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, up from 55% the year before. That sounds transformative.

Here's the caveat buried in the methodology: "using AI" includes one marketing manager who has a ChatGPT Plus subscription and generates email subject lines with it. It includes a developer who has GitHub Copilot autocomplete turned on. It includes the CEO who asked Siri to set a reminder.

Real AI adoption at the organizational level looks different. It's messier, more specific, and more interesting. Here's what we've actually seen working, and where companies have gotten burned.

Marketing: The First Mover Advantage Is Already Gone

Two years ago, teams that adopted AI writing tools had a real competitive edge. Content volume was the bottleneck, and AI dissolved it. Today, that advantage has normalized. Everyone's using AI for first drafts. The content quality bar has raised, not lowered, because more of it exists.

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

Where it fails: AI-generated content that isn't reviewed. More than one company has published factual errors at scale because someone turned off the human review step to move faster. The errors that make it through tend to be subtle enough that individual readers don't catch them, but they accumulate into a credibility problem over time. The McKinsey survey found that hallucination and accuracy were the top concerns for marketing teams using AI, ahead of even cost or compliance.

Engineering: The Productivity Gains Are Real But Uneven

The developer productivity data is starting to come in, and it's interesting. GitHub's own research found Copilot users completed tasks 55% faster. A more recent study from MIT found productivity gains of 40% on coding tasks among developers who adopted AI assistants. These numbers get cited a lot.

What gets cited less: the gains are heavily skewed toward junior developers and routine tasks. Senior developers working on complex, novel problems see smaller improvements. One engineering manager at a mid-size startup described it to us as: "My junior developers got dramatically better. My senior developers got marginally better. The gap between them shrank."

That has organizational implications. Teams have been able to maintain output with smaller headcounts, mostly through natural attrition rather than layoffs. A 4-person team that needed to hire becomes a 3-person team that doesn't. This is happening quietly across a lot of companies but doesn't show up in headline numbers about AI-driven employment changes.

The failure mode: accepting AI-generated code without understanding it. Companies that skipped code review processes because "the AI wrote it" found themselves with security vulnerabilities and architectural debt that took longer to fix than the time they saved. The productivity tools work; the oversight can't be dropped.

Customer Support: Automation of the Repetitive, Escalation of the Complex

This is probably the most mature and consistent AI use case across industries. The pattern is almost identical wherever you look: deploy AI for first-line support, handle 60-80% of tickets automatically, route the rest to humans who now deal almost exclusively with edge cases and emotionally complex situations.

For most support teams, 70-80% of incoming tickets are variations of the same 10-15 questions. "Where's my order?" "How do I reset my password?" "Can I change my subscription?" AI handles these reliably. The questions that require empathy, judgment, or genuine problem-solving get to human agents faster because the agent's queue is cleared of noise.

Net promoter scores have actually gone up at several companies we've tracked post-AI implementation, which seems counterintuitive. The explanation: humans who previously spent 6 hours/day answering "where's my order" were burned out. Now they spend 6 hours on problems they can actually solve well, and that quality shows.

The failure case is the edge situation that AI handles confidently and wrongly. A customer in a genuinely unusual situation who gets a polite, firm, and incorrect answer from an AI that can't recognize it doesn't know what to do. Escalation paths need to be obvious and easy to reach. The companies that hid them behind AI layers found it in their churn numbers later.

Video Production: The 10x Cost Reduction

Enterprise video production costs have collapsed for certain content categories. Not for everything -a brand campaign or live event still needs human production talent. But for explainer videos, product demos, training content, and localized versions of existing videos, the math has completely changed.

A standard 90-second explainer video with a human production team runs $5,000-$15,000 and takes 3-4 weeks from brief to delivery. The same video using Synthesia or HeyGen for the avatar, ElevenLabs for voice, and a human for scripting and review runs $200-$500 and takes 3-5 days. That's a 10-30x cost reduction depending on the baseline.

The adoption pattern: companies aren't replacing their production agencies. They're segmenting their content. High-visibility, brand-critical content still goes to professional production. High-volume, functional content (tutorial videos, internal training, localized variants) goes to AI tools. Total video output has increased substantially while production budgets have stayed flat or decreased.

Localization is particularly notable. Producing a video in 8 languages used to require 8 recording sessions, 8 sets of subtitles, and 8 video edits. With AI voice cloning and avatar tools, it's one session plus generation time. Several global companies are using this to reach markets they previously couldn't justify the production cost for.

What the Success Stories Have in Common

After looking at enough of these cases, a few things stand out about the companies making AI work versus the ones that tried and retreated.

The winners started with a specific, measurable problem. Not "integrate AI into our workflow" but "we spend 200 engineer-hours per month on documentation, let's see if we can cut that in half." Specific target, specific tool, specific person responsible.

The winners kept humans in the loop for consequential outputs. Automation handled the volume, humans reviewed the stuff that mattered. This adds friction deliberately, which is uncomfortable for the people who got excited about AI removing friction. But the companies that skipped review found out the hard way why it exists.

And the winners measured something. Productivity hours saved, cost per unit of output, ticket resolution time, whatever made sense for the use case. Without measurement, AI adoption tends to drift into vague optimism or vague skepticism. With it, you know what's actually working.

That's not a particularly exciting conclusion, but it's what the cases actually show.