Kantar LINK AI on Azure: AKS Architecture, API Limits and Content Optimizer

Advertising is not the usual WindowsForum beat. The architecture is the useful part for IT pros, though. It shows how a data-rich company is combining older machine-learning models with large language models, a Kubernetes compute layer, message queues and managed databases, and how it keeps people in charge of what ships. It also shows the gap between what a vendor case study says the system can do and what customers can actually reach through the product today.

From "Will this work?" to "How can we make it better?"​

Kantar has spent decades measuring ad effectiveness from its own historical data. A global advertiser would bring in a TV spot or a digital video and ask whether it would land. Generative AI has changed the job. Brands now produce far more content, in more formats and channels, and faster than before. Marketers want to know how to improve a piece of creative before they commit media budget.

Ashok Kalidas, Kantar's Chief AI Scientist, puts the goal in terms of doing something with the score. That means improving low-scoring ads, creating new versions, personalizing and adapting them for different platforms, and then rescoring to show the new versions actually do better.

Microsoft says the work was done with its Frontier Company Forward Deployed Engineers (FDEs). The result is described as a creative intelligence platform that ties Kantar's proprietary data and AI models into creative workflows before content goes to market.

The feature has been visible for a while. Research Live reported in June 2026 that Link AI will also have an AI-supported content optimiser to guide teams on how to improve creative prior to launch, including visualised recommendations for static and video content. Kantar's own support documentation describes the content optimiser as combining predictive evaluation with actionable, AI-driven content optimisation in a single workflow to test more, learn faster, and continuously improve creative at scale.

Section summary: LINK AI is moving from a scoring tool to a score, fix and rescore loop. Azure is the foundation and Microsoft engineers were embedded in the project.

The Azure stack behind LINK AI​

According to Microsoft's customer story, the platform was co-engineered as a cloud-native system on Azure. The listed components are:

Most AI projects end up looking something like this, even though a slide showing "LLM goes brrr" is more common.

• AKS runs a group of models, not one. More than 20 models under variable load is a normal Kubernetes job. Batch uploads arrive in spikes, and individual jobs need fast turnaround.
• Service Bus separates intake from processing. When someone uploads a large batch of video, a queue stops the burst from overwhelming the scoring services.
• The LLMs sit alongside the traditional ML. Microsoft says the FDEs helped Kantar adopt LLMs "as key components" that augment its existing ML models, rather than replacing them.

The organizational side is part of the story too. Jonathan Adler, Kantar's Director and Enterprise Architect, credits the FDE team with technical skill and says they also helped convince colleagues to accept a new architecture. Anyone who has tried to move a long-established analytics team off an architecture it trusts will know how much that matters.

Section summary: The design is a familiar enterprise AI pattern: proprietary data plus traditional ML plus LLM reasoning and generation, run on AKS, fed by queues and governed through Foundry.

The numbers, and what they leave out​

Microsoft reports these results from the rebuild:

• 20× faster batch scoring
• 4× lower platform operating costs
• 6× faster single-video scoring
• More than 100,000 video ads scored in production
• Customers move "from insight to action in minutes, not hours"

Microsoft's companion technical material has a separate set of figures. It says that after rebuilding its LINK AI creative effectiveness platform on Azure, Kantar reduced ad-scoring time from as long as an hour to just a few minutes and enabled tens of thousands of ads to be evaluated within hours. Microsoft's September 8 Cloud Blog post ties that throughput to an unnamed global beverage client and gives no exact test count.

Kantar's own product page says individual predictions are delivered in minutes, with large-scale testing across thousands of assets completed in hours. It describes the underlying data as a database of 300k+ ad tests, trained on 35 MM+ human interactions.

