Does AI in video production pay off? The real ROI, no smoke

Every AI production pitch promises the same thing: faster and cheaper. Almost none of them show you the return, and that is where the problem starts, because "efficiency" is not the same as "profitability" and mixing them up is the most expensive way to adopt AI.

This article will not tell you that AI always pays off. It will tell you where it pays off today, where it still doesn't, and how to calculate your real return instead of trusting the demo. If you have to defend a budget to management, this is what you need to do it with data.

The problem: everyone promises efficiency, almost nobody shows the return

The most telling signal of 2026 came from inside the ad industry itself. At Cannes Lions, Publicis Groupe (hardly an anti-AI voice) warned that the technology has widened the gap between what gets promised in pitches and what actually gets delivered, and called for moving the conversation from AI demos to business proof. The line that made the rounds at the festival says it plainly: the hype era is over and proof is the new flex.

For you the practical takeaway is simple: from now on, "we use AI" impresses no serious client. What impresses is showing which task gets faster, by how much, and at what real cost.

Where AI does generate measurable return today

McKinsey places AI's biggest immediate impact at two ends of the pipeline: development and pre-production, and post-production, with early productivity gains it puts at roughly 5%–10% in specific use cases. This is not a promise of revolution, it is a real, compounding improvement. These are the fronts where that return shows up first.

Pre-production and previs: the biggest immediate leap

This is the best effort-to-return ratio available today. Clarifying a pitch deck, generating visual directions, breaking down a script or testing several story openings before the shoot costs a fraction of what it used to, and it cuts the expensive rework: the kind that happens on set. The structural shift McKinsey points to is exactly that, moving work from "we'll fix it in post" to "we solve it in pre". We develop this in AI with human direction.

Variants and personalisation at scale in advertising

Producing dozens of versions of the same piece for different formats, platforms and audiences was, until recently, a bottleneck measured in hours. Today it is where the investment in campaigns is recovered fastest. The detail, how to do it without breaking the brand, is in AI advertising at scale.

Localisation, dubbing and catalogue reactivation

McKinsey identifies dubbing, localisation and library filtering among the applications with the most traction. For anyone with a catalogue or international ambitions it is almost direct return: the same asset, monetised in more markets. We also cover it in AI advertising at scale.

Where the ROI does not appear yet (and why)

This is where most providers go quiet, and where we prefer to be clear.

Fully generated premium. A film piece or a high-end brand piece produced autonomously still does not pay off: quality is unstable and rework eats the saving. McKinsey confirms this at a general level. If the piece carries high reputational risk, the apparent return evaporates the moment you count the correction hours.

Hidden costs. The render is not the cost, it is the tip of the cost. Human supervision, iterating until you reach something approvable, and rework are the part nobody puts on the demo invoice. A figure from a neighbouring field illustrates it well: in architecture, the Chaos and Architizer report found that around 70% of professionals believe AI visualisations still need constant professional supervision, and that reliability remains the top barrier to adoption. That supervision is work, and work costs.

How to calculate your real ROI: the framework we use

To stay out of the smoke, we calculate return with three rules.

Rule	What it measures	Why it matters
Total cost, not render cost	Licences + direction and supervision hours + rework	If you only count the render, the saving is fictional.
Metrics that matter	Time-to-market, number of genuinely useful variants and cost per approved deliverable	A piece the client rejects is not production, it is cost.
The right question	Which specific task in your pipeline gets faster, and by how much	ROI lives in the task, not in the tool.

How we measure it on real projects

On every project we define upfront which phase AI optimises and compare it against the cost of the equivalent traditional process, supervision hours included. If AI doesn't improve the number, we don't use it in that phase, and we tell you. That is the difference between a provider that sells AI and one that sells results.

Want to know which part of your production would pay off with AI and which wouldn't? We calculate it with your numbers, not ours.

Sources

We cite what we claim. These are the references behind the figures in this article.

• McKinsey, state of AI in media and entertainment (productivity gains in pre-production and post-production).
• Cannes Lions 2026, Publicis Groupe on the gap between demos and business results.
• Chaos × Architizer, report on AI in architectural visualisation (professional supervision and reliability).