AI made creative cheap for mid-size brands. Knowing what worked, and who signed off, still isn't.
BetMGM runs a 120-person in-house creative department and uses generative AI, including the Arcana AI suite, to produce imagery and video. One category of content still stays away from the machine: anything carrying betting odds. Victor Gonzalez, the company's senior director of creative, gave Digiday the reason in five words. "You can't hold AI accountable."
I've read a lot of AI-in-advertising coverage this year. That's the most honest sentence in any of it, and it came from a creative director.
BetMGM appeared in the story alongside Wyndham Hotels & Resorts and consumer health company Opella. The three have little in common beyond broad product portfolios and a constant need for digital creative. All three now produce that creative in-house with AI, at a scale that until recently belonged to first movers like Unilever, Coca-Cola and L'Oréal. None of them is small. Wyndham reported $1.43 billion in revenue for 2025. But they sit well below the tier that started this.
The signal goes beyond three companies. Brandtech's Pencil platform counts several clients with more than $10 billion in revenue, yet CEO Will Hanschell told Digiday his fastest-growing segment is companies with less than $1 billion in annual income. In mid-2025, video ad buyers surveyed by IAB projected that generative AI would account for 40 percent of all ads by 2026, with small and mid-tier brands adopting faster than the biggest ones.
So generation is becoming a commodity. What's scarce is clean performance data, honest measurement and somebody with the authority to stop an asset from shipping. Most of the industry is paying for generation and improvising the rest.
Three companies, three different bets
Wyndham started from what it already had: real photography of its hotels. For about a year, its in-house team has used Adora to edit those images into multiple versions for paid and owned Instagram. The team then tracks how each asset performs, combining media-mix modeling with data from Meta's API. Mike Shiwdin, Wyndham's group vice president of loyalty, guest engagement and strategic partnerships, says the team now produces 15 times as many assets as before. He says it has cut the time spent on concepting, production and approval by 75 percent.
Opella went fully in-house. An 85-person team uses generative tools to write strategy briefs and to generate and animate marketing assets, mascots included. Nigel, the owl that fronts the allergy brand Xyzal, is now AI-made. Julien Leterrier, Opella's global head of martech, says output is up twentyfold, past 20,000 pieces of content. Most of it is variants and adaptations of core assets for different languages and markets.
BetMGM spends around $160 million a year on media and moved most carefully. It uses AI for imagery and video but reserves some production for people. According to Gonzalez, the company hasn't yet built enough trust in the machine to hand it everything in a category this tightly regulated.
I wouldn't rank these by output. Wyndham built a system around authentic property imagery and platform feedback. Opella built a localization engine. BetMGM drew a line around the information a regulator would look at first. Each of them answered the same question, knowingly or not: does the production model fit the data they actually have, and what does it cost when the system gets something wrong? That's the benchmark worth copying.
Fifteen times the assets is a capacity number
AI makes output easy to count. Assets, versions, languages, hours saved. Those numbers tell you about capacity and cost. They tell you nothing about whether the extra creative made the business any money.
Wyndham has the strongest performance claim of the three. Shiwdin told Digiday the company's site-visit rate is six times better than its benchmark. He also said Wyndham was increasing its Instagram media spend over the same period, and he didn't share financials. We don't know how the benchmark is defined, how the test was designed or whether spend was controlled.
I'd treat the result as a promising, company-reported signal. Proof would need all three.
Opella's 20,000 pieces call for a different caution. They show production and localization capacity the company didn't have before. Nothing in the public reporting ties that volume to conversion, sales, brand lift, or incremental demand. A content factory can run at full speed and still make far more than any audience needs.
Three separate questions get blurred constantly here. Did production get faster or cheaper? Did a given asset beat the alternative on the channel where it ran? Did the campaign produce an outcome that wouldn't have happened anyway? A team that answers the first and presents it as the third has turned an honest efficiency gain into a growth story it can't defend.
Who keeps what the system learns
The real advantage shows up when generation and measurement run as one loop. IAB's Generative AI Playbook describes dynamic creative optimization in those terms. You launch many variations at once, collect signals like click-through rates or conversions and let the feedback pick the stronger versions. Its implementation roadmap adds clear success metrics, a maintained source of truth and shorter feedback cycles.
Writing that down is easy. Doing it means every asset carries a stable identity and usable metadata. The team knows which elements changed, which audience saw the work, how delivery differed and which outcome was measured. Tests are comparable enough to separate a creative effect from a shift in spend, placement or targeting. Approval decisions stay traceable after the fifth iteration.