These figures deserve some caution:

1. All the figures come from the vendor and its cloud provider. None of the reviewed material offers independent benchmarks or an audit.
2. The 100,000-ad figure has no scope. The story gives no time period, customer count or definition of what counts as a score.
3. Faster scoring does not prove better ads. The speed and cost gains are platform metrics. The reviewed sources give no measured increase in campaign ROI, sales or effectiveness from AI-generated variants.
4. The 400-brand example is hypothetical. Microsoft says a company with more than 400 brands may run tens of thousands of campaigns and millions of assets each year. That describes the scale of the opportunity, not a named Kantar customer or a result.
5. Heath Greenfield's revenue comment is a forecast. Kantar's EVP of Solutions & Platforms says the work will grow its creative solutions portfolio and bring in incremental revenue rather than cannibalizing existing business. That is an executive's expectation, not a reported financial result.

Section summary: The platform efficiency figures are specific and credible as vendor claims. Claims about business outcomes are still expectations.

What developers can use today​

Kantar's public LINK API documentation fills in an important practical detail:

• Customers build their own integrations. Kantar supplies the API, documentation and support materials. Building the integration falls to the customer, its agencies or technology partners, so a development team is needed.
• Where it connects: digital asset management (DAM) systems, compliance platforms, creative quality tools, activation and demand-side platform (DSP) environments, BI tools and data lakes.
• How tests start: automatically, through rules you define (for example, when an asset reaches a certain stage in a DAM), or manually.
• Getting access: request a Client ID and Secret from Kantar Support, generate an access token, then start calling the API. Kantar recommends first agreeing your workflow, use case, required metrics, expected volume, and whether the API will only return scores or also trigger tests.
• What the API returns: a predefined set of LINK AI and LINK+ metrics, plus any custom metrics for your setup. Not every dashboard metric is available. Heat maps and trace lines stay in the Kantar Marketplace dashboard.
• The main limit: the content optimiser is not in the API. Kantar's FAQ says the LINK AI content optimiser is currently used through the Kantar dashboard. The API returns scores and results with links back to the dashboard, where optimization continues.

So the generative loop described in the Microsoft story (recommend, generate, rescore) is a dashboard feature today. A pipeline that pulls scores into your DAM will not also generate improved variants.

Kantar's support documentation also lists practical upload limits. Standard LINK AI projects take up to 20 ads, and LINK AI Batch takes 100 to 1,000 TV or digital ads per project. Video files must use supported containers such as MP4, MOV or WEBM with H.264 or MPEG-1 encoding, and AV1 is not accepted.

Section summary: Scores can flow into your systems through the API. The optimize-and-generate step currently happens in Kantar's dashboard.

Human sign-off stays in the loop, and "in-flight" is still ahead​

Microsoft describes the platform as agentic, but the story is careful on one point. Generation is "controlled," and people retain final authority over what goes to market. Nothing in the reviewed material suggests the system publishes ads by itself.

The most ambitious capability is still in the future. Kalidas says real-time, in-flight optimization, which would tweak and personalize ads using continuous market feedback, is the capability he is "most excited to validate." He calls it "within reach," which means it is not shipping yet. Today the system optimizes creative before launch, and that is how it should be judged.

Why this matters beyond ad agencies​

For WindowsForum readers who build or run enterprise systems, Kantar is a useful example for three reasons.

• A proprietary dataset is the real asset. A generic image model cannot score an ad against decades of historical test results. The LLMs make Kantar's existing data more useful, but they do not replace it.
• Separating the layers pays off. Foundry handles orchestration and governance, AKS handles compute, Service Bus handles flow and the databases handle state. That lets each part scale separately, which is where the 4× cost reduction would plausibly come from. That last point is our reading of the architecture, not a claim Microsoft makes.
• Getting people to adopt it is part of the work. Microsoft and Kantar both highlight persuading internal teams as much as the code. Organizations weighing the Microsoft FDE model should expect the same challenge.

The bottom line​

Kantar's LINK AI rebuild is a solid Azure case study with named services, specific throughput and cost figures, and a clear commercial reason behind it. Treat the numbers as vendor-reported. Remember that the content optimiser currently lives in Kantar's dashboard, not its API, and that real-time in-flight optimization is still a goal. The architecture is worth studying. The claim that generative AI will make every ad better has not been shown.