The Maze: AI is already woven into content production, but not evenly. An NP Digital survey of 100 companies puts AI integration at 91.4% for images, 84.1% for text and 82.8% for audio. Video sits at 59.7%. The average is 79.5%, yet that single number hides a 31.7-point gap between the easiest format to automate and the hardest. The useful question is no longer whether a marketing team uses AI. It is where AI sits in each workflow, what it is allowed to decide, and where a human still has to own the result.
Images have become the default AI entry point. At 91.4%, image workflows lead video by 31.7 percentage points and stand 11.9 points above the four-format average. That does not mean nine in ten finished images are machine-made. “Integration” can cover anything from ideation and background removal to resizing, retouching and full generation. Still, the direction is clear: visual production now has the lowest operational barrier to AI assistance. A team can test more variants, localize assets faster and compress routine studio work without rebuilding its entire publishing stack.
Text and audio have converged into the same operating band. Text reaches 84.1%; audio reaches 82.8%. A 1.3-point difference is too small to carry a grand story, especially with rounded data and limited methodology. The stronger signal is that both formats sit above 80%. Their workflows are structured, editable and easy to review in stages. Drafts can be checked against a brief. Transcripts can be corrected. Voice can be regenerated. That makes human oversight cheaper than in video, where one error may infect script, footage, timing, lip movement, sound and brand presentation at once.
Video is the constraint, not the laggard to shame. At 59.7%, video remains widely touched by AI, but it is 19.8 points below the average. Production has more moving parts, higher rendering costs and more visible failure modes. A strange hand in an image is embarrassing. A strange hand moving for 30 seconds beside a mismatched voice becomes the whole ad. Commerce teams also need product accuracy: shape, color, fit, motion and usage cannot drift without creating return risk. The lower rate may therefore reflect rational quality control rather than weak ambition.
Provenance will expose process, not rescue mediocre work. Neil Patel connects the adoption data to SynthID and the danger of using AI merely to turn two hours of thinking into ten minutes of output. The distinction matters. Google DeepMind describes SynthID as an invisible watermark for supported generated media, while OpenAI combines supported watermarks with Content Credentials. These signals can help establish origin. They do not tell a customer whether a claim is true, a concept is original or a campaign sells. Detection makes lazy production easier to identify; it does not make good judgment automatic.
Why it matters: A single “AI adoption” KPI is now almost useless. The operating model has to be format-specific. Images may need fast brand and rights checks. Text needs evidence and voice review. Audio needs consent, pronunciation and disclosure controls. Video needs the deepest product-truth and continuity QA. The teams that win will not be those with the highest automation percentage. They will know which steps machines can accelerate, which decisions humans must own, and which errors customers will punish. Speed is abundant. Editorial responsibility is still scarce.
