Problem Solved: Finding the ‘Fun in a Framework’ with Snickers’ AI-Powered Banter

Mars Wrigley challenged us to grab the attention of football fans for Snickers in a way they wouldn’t have seen before. We quickly realized that to get real fans talking, we needed to speak to them where they’re chatting authentically – in closed social groups like WhatsApp. We also realized they’d quickly dismiss any brand trying to speak to them here. But might just listen to one of the world’s greatest coaches.  

While we initially thought the idea was ambitious – a funny, clever, AI script writing tool that then generates cloned audio and video – this part of the tech solution was (relatively) simple. We had a working prototype of each element within weeks, and all that was left was to fine-tune the output.

The hardest bit was making it both safe, and funny. The more open the tool was, the funnier the scripts and the outputs were, but it also opened us up to risk. But locking it down to be fully safe and pre-prepared killed the magic. We needed a middle ground.

Around the time of the brief, lots of celebrities were being negatively deep faked in the news. Which made us think about the potential for what an authorized deep fake of footballing talent could do – especially if you have it call out and respond to the dumb things people have done.

We ended up using a blend of human and artificial intelligence to create a system that could get the balance right. They say that ‘explaining a joke ruins it’ but that’s what we needed to do in order to teach our AI to be ‘funny’. So we tried to dissect what made José’s tone and style of delivery funny – and through a series of iterations and scripts, we created a language model which could think and write a bit like José, but within the confines of the Snickers brand idea.

We also had to sacrifice some of the funniness of the idea and the earlier prototypes – because to make it safe, we needed it to be a bit more formulaic. Eventually we found our ‘fun in a framework’ that would work.

To make the tool safe, we needed a ‘red-team’ of up to 30 people, trying to get the tool to say offensive and inappropriate stuff. From collating long lists of famous British criminals and serial killers, to coming up with creative ways to try and game the system to create something offensive – the team had to explore some unexpected places to test the edges of the tool.

We ended up with eight different inter-related AI models that would test, refine, and create. This could catch and change offensive or unclear inputs, and then it would generate a script, and then re-check the output against another safety layer.

The tool ended up blending together language model capabilities from Open AI, with voice generation from ElevenLabs and video cloning from Synch Labs, with custom safety layers and another tool called Zefr for brand safety. But the manual tuning and testing was where the magic happened.

We even spent some time ‘auditioning the LLMs’ for the main script writing – using the same prompts and workflows with some different LLM options – and seeing a spectrum of interpretations and tones.

While the initial prototype took weeks, the testing and refining took months – and required a lot of careful and close collaboration between legal teams, José Mourinho’s agents, and internal teams. Fine-tuning an inherently unpredictable and in ways unknowable AI system was an iterative process until we hit the sweet spot – and we were ready to launch.

We had daily sessions with client and agency individuals, to triage and either approve or reject any edge cases in the build up to going live. This led to lots of unusual meetings discussing vulgar British slang with a multinational team at 9am. There was a lot of blushing and sending of urban dictionary definitions in advance of the meeting!

It also forced us to properly debate what was deemed as ‘too rude’ for our tool – so the entire brand experience was very much iteratively adapted and evolved in real-time.

Once live, the custom content and the potential for personalised humour from Jose, cut-through and caught the imagination – meaning that it got shared far and wide. 

The results were incredible – with huge amounts of time spent viewing the outputs of the platform, and each video getting shared on average three times with different individuals. 

We also picked up a lot of positive news coverage – reaching national titles and getting further social sharing.

The most interesting part of the project was how it gave us a glimpse at how we might need to work with agentic systems and collaborate with non-human intelligence and creators. Learning how to manage the output in an unpredictable system, and get the balance right between safe and funny, led to a really powerful piece of work that we’re all very proud of.