Could open-source AI models cut upfront costs for agencies seeking an edge?
Agency groups have made big investments and pursued strategic alliances with AI players, but open-source models could be another route to establishing an AI edge.
Agenices could reduce some of the up-front cost of AI experimentation with open source models / Unsplash
WPP and Publicis are putting eye-watering sums towards AI development this year. But most shops can’t put aside hundreds of millions of dollars and most don’t already have the in-house expertise to build a proprietary AI solution like CoreAI.
One route around those obstacles is open-source software. It’s cheaper and costs the environment less for agencies to build in-house AI solutions on top of large models already created by other organizations. Open-source AI models have been released, in some cases free for commercial use, both by private companies and academics. Apple released the open-source AI model MGIE this week, while Meta’s Llama large language model is available free of charge to private companies.
Brian Klochkoff, executive vice-president, innovation and emerging technologies at Dentsu International, tells The Drum: “We believe that there is tremendous value that can come out of third-party open-source models.”
Advertisement
Julia de Sainte Marie, managing partner of experience at Ogilvy Paris (where the network’s AI labs are located), says that open-source tools can help agencies build applications and tools closer to their particular specifications than tools licensed from other providers. That can be key when taking an AI solution from pilot stage to market.
“Closed-source models allow our people to quickly test how gen AI could change the way we and our clients can work. When we need to turn this into a scalable service, open-source models allow for complete and granular control on how we want it to behave as we can completely customize it and run it anywhere we want,” she says.
Integrating open-source tools isn’t necessarily straightforward, though. De Sainte Marie says that agencies considering this approach need to navigate practical and legal issues.
“From a practical standpoint, open-source models would require developing a framework and interface around it to scale, which requires people, money and time investment,” she says.
“From a legal safety point of view, as we are customizing a layer on top, with our own data, it greatly minimizes legal risk. In terms of client confidentiality, using open-source tools doesn’t inherently pose a risk as the model can be run in any closed environment, detached from any third-party platform that could intercept the data.”
Advertisement
A partnership recently inked between Dentsu and Amazon Web Services (AWS) was pursued, in part, because the US firm’s tech was better suited to the use of open-source AI tools than competitors, Klochkoff says.
As part of the partnership, Dentsu has gained access to AWS’s SageMaker and Bedrock tools – software that can be used to integrate third-party, open-source AI models.
“The Microsoft and Google ecosystems simply didn’t have the administrative controls we wanted to deliver third-party open-source models in a way that met the rigor we have around security and compliance,” he explains.
When vetting a given tool or model, Klochkoff says Dentsu’s general counsel and legal teams get involved to review terms of service, security measures and any copyright or intellectual property issues concerning the data used to train a model.
Suggested newsletters for you
Dentsu and Ogilvy have plenty of resources at their disposal for AI development. Ogilvy’s parent company, WPP, for example, is investing hundreds of millions of dollars in its AI capabilities this year alone.
Open-source models can, however, be a means for small agency businesses to explore the space, says Klochkoff. “I think it comes down to risk appetite. At the scale at which we’re operating, we want to make sure we’re able to explain and mitigate the risks to our customers.”
Agencies should test out new tools on “dummy” data that can’t possibly compromise a client’s business, he adds. That means an agency can tease out specific applications of a given tool and take time to kick the tires on a model to see if it meets the necessary thresholds for sample bias.
“We want to make sure that we have hard and fast use cases,” he says, noting that, should Dentsu find an open-source tool that is worthy of being rolled into its AI tech stack, the firm would seek to collaborate with or reimburse the original creators through the Dentsu Innovation Initiative, the tech arm of the holding company’s venture investment business.
“I don’t think that white labeling other people’s work is the type of thing that we want to do.”