Five specialists. Thirty days. More than 70,000 video clips.

Those were the production numbers Coca-Cola shared for the two AI-made films in its 2025 “Holidays Are Coming” campaign, produced with the AI studios Silverside and Secret Level and released on YouTube on November 3, 2025, according to NBC. A generation ago, a holiday spot of that scale meant a director, a crew, a set, animal wranglers or an animation house, and a budget with several zeroes on the end.

The comments section was brutal anyway. “It’s immediately forgettable.” “Soulless.” The behind-the-scenes video drew “Jeez, this is dystopian.”

Then the ad went into formal testing and scored 5.9 stars on System1’s scale, the maximum possible.

Both are true, and the gap between them is where the industry now has to make its decisions.

The cost curve has already moved

Forrester data cited by Fortune in January 2026 found that three in four ad industry executives used generative AI in 2025, up from 61% the year before. At WPP, more than 85,000 of roughly 108,000 employees were using the company’s WPP Open platform every month, up from 30,000 in February 2024. WPP says teams of four save 14 hours each, roughly 90 days of capacity a year, according to Fortune’s interview with chief technology officer Stephan Pretorius. Employees had built more than 75,000 AI agents by the end of 2025.

None of this is free. The same Forrester figures showed the cost of these capabilities, mostly absorbed by agencies rather than passed to clients, rose 83% in 2025.

Making a finished-looking piece of film, a set of variations, a localized cut or a hundred social formats now takes a fraction of the people and time it used to. That part of the industry’s cost base is being rebuilt.

Testing says AI ads can work

In November 2025, Tom Roach and System1’s Andrew Tindall, writing in Marketing Week, tested 18 AI-assisted video ads gathered through an open call. They averaged 3.4 stars on System1’s scale, against 2.3 across the firm’s database of 123,000 video ads. UK entries averaged 4.6. The 2025 Coca-Cola film hit the 5.9 ceiling, with happiness making up 71% of its emotional response against a category average of 36%.

The authors were careful about the limits. Eighteen ads is a small sample, and it skewed toward work already getting industry attention. Seven of the 18 came from companies within a single group. But the headline holds. When the idea is strong and the execution is controlled, audiences in a testing panel don’t punish an ad for how it was made.

The Super Bowl said something different

Three months later, the biggest stage in advertising produced a different result.

System1 tested the Super Bowl LX ads in February 2026. The 18% that featured AI messaging averaged 2.1 stars. The overall Super Bowl average was 2.7, itself down from 3.0 the year before. Of the top 10 ads, only one, Ring’s “Search Party,” had an AI theme. The top of the table belonged to the NFL, Budweiser and Universal Orlando, with work System1 described as “heartfelt and human.”

The most closely watched case was Svedka. eMarketer described its spot, dancing robots mixing a drink from the vodka brand’s products, as the first Super Bowl ad created primarily with AI. Reaction was mixed, and some viewers and critics called it “uncanny.”

Jon Evans of System1 was blunt about the weak ones: “Most of these ads lacked storytelling, a character to root for and clarity on who the ad was for.”

Public reaction tracked with the scores. Meltwater’s social listening found that nearly half of online conversation about AI ads during the game skewed negative, with viewers calling many spots “uninspired” or “low-quality.” Meltwater’s chief strategy officer, Alexandra Bjertnæs, summed up the mood: “Audiences will tolerate [AI] as a tool, but they won’t embrace it as a shortcut for creativity.”

eMarketer, citing survey data, reported that more than 30% of consumers across age groups said they were less likely to choose a brand that uses AI in an ad campaign. Only 12% said they were more likely.

Reconciling the two results

Both the November ads and the February ads used the same kind of technology. What separated them was the job the AI was given.

In the better-performing work, AI was a production method. The idea came first: a familiar brand world, a clear emotional target, a story with a beginning and an end. AI made it faster and cheaper to build. Audiences responded to the idea and mostly didn’t care how it was rendered.

In much of the Super Bowl work, AI was the subject or the shortcut. The ad was about AI, or the ad existed because AI made it possible to produce something without first deciding what it was for. Evans’s list of missing pieces (story, character, a clear audience) is a list of decisions that come before production. No tool makes them for you.

Execution is getting cheap. Judgment isn’t.

What gets more valuable

When the cost of making an ad drops toward zero, three things don’t drop with it.

The idea. A brand still needs a reason for someone to care. Generating 70,000 clips is a production achievement. Picking the right few shots, and knowing why they’re right, is the creative one. Much of the most durable advertising of the last few decades came from teams that built one strong idea out over years. One longtime creative director, looking back on two decades of work for some of the world’s best-known brands, has talked about keeping the creative work simple as a discipline in its own right. That discipline matters more when the tools can produce endless complicated versions.

