies that buried them found it in their churn numbers.

Video: a 10x cost reduction, but only for certain content

A 90-second explainer video with a professional production team: $5,000-$15,000, 3-4 weeks. The same video using Synthesia or HeyGen for the avatar, ElevenLabs for voice, a human for script and review: $200-$500, 3-5 days.

Companies aren't replacing agencies. They're segmenting. Brand campaigns and live events still go to professionals. Tutorial videos, internal training, localized variants go to AI tools. Total output is up; budgets are flat.

Localization is particularly worth noting. Eight-language video used to mean 8 recording sessions, 8 subtitle tracks, 8 edits. With AI voice cloning it's one session. Several companies are now reaching markets they couldn't previously justify the production cost for.

What the companies that made it work have in common

They started with a specific problem. Not "integrate AI into our workflow" but "we spend 200 engineer-hours a month on documentation - can we cut that in half?" A target, a tool, a person responsible.

They kept humans in the loop for consequential outputs. This adds deliberate friction, which annoys the people who got excited about AI removing friction. But the companies that skipped it found out why it exists.

And they measured something. Hours saved, cost per output, ticket resolution time - whatever made sense. Without a number, AI adoption drifts into vague optimism or vague skepticism. With one, you know what's working.

Not a glamorous conclusion. But it's what the cases actually show.