- Wonga used AI video generation to extend its 2026 ‘Yoh! to Yebo’ campaign across TV, VOD, digital and out-of-home.
- Google Veo outperformed Sora and Grok Imagine, delivering high-quality 4K video and scene extension capabilities.
- AI still needs human oversight, with more than 400 video outputs rejected before 10 were campaign-ready.
- Human creativity drives effective AI advertising, from prompting and editing to local humour and cultural nuance.
- AI enhances, not replaces, creative teams, enabling faster, more scalable content production while keeping people in control. *
With the launch of the 2026 marketing campaign that illustrates how Wonga’s short-term loans can help South Africans move from ‘Yoh! to Yebo’, our team thought creatively about how best to put these stories in front of audiences across the country and make them relatable.
The Wonga brand story has a rich history and is an exciting opportunity to play on the sense of fun and comedy that South Africans tend to appreciate so dearly. Sensing this, our team felt it was time to take the brand back above-the-line.
Our return began with two TV adverts produced with our talented agency partners that centred on the moments life throws our way, from the relatable to the absurd – whether that’s falling into a pothole or losing a smartphone.
Natural fit
For the silver screen, this seemed the natural fit – but we quickly realised that across social and video on demand (VOD), there was a smaller, nimbler, but just as rich canvas to create short-form video assets that captured the moments our TV adverts couldn’t.
We settled on the idea of trialing ten AI-created video ads to tie together our efforts across streaming and linear TV, out-of-home, and static ad assets to see how far we could push the envelope.
The idea of an AI-created advertisement is hardly novel; however, we saw an opportunity to use AI platforms to put together the scenes and ideas that time, and our budget, would not typically allow in a way that amplified the message of our existing ad inventory.
With the goal of keeping our execution cost-effective, scalable, and accessible to the ideas we developed in concept, we explored several AI models in ‘off-the-shelf’ environments; that included Sora (OpenAI), Grok Imagine (X), and finally, Veo (from Google Gemini).
Striking paradox
We selected Veo for the breadth of its output and availability to serve high-quality visuals up to 4K. We used the model through Google’s proprietary Flow platform to extend or remaster scenes without immediately heading to the editing booth.
What emerged was a striking paradox; while it’s impressive to reflect on how quickly video generation capabilities have developed, it’s equally disappointing to reflect on how far they have yet to go. In generating 10 successful files, we discarded over 400 outputs.
As it goes for text generation, AI tooling is only as effective as the information it’s been trained on. Developing funny, and slightly absurd scripts (such as one abandoned concept where an unfortunate male loses an AirPod to a urinal) is difficult where few reference scenes of that nature exist.
This can be worked around by providing reference frames generated from other platforms, but built-in modesty restrictions and content filters prevent AI platforms from going to the frontiers where TV adverts may play.
Quality prompting
Some of those abandoned 400 outputs were short-lived; while AI can typically generate eight-second scenes, certain platforms (such as Veo) are capable of extending them gradually, up to a minute.
However, this is where it becomes make or break: models build up context across multiple prompts, and the many attempts needed to extend a scene meaningfully often dissolve into a muddle.
Prodding an AI to deliver quality output does require human intelligence to judge it ‘good enough’ – examples of the edge cases we worked through included the decision to eliminate dialogue for its unreliable quality or abandoning great output because it featured international (and very identifiable) trademark licensed brand products.
Quality prompting on behalf of the director is also key – having a vision is one aspect, but recklessly worded input can turn a scene about a flooded apartment into a world where physics is abandoned and household objects, rather than floating on the water, float through the air in scenes out of Star Wars.
Consistency an issue
Another aspect of video generation that AI hasn’t yet cracked is consistency from one scene to the next; characters, locations and other visual elements were rarely consistent across generations.
Some common challenges we encountered were that character faces would change, outfits would swap, and new props would be added unprompted – which pushed us towards single-scene outputs rather than trying to extend scenes further.
While Veo (and other AI tools) are capable of generating some worthwhile concepts, they still require the same editing expertise as a fully-fledged production to come to life.
One key example is the soundstage – adding the right foley to a scene could transform a viewer’s understanding of the message from a chuckle into a laugh; and this is where AI has yet to deliver without combining multiple platforms – and, ultimately, human hands – to land the message.
Local nuance
The rest – local nuance, timing, and fit to platform – remain in the hands of the “director”. So much of what makes humour land here is distinctly local – and bringing that to life with care and taste requires a set of eagle eyes and ears to ensure that concepts that poke fun aren’t taken too far or lose their meaning.
Is AI ready to take the stage and develop full-blown TV adverts that replace human hands? Decidedly not. Brands looking to invest in creative AI capabilities should match that investment with capable teams or agencies to enhance their executions, rather than replace them.
That being said, the good news for the Wonga brand – and many other brands out there – is that there is now aa
SIDEBAR
Four things to know before you make creative with AI:
- Plan to throw most of it away. We kept ten clips from more than 400. Generation is cheap enough to fail fast – but only if your process expects a brutal hit rate, not a tidy one.
- The funniest ideas are the hardest to make. Models can only draw on what they’ve seen, and content filters steer them away from the edgy, absurd territory great advertising lives in. The more original your concept, the harder you’ll fight the tool for it.
- Think in scenes, not stories. AI still loses the thread across cuts – faces, outfits and props drift between generations. Build in short, single scene beats and stitch them together yourself, rather than asking the model to carry a whole narrative.
- Buy the tool, keep the talent. AI compresses the distance from idea to frame like never before, but it can’t judge what’s ‘good enough’, land local nuance, or make an audience feel. Treat it as an accelerator for your team – never a replacement for it.
*Summary created by AI.
Bryan Smith is Wonga’s digital media manager, a fintech marketer of more than ten years who bridges creative storytelling and performance marketing, and still believes the best growth starts with a story worth telling.