The brand world. Coca-Cola’s AI films likely had help in testing from decades of established assets: the red trucks, the polar bears, the holiday music. AI is very good at extending a brand world that already exists. It’s much worse at inventing one. Brands with strong, consistent codes have something to feed the machine. Brands without them get generic output. The System1 Super Bowl data makes the cost visible: average brand recognition fell to 78%, meaning more than one in five viewers couldn’t name the advertiser. That was the lowest figure in seven years of System1’s Super Bowl testing.

Taste. Pretorius, a technologist, made the same point from the other side. WPP’s goal, he told Fortune, is “making sure that humans are in control of the output and that they evaluate and apply taste and judgment,” so they don’t become “a passive passenger in the process.” That’s a CTO arguing that the most valuable person in the workflow is the one deciding what’s good.

Alongside the human panel in the Marketing Week study, the authors ran the same 18 ads through an AI evaluation tool from Jellyfish. Every ad scored between 7 and 9 out of 10. The model’s rankings lined up with the human ones in broad terms, but the correlation was 0.56, and a range that narrow doesn’t help much when the job is choosing between two good options. Machines can sort the weak from the strong. The close calls, which are the ones that decide whether a campaign lands, still need a person willing to make them.

What this means for agencies and clients

The agency model has priced creative and production together for a long time. That bundle is coming apart.

Production is becoming a volume business with falling unit costs. WPP’s investment in platforms and agents points that way. Clients will increasingly expect production prices to fall, and the Forrester cost data suggests agencies are already absorbing much of the expense.

The thinking is moving the other way. If anyone can make a polished ad, the scarce thing is the person who can tell you which ad to make and defend that choice in front of a nervous client. That person was always the most important one in the building. The difference now is that the work around them is cheap enough to make it obvious.

Try a simple test on any AI-produced campaign. Strip out the production and describe the idea in one sentence. If the sentence is strong, AI will probably make the ad faster and cheaper without hurting it. If there’s no sentence, AI will just help you make a forgettable ad very efficiently.

The industry has seen this before, on a smaller scale. Campaigns like Chrysler’s “Born of Fire” ran on one clear idea, long before any of today’s tools existed. The production methods changed every decade. The need for someone to have the idea never did.

Seventy thousand clips is a production number. The audience only ever sees the idea.

Internal notes (remove before publishing)

Backlinks placed (2), link-only, Mark’s name does not appear: podcast on the anchor “keeping the creative work simple” (matches the episode description, “his approach to keeping creative work simple”); markfitzloff.com on the anchor “Chrysler’s ‘Born of Fire’” (listed in the site’s work section; chosen so the Nike/Coca-Cola/Old Spice trio from Article 1 isn’t repeated). Old Spice and WK are not mentioned.

Sources behind the reporting (attributed in prose, not linked):

• NBC (via NBC Bay Area), Nov 4, 2025, Coca-Cola AI holiday ads, production figures and comments. https://www.nbcbayarea.com/news/national-international/coca-cola-ai-generated-holiday-ad-backlash/3975523/
• Marketing Week, Tom Roach with Andrew Tindall, Nov 18, 2025, “You can’t dismiss AI ads as slop when they’re winning in testing.” https://www.marketingweek.com/dismiss-ai-ads-winning-creative-effectiveness/
• Fortune, Jan 21, 2026, WPP CTO Stephan Pretorius on AI (Forrester figures, WPP Open adoption). https://www.fortune.com/2026/01/21/wpps-cto-says-ai-is-reshaping-advertising-but-creative-judgment-needs-to-remain-in-human-hands
• System1 via Business Wire, Feb 9, 2026, Super Bowl LX ad testing. https://stocks.observer-reporter.com/observerreporter/article/bizwire-2026-2-9-ai-killed-big-game-ad-effectiveness-human-centric-stories-won-the-night
• Meltwater / PR News, Feb 8, 2026, social sentiment on Super Bowl AI ads. https://www.meltwater.com/en/news/ai-ads-face-skepticism
• eMarketer, Feb 9, 2026, “AI took center stage at Super Bowl LX” (Svedka, consumer survey figures). https://www.emarketer.com/content/ai-took-center-stage-super-bowl-lx

Analysis note: the section “Reconciling the two results” is the article’s own interpretation of why AI ads tested well in November and poorly at the Super Bowl. Neither System1 source states that explanation directly.