- Is it worth bringing creative production in-house with AI?
- What did companies that have already been through this actually do?
- What goes to the internal team and what stays with the agency?
- How much stuff do you need to make for it to be worth it? Three examples
- What tools are used?
- What rules are needed before starting?
- Why is everyone talking about this right now?
- How do you get started, step by step?
- The most common mistakes
- How to tell if it's working?
- So, is it worth it for your company?
- How we see it at SHM
- Frequently Asked Questions
- Bringing creative production in-house with AI only makes sense if you're making lots of similar stuff: variants, formats, languages, channels.
- The companies that did this (Wyndham, Opella, BetMGM) report 15 to 20 times more assets, but none of them fired their agency.
- The agency keeps the idea and the brand; the internal team handles mass production, always with someone giving the green light.
- Before getting started, you need three rules: what AI doesn't generate, who signs off, how results are measured.
Updated October 3, 2026
Five people, fifteen times more advertising materials, approval times reduced by three-quarters. This isn't a lab study: it's what a listed hotel chain did by bringing creative production with AI in-house. The real question for an Italian company is different: is it worth it for you too?
Is it worth bringing creative production in-house with AI?
It's worth it if you produce many similar materials (variations, formats, languages, channels). It's not worth it if you run few campaigns a year, each different from the other.
Generative AI, i.e., software that creates images, videos, and texts from written instructions, speeds up mass production. It doesn't replace the idea. So the point isn't "AI yes or no," but how much of your creative work is repetitive.
Think about the last campaign you ran. How many versions of the same message did you have to prepare? One format for the feed, one for stories, a banner for the website, maybe the English version. Each of these is work that someone does by hand today, and that you pay for by the hour. This table is the quickest way to understand where you stand.
| Your situation | Is it worth it? | why |
|---|---|---|
| Large catalog, multiple languages or markets | Yes | Every product needs to be adapted for formats, languages, and channels: it's repetitive work |
| Campaigns with many tests (ads on Meta and Google) | Yes | More variations to compare = higher chance of finding the one that performs |
| Few campaigns a year, each one different | No, for now | The value lies in the idea, not the volume |
| No one in-house who can check the materials | No, for now | Without human approval, brand risk outweighs the savings |
If you identify with the first two lines, it's worth reading the rest carefully. If you're in the last two, the investment probably won't pay off today, but it's still worth knowing how the market is moving, because the situation changes quickly.
What did companies that have already been through this actually do?
They created a small internal team that uses AI for mass production, kept the agency for strategy, and wrote down what AI should not do.
The three cases shared by Digiday on October 2, 2026, they are mid-to-high-end companies, not giants with billions in advertising budgets. This is precisely what makes them interesting: until a year ago, systems of this type were the stuff of multinationals.
| Company | sector | In-house team | Stated result | Rule they set for themselves |
|---|---|---|---|---|
| Wyndham Hotels & Resorts | Hotel chain | 5 people | 15 times more materials, production and approval times −75% | Still being debated whether to generate people in hotels |
| Opella (ex Sanofi) | Health products | 85 people | 20 times more materials, over 20,000 contents | Forbidden to generate doctors; AI disclosure in EU markets |
| BetMGM | Online betting | 120-person creative department | Use AI in production, not everywhere | Quotas and regulated content made only by humans |
The declared numbers are strong. Wyndham, with a team of five, produces fifteen times the previous materials. Opella, born from Sanofi in 2025, has built a dedicated team of 85 people and has surpassed 20,000 contents, because it needs to adapt the same products for many markets and many languages.
But the multiplier alone says little. The most important data for decision-makers is time: in Wyndham's case, production and approval take a quarter of the previous time. This means being able to launch a campaign, or correct it, in days instead of weeks.
The most interesting data, however, is neither of those. It's that none of the three companies left the agency, and all three set precise limits before starting. We'll talk about it later.
What goes to the internal team and what stays with the agency?
Repetitive production goes to the internal team; the agency keeps the idea, the brand, and the outside perspective.
Wyndham, for example, continues to entrust the largest and most traditional creative briefs to its long-standing agency, and uses AI for variants intended for social media. This division of labor also makes sense for a medium-sized Italian company: the person who conceives the campaign doesn't necessarily have to be the one who produces every single version of it.
The diagram below summarizes the flow these companies have adopted. People decide, AI multiplies, data chooses what to keep.
- 1 · People and agency Brief and idea Goal, message, creative concept.
- 2 · People and agency Approved master A "master" image or video that respects the brand.
- 3 · Internal team AI generates the variations Formats, languages, channels, products: dozens of versions from the same master.