With that in place, last cycle's results become next cycle's brief. Without it, the brand produces variants and lets the platform's delivery algorithm sort them. The platform learns which creative works, and the brand gets a dashboard.
This is the part that bothers me most. A company can pay for the production, pay for the media and still end up with the most valuable insight living inside Meta or a vendor's system instead of its own.
What agencies keep
In IAB's 2025 State of Data research, half of agencies said they worry about brands bringing AI capabilities in-house and relying less on partners. That worry is rational. Once a brand team can turn an approved concept into channel and market variants on its own, nobody needs to commission each execution.
Yet Wyndham, Opella and BetMGM all still work with creative and media agencies.
Wyndham's agency of record is Havas. What goes outside is the larger, more traditional work and anything that benefits from an outside perspective, like BetMGM's recent above-the-line campaign with Jon Hamm.
The split is getting clearer. Brands have every reason to own routine production, their own performance data, the taxonomy and final approval. Agencies stay valuable where distance makes the work better. That means the big campaign idea, the challenge to an internal assumption, specialist craft, cultural reading across markets and independent testing of whether a brand's system is producing off-brand or unsafe material. The business is moving from selling capacity to selling judgment. Whatever the conference panels say, that isn't the death of the agency.
Every asset is a review
More output means more to review. A team producing thousands of assets can't run on the informal checks that worked for a handful of campaigns, and the data suggests plenty of teams are trying anyway. In a 2025 IAB and Aymara survey of 125 U.S. advertising executives, 70 percent reported at least one AI-related incident, including hallucinated, biased or off-brand output. Nearly 90 percent said they feel prepared to catch problems before launch. Fewer than 35 percent planned to increase investment in AI governance or brand-integrity oversight over the next year. And 14 percent said nobody at their company owns AI governance at all.
Seven in ten have already had an incident, nine in ten expect to catch the next one, and only about a third plan to spend more to make sure. IAB read that as overconfidence. I'd call it a budget decision somebody will eventually have to explain.
The three companies show what sensible boundaries look like. Opella won't let its AI invent doctors, and it adds AI disclosures in markets that require them, such as the European Union. BetMGM keeps odds out of AI-generated assets. Wyndham is still debating whether generated people belong in images of its hotels and resorts. Each boundary is drawn around what the ad claims or who it depicts, rather than whether AI touched the file somewhere in production.
That logic matches IAB's AI Transparency & Disclosure Framework V2, published in August. It uses risk and materiality to decide when AI involvement should be disclosed in consumer-facing advertising. It also matches the legal baseline. The FTC's position is that ad claims must be truthful, can't be deceptive or unfair and must be backed by evidence, with extra rules for some specialized products and services. Using AI doesn't shift responsibility for a false claim away from the advertiser, and a disclosure label doesn't rescue a misleading ad.
A working control system starts with named owners for model selection, source material, claims, likenesses, localization and release approval. It keeps a record of which tool and which inputs produced each asset, and it preserves the approved version. Higher-risk material goes to reviewers qualified to judge it. IAB's playbook covers much of this in its responsible-AI checklist, from documentation and factual validation to human review, regular audits and a cross-functional governance team. NIST's voluntary Generative AI Profile frames the same work across the AI lifecycle rather than as a single check at launch.
This is the gap I keep writing about. A policy that says "human review required" is easy to write. At 20,000 assets it means nothing unless the workflow forces the review to happen and records that it did.
Where most companies actually are
IAB's 2025 State of Data study, published in March of that year, found that only 30 percent of agencies, brands and publishers had fully integrated AI across the media campaign lifecycle. Half the industry lacked a strategic roadmap. Nearly two-thirds named data quality, data protection and fragmented tools as top barriers. I suspect the integration number has moved since. I doubt the barriers have, and they explain why buying a generation platform is so much easier than building the operation around it.
The sequence that works is unglamorous. Start with low-risk tasks and learn where the tool fails before volume hides the failures. Connect approved source assets, brand rules and workflow permissions, and get legal, compliance, creative and data owners to agree on restricted uses before anything goes live. Put asset metadata and media results on one taxonomy so tests are comparable and channel response doesn't get mistaken for incremental business. Then let controls tighten as claims get more sensitive, while low-risk adaptations move faster. The generation platform is the shortest step on that list.
Accountable by name
What Wyndham, Opella and BetMGM show is that creative abundance has reached companies outside the global-brand elite. That part is settled. From here, the advantage goes to whoever knows which creative to make next, which evidence to trust and where a human decision is still worth the delay.
Gonzalez's five words are the operating principle. The machine can't be held accountable, so someone in the building has to be, by name, for every asset that makes a claim. Without that person, 20,000 pieces of content is just noise with a media budget.