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

72% of companies reported using AI in at least one business function in 2025, according to McKinsey's annual survey. That number is cited constantly in board decks and earnings calls as proof of transformation. It is also nearly meaningless.

"Using AI" includes one marketing manager with a ChatGPT Plus subscription. A developer with GitHub Copilot autocomplete turned on in VS Code. The CEO who used Siri to set a calendar reminder. All of that counts toward 72%. Real adoption - the kind that changes headcount decisions, restructures workflows, or creates measurable cost differences - is a much smaller number inside that figure.

Here is what that smaller number actually looks like.

Marketing: the volume advantage is gone

Two years ago, using AI for content gave companies a real competitive edge. Content volume was the bottleneck. AI removed it. Teams that figured this out early moved faster than competitors still doing everything by hand.

That window has closed. Content at scale is now table stakes. The bar has risen, not dropped, because every competitor is also producing more content and competing for the same attention.

The teams performing well now made a second-order move: using AI to test faster rather than just produce more. One mid-size e-commerce company generates 50 variants of a product description, A/B tests them across audience segments, feeds results back into the next batch. Manually that operation requires a small team and weeks of calendar time. With Jasper handling generation and one person managing the process, it runs on a two-day cycle.

The failure mode is the one companies hit when they remove human review to move faster. Factual errors in AI output are subtle enough to slip through individually but accumulate into a credibility problem at scale. McKinsey found accuracy was the top concern for marketing teams using AI - ahead of cost or compliance. That ranking reflects hard experience, not theoretical worry.

Engineering: the gains are real, but distributed unevenly

The productivity numbers from GitHub and MIT studies get cited relentlessly. What gets cited less is the shape of those gains. One engineering manager described it directly: "My juniors got dramatically better. My seniors got marginally better. The gap between them shrank considerably."

The downstream effect is structural and quiet. Teams that would have hired a fourth developer are staying at three. Headcounts are holding flat through natural attrition rather than growing to match workload. This does not register clearly in employment data, but it is showing up in hiring freezes and flatter org charts at companies that use GitHub Copilot or Cursor seriously.

The failure case is accepting AI-generated code without understanding it. Engineering teams that dropped code review because the AI wrote it found themselves with security vulnerabilities and architectural debt that took longer to fix than the time originally saved. The tools are effective. Removing oversight does not work.

Customer support: consistent results across company sizes

If there is one use case with near-uniform outcomes across industries, it is first-line customer support. The pattern is almost identical wherever you look: deploy AI for initial ticket handling, automate 60-80% of volume, route the rest to human agents who now spend their time on cases that require actual judgment.

The economics are straightforward. Most support queues are dominated by 10-15 question types. Order status. Password resets. Plan changes. AI handles these well. The queue clears. Complex or emotionally difficult cases reach human agents faster because they are not buried under routine volume.

Something unexpected keeps appearing in the data: NPS scores have improved at several companies after AI implementation. The explanation makes sense once you think through it. Agents who spent six hours per day answering the same questions were burned out, and burned-out support agents show it. Free them to spend time on actually complex problems and the quality of those interactions rises.

The failure case is the confident wrong answer. A customer with an unusual situation who receives a polite, firm, and incorrect AI response that does not recognize its own limits or offer a clear escalation path. Companies that buried the escalation option found the consequences in their churn numbers.

Video production: a 10x cost reduction for the right content

A 90-second explainer with a professional production team runs $5,000-$15,000 and takes three to four weeks. The same video using Synthesia or HeyGen for the presenter, ElevenLabs for voice, and one person handling script and review costs $200-$500 and takes three to five days.

Companies are not replacing agencies with AI tools. They are segmenting. Brand campaigns and live events still go to production teams. Tutorial content, internal training, and localized variants go to AI tools. Total video output has increased substantially while budgets have stayed flat.

Localization is where the economics are most striking. Eight-language video used to require eight recording sessions, eight subtitle tracks, eight edit passes. AI voice cloning turns that into one session. Several companies have now reached markets they could not previously justify the production cost for.

What the companies that got results have in common

They started with a specific, measurable problem. Not "integrate AI into our workflow" but "we spend 180 engineer-hours per month writing internal documentation - can we reduce that by half?" A defined target, a specific tool, a named person responsible for the outcome.

They kept humans reviewing consequential outputs. This introduces deliberate friction, which annoyed the people who got excited about AI removing friction. But the organizations that skipped human review found out why it exists.

They measured something. Hours saved per week, cost per piece of content, ticket resolution time, something with a number attached. Without a metric, AI adoption drifts toward vague enthusiasm or vague skepticism. With one, you know within 60 days whether you have a result.

By 2027, at least one major enterprise will publicly walk back a broad AI deployment - not because the tools failed, but because the measurement infrastructure was never in place to show whether they worked. That is a near-certain outcome of the current "implement AI everywhere" approach that skips the measurement step.