- 4 · Internal team Human oversight Brand rules, no regulated content, a human approver.
- 5 · Internal team Publishing and testing The variations go online and compete against each other.
- 6 · Data Measurement and choice Keep the winning variants, drop the rest, and go back to step 3.
| To the internal team with AI | To the agency (or strategist) |
|---|---|
| Variants of the same ad | Creative concept and campaign idea |
| Format adaptations (stories, posts, banners) | Brand campaigns and major launches |
| Versions in other languages and for other markets | Brand guidelines and tone of voice |
| Quick tests on headlines, images, colors | Results review and budget decisions |
Translated into budget: if today you also pay the agency for the tenth version of the same banner, that expense can be shifted to work that brings more value, such as the idea for the next campaign or the analysis of results. Be careful though, because some items they don't disappear, they shift : tool licenses, training, reviewer time.
How much stuff do you need to make for it to be worth it? Three examples
The calculation is simple: products (or services) × formats × languages × channels. If the result per campaign is tens or hundreds of items, in-house AI makes sense. If it's a handful, it doesn't.
We took three typical situations of the companies we follow and did the math. The numbers are examples, but the reasoning is what we actually do when a client asks us if it's worth it.
| Scenario | The bill | What we would do |
|---|---|---|
| E-commerce with 40 products on promotion, site in 3 languages, Meta ads | 40 products × 3 formats × 3 languages = 360 pieces per campaign | Internal system with AI : without it, those materials don't get made or cost too much |
| Retail network or studios with local campaigns per city | 10 locations × 2 formats × 4 releases = 80 pieces | Middle ground : master made by pros, local variants with AI |
| B2B manufacturing company, 2 trade shows a year, catalog in 2 languages | 2 trade shows × 6 materials = 12 pieces, each carefully crafted | Agency or designer , AI only as help (drafts, translations to be reviewed) |
Note that the criterion is not the size of the company but the repetitiveness. A small company with a large catalog and three languages may need an internal system more than a large company that runs two ads a year. And the cost changes over time: if you start testing more ads on Meta or Google, the variants multiply even if the catalog remains the same.
What tools are used?
Two paths: specialized platforms that manage the entire process, or a workflow built on generic tools. The first is faster, the second more flexible.
Source companies have chosen dedicated platforms. Wyndham works with Adora, an AI company that also helps it understand which materials perform best. BetMGM uses Arcana AI, a suite of generative tools. Other brands use Pencil, a platform from the Brandtech group. These are tools designed for high-volume producers who want integrated brand rules and approvals.
For a medium-sized Italian company, it often makes more sense to start with generic tools (image and video generators, editors with AI functions, text templates) connected by a simple workflow: a master, a rule library, an approval step. It costs less, adapts better, and you can switch to a dedicated platform when volumes justify it.
| Specialized platform | Workflow with general tools | |
|---|---|---|
| Examples | Pencil, Adora, Arcana AI | Image and video generators, editors with AI features, text models |
| For whom | High volumes, many markets | Medium volume, first tests |
| Brand rules and approvals | Already integrated | To set up (rule library + approval step) |
| Flexibility | Lower | Higher |
| When to switch to the other | — | When volumes grow and approvals become a bottleneck |
What rules are needed before starting?
Three rules, decided before the first asset: what AI cannot generate, who approves before publication, how to measure what works.
The source companies learned them in the field:
- Regulated content stays human. BetMGM does not use AI for betting odds: an error there has legal consequences. As their manager said, there is not yet trust in the machine for these contents. In Italy, the same applies to prices, sales conditions, and claims about health and safety.
- No “fake” people where they matter. Opella prohibits generated images of doctors or healthcare professionals and, in European markets, adds an indication that the content is made with AI. Wyndham is still discussing whether to generate people within its hotels.
- Brand consistency. A single reference is needed (colors, tone, what's forbidden) and a person to approve. Without them, you produce a lot and communicate inconsistently.
Then there is the issue of platform reaction: some are already penalizing AI-generated content , and the audience preference it's not a given. That's also why human oversight isn't just a bureaucratic step: it's what protects the result.
Why is everyone talking about this right now?
Because it costs less, you test more, and you need more content. And because a growing share of advertising is already done this way.
Four numbers give an idea of the moment:
The latest data is the most important, and should be read together with the others. If creative quality accounts for about half of an advertisement's effectiveness, as Kantar claims, then producing more is only useful if it helps you find the best version. More tests mean more variants, and that's precisely where manual production becomes the bottleneck.
