No doubt about it: Generative AI exploded in 2023. Agencies have been scrambling to figure out how to use it, and, more importantly, how to use it safely. Omnicom is no exception.
Since the agency debuted Omni Assist, a generative AI–enabled virtual assistant, as part of its technology platform Omni last summer, Omnicom has been working to build applications assisted by generative AI that are inclusive in both their makeup and implementation.
Clarissa Season, chief experience officer at Omnicom’s data and analytics shop, Annalect, said that part of her team’s work, which builds technology for other Omnicom agencies, is to make sure the data itself is inclusive, so that any apps built on top of the data are, too.
“We are creating data sets at the base, underneath all of our generative AI capabilities, that are representative of people and races and different sexual orientations or gender orientations,” Season told Marketing Brew.
Houston, we (do not) have a problem
The first step in making Omnicom’s generative AI applications is making sure the audience data the apps are built on is inclusive of different demographics, she said. For example, the Hispanic population, as well as some Asian populations, are often underrepresented in US data sets, Season said. For Hispanic folks specifically, to be truly inclusive, according to Season, the data should represent both those whose primary language is English and those whose it isn’t; the latter, she said, is a category that is sometimes overlooked.
After Season’s team ensures the data is inclusive, they then use a checklist to evaluate the generative AI–assisted applications it builds on top of the data, she said. A chief priority for her team is making sure the applications can provide outputs that can be applied to different industries, like auto and pharma. The checklist has about 25 items across three buckets:
- Response quality: The AI’s output is evaluated for factors like its coherency, how well it conforms to linguistic norms, and whether the output is insightful, rather than just a statement of the input’s content.
- Narrative and technical guardrails: This category’s metrics include evaluating whether the generative narrative aligns with expectations, whether data is being retrieved consistently, whether the output “conforms to truth as we know it, based on the data sets,” and if code is being reliably generated.
- Performance efficiency: The generative AI’s efficiency is tested. For example, the team checks the amount of data being processed to form responses. Another factor evaluated is how long the application takes to create responses.
Get marketing news you'll actually want to read
Marketing Brew informs marketing pros of the latest on brand strategy, social media, and ad tech via our weekday newsletter, virtual events, marketing conferences, and digital guides.
After evaluating an application’s inclusivity and overall effectiveness, Annalect releases it in alpha to a group of Omnicom employees that use the company platform Omni, who then work with clients across various industries to test the application, Season said. The alpha testers provide weekly feedback on the applications and participate in related focus groups.
“We are very cognizant, when we develop our alpha test groups, to be as representative as possible of the population, but also of our clients. Some things that might work in automotive might not work for pharma clients,” Season told Marketing Brew.
3, 2, 1, go!
Once an application is through the testing phase, the Omnicom agency it was built for can use it to support its projects. Omnicom “wants to be very transparent with clients…when generative AI is being applied and when [an] output is based on generative AI,” Season said. “If we generate an insight using our Audience Explorer tool that leads the client to make a decision about what the creative should look like, they need to know that that insight was generated via generative AI.”
While Annalect is focused on making sure its approach is “multifaceted in terms of inclusivity” by creating more representative datasets that can help its apps provide useful outputs for clients of different industries, in the meantime, Omnicom agencies continue to use them on behalf of clients, according to Season.
However, the agency soon plans to train its first client to use the technology directly, Season said.