Dentsu reduces time to media insights by 90% using Azure AI

Multinational advertising and media company Dentsu is known for delivering innovative marketing driven by unusually robust media and consumer analytics. A comprehensive team of data scientists analyze data from multiple systems to develop media performance insights that guide client planning and budget optimization.

As the complexity of media metrics and the variety of consumer touchpoints have exploded, the needed data insights were taking too long to achieve. In an industry where innovative campaigns often take advantage of trending topics and emerging technologies, cumbersome data analytics put client satisfaction at risk.

“Delayed campaigns slow customer service and may result in missed opportunities,” says Becca Kline, Senior Director of Analytics at Dentsu.

To speed its data analytics, Dentsu turned to generative AI. Callum Anderson, Global Director for DevOps and SRE at Dentsu, says, “We look to generative AI for efficiency and performance solutions, all underpinned by excellence in craft, ethics, and responsibility.” In early 2023, a team began planning a chat-based, predictive analytics copilot that could interact using natural language while referencing Dentsu’s voluminous media metrics and expertise, modeled client data, and best practices. The goal was to create a copilot that would help users interpret results as they optimized media plans for clients.

Relying on a familiar ecosystem

Dentsu wanted its new copilot to interact smoothly with its suite of business apps. The application architecture consists of loosely coupled microservices and multiple generative AI agents that act autonomously but can collaborate using API-based integration and the GraphQL protocol.

“The idea for this copilot was to sit within Dentsu’s suite of applications as another kind of microservice that maintains the same look and feel and inherits a shared set of governance modules,” says Anderson. An established, shared governance approach was important. “We have to be responsible in how we use AI for all our clients,” he explains. “Everything we did, we considered through the lens of how we governed and ensured the AI is responsible.”

After assessing its options for an AI platform, Dentsu chose the Microsoft Azure AI platform, including Azure OpenAI Service and Azure Machine Learning. The company’s familiarity with the Azure ecosystem and access to the latest generative AI models gave the team confidence to move ahead with creating the copilot using Azure AI Foundry.

Proof of concept to production in under 12 weeks

The project provided an opportunity for Dentsu to strengthen its language model operations as it built the new copilot. According to Anderson, the copilot workflow represented a new paradigm for how Dentsu developers thought about assembling applications and app logic. Anderson says, “We had experience with machine learning operations (MLOps), and generative AI operations (GenAIOps) was a step further.”

Microsoft Visual Studio made it easy to set up a local development environment. Then the GitHub GenAIOps with Prompt Flow template helped speed prompt crafting and automations. Simon Ransom, Lead DevOps Engineer at Dentsu, says, “It was great to get a head start with the scripts that were part of the Azure AI prompt flow template.”

With that support, the Dentsu team developed the copilot in under 12 weeks. Dentsu DevOps Engineer Jon Crocombe says, “The people doing prompt orchestration were learning as they were developing. We had our existing app architecture in front of prompt flow, and we ran everything on Azure Kubernetes Service (AKS) at the front end and connected that with our Azure AI hub in AI Foundry to provide prompt flow as a third tier for the app.” The team built the copilot in directed acyclic graph (DAG) flows using Python files.

“The trace facility of prompt flow is really useful for being able to visualize what was going on in the DAG flow,” says Ransom. He also appreciates the solution’s flexibility. “Even if you like to use LangChain, you can still get benefits from prompt flow,” he notes. “It works with other frameworks and tools.”

Evolving together

Throughout development, the team collaborated closely with Microsoft, incorporating user feedback to improve their continuous integration and continuous delivery (CI/CD) tooling, the GitHub template, and Azure AI Foundry. Although working with an evolving product and prompt flow template introduced challenges, Katie Jenkins, Director of Project Management for Dentsu, says, “It was really cool to see Microsoft and Dentsu engineering brains coming together, troubleshooting things in real time, working on cutting-edge products, and learning from each other.”

The new copilot recently went into production, with an Azure OpenAI Service model driving its chat capability. A custom implementation of Azure API Management supports generative AI logging and monitoring to help ensure responsible AI.

“Rolling out was very easy, with zero downtime,” says Ransom. “It supported a blue/green deployment model, which was nice. You don’t always get that out of the box.” Developers are currently adding functionality, reorganizing logic to increase answer accuracy, and incorporating response evaluations, which the initial time frame didn’t permit.

As the project proceeded, the Azure solutions and tools helped the team make the most of their agentic architecture while building the processes that comprise their AI operations. Anderson notes, “I wouldn’t say we’re experts, but we’ve certainly moved from being competent in MLOps to starting to understand GenAIOps challenges and how they’re different.”

Cutting analysis time from weeks to minutes

The resulting copilot makes media insights available to more Dentsu employees via conversational chat. It slashes analysis time by 80% and overall time to insight by 90%, reducing analysis costs. Jenkins says, “Before, our client-facing media planners might have to wait weeks. Once we roll this out, it will take minutes. It’s a huge time-saver.”

The copilot also aids people without specific analytical skills. Kline says, “The copilot agent is super important because it helps the user build and optimize media plans with little to no training.”

By improving model run accuracy and increasing media effectiveness, the new copilot is expected to improve Dentsu clients’ return on advertising spend (ROAS).

“We can complete media plans faster and with more confidence to help clients optimize their media spend and exceed their goals for customer response,” says Kline. “That makes everyone who contributes to the campaign successful.Ultimately it empowers our clients to engage with more customers and grow their businesses.”

The project also helped strengthen Dentsu’s AI operations know-how for the future. Dentsu plans to make additional investments in AI apps using a similar framework. The team channeled what they learned back into the GitHub template so others using Azure AI Foundry in the future can adopt AI at scale with more speed, confidence, and predictability.

“This AI project was a significant step forward for how we build and manage our generative AI solutions,” says Kline. “It’s also how we drive impact for our clients.”

Anderson concludes, “With Microsoft, we’re turning our media expertise into a competitive advantage—and harnessing data to build brands and drive business growth.”

Discover more about Dentsu on Instagram, LinkedIn, X/Twitter, and YouTube.