The market is moving in this direction. According to the latest report from IAB, the American digital advertising association, one-third of video advertising materials in 2026 will be made with generative AI, and almost two out of three buyers are already using it.
And the most common use is precisely what we're discussing in this article: creating versions of the same asset for different audiences.
Meanwhile, distribution is also changing: think about ads on ChatGPT with click-to-chat . More channels means, once again, more formats to produce.
How do you get started, step by step?
With a small, measurable test: a single campaign, written rules, and an honest comparison with how you work today.
- What you produce today matters. Take the campaigns from the last quarter and count the assets: how many, in what formats, how many languages, how much time from brief to publication. This is the basis for understanding if there's room for improvement.
- Write down the rules. Colors, tone of voice, what AI cannot generate, who approves. Without this step, the rest falls apart.
- Choose a pilot campaign. Better a campaign with many variants and a clear goal, for example Meta ads for a product line.
- Produce the master carefully, then the variants with AI. The master is made by humans (in-house or agency). The variants are generated by the internal team, and every single one goes through approval.
- Measure and compare. Time, cost per asset, variation results. Then you decide whether to scale, tweak, or stop.
The most common mistakes
Starting with tools instead of rules, and measuring how much is produced instead of how much it pays off.
- Buying the platform before knowing what is needed. You pick the tool after counting the assets and writing the rules, not before.
- Skipping approval "just because it's a variation". Errors slip through precisely on the variants: an old price, a wrong color, an off-tone phrase.
- Generating people where real people are needed. Testimonials, staff, professionals: here AI does more harm than it saves.
- Measure volume. Producing ten times as much isn't a result. Finding the version that brings in the most leads sooner is.
- Firing creativity. Without a strong underlying idea, AI just multiplies a weak message.
How to tell if it's working?
With a few indicators, measured before and after: time from brief to publication, cost per asset, variants tested, and campaign results.
| Indicator | What it measures | Sign that it works |
|---|---|---|
| Time from brief to launch | Process speed | It shortens without increasing errors |
| Cost per asset | Hours and licenses divided by assets produced | Goes down, with equal quality |
| Variants tested per campaign | How much you manage to experiment | Increases |
| Materials rejected during approval | Quality of what comes out of AI | Stays low or drops over time |
| Results (clicks, leads, sales) | If the variants actually perform | They are improving compared to previous campaigns |
The comparison should be made on the same campaign or similar campaigns, and for a few weeks. A good sign isn't just "we spent less": it's "we found the ad that works sooner, and we were able to adapt it to all channels."
So, is it worth it for your company?
It is worth it if each campaign requires dozens of variations, you already have someone in-house who can learn to use the tools, and you can link every piece of material to the results data (clicks, leads, sales).
A middle ground exists: have the system built by those who already know how to do it (brand rules, templates, approval workflow) and then manage it internally, instead of starting from scratch.
Not yet if run few campaigns a year, each different, or if you don't have anyone who can approve materials continuously. In that case, AI produces more stuff, not more results.
How we see it at SHM
For us, technology is useful when it frees up time to do the things that matter better, not when it becomes a way to produce more without asking if it works. A creative production system without connected performance data is a machine running in neutral.
Because of this, when we set it up for a company, we start with the rules: a Digital brand kit that AI must follow, a workflow of integrating AI into processes with who approves what, and a Training for the team that's going to use it. Campaigns are measured afterwards, and stuff that doesn't perform gets dropped. More insights in the hub Marketing and ADV performance .
Frequently Asked Questions
It depends mainly on the volume and who manages it: tool licenses, team training, and approval time. The cost of using models has decreased by 41% compared to the March 2026 peak, but the heaviest item remains people's time.
No. None of the companies mentioned have done so: the agency keeps the idea, brand, and important campaigns, while the internal team handles mass production.
With rules written before starting (colors, tone, what's forbidden), an approved master to generate variants from, and a person to check every output.
Yes, if they have a lot of repetitive materials: catalogs in multiple languages, product sheets, LinkedIn campaigns with variations. For a few releases a year, the advantage is low.
Prices, sales conditions, regulated content, and images of professionals (doctors, technicians) that look real. These are the limits that the same companies using AI on a large scale have set for themselves.
Time from brief to publication, cost per asset, number of variations tested, assets rejected during approval, and, above all, campaign results: clicks, leads, sales.
The right question isn't how much you can make with AI, but how much of your output is just repetitive, and who checks that it actually works.
Related articles
Discover more articles exploring similar topics, selected to offer you a more complete and stimulating perspective. Each piece of content is carefully chosen to enrich your experience.
