Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

CTV contextual AI is at an inflection point.

So why does that matter for reaching an Hispanic audience?

[Photo: Getty Images]

Profile photo of Tony Gonzalez

CEO & co-founder · Mundial Media

The money flowing into contextual artificial intelligence connected TV right now is real. The capability backing most of it is uneven. That is the honest state of the category, and brands committing serious budget to this channel deserve a clearer picture of what they are buying.

The gap between where the technology genuinely delivers and where it is still catching up to its own pitch has closed over the past two years. But it still exceeds what most vendors will acknowledge in a sales conversation.

THE PROBLEM WITH LATENCY

Latency is the category that needs to improve most. U.S. CTV ad spending hit $26.6 billion last year, with nearly half of all inventory moving through programmatic systems. Those systems run on bid response windows measured in fractions of a second. Real-time contextual analysis often cannot finish in time for that window, so the platform falls back on a static label it generated by pre-classifying the content hours or days earlier. The system matches the ad to that label rather than to what is playing when the bid fires.

Halfway through a thriller episode, the content has already shifted several times—tension, quiet, something close to humor, back to tension. One label sitting atop all of that adds very little context.

Every partner conversation I have eventually comes down to one question: What is the millisecond latency your contextual engine introduces during a live bid stream? I have had that conversation enough times to know that the answer—or the way someone avoids giving one—tells you most of what you need to know about the system.

The pitch says real-time. In most cases, the system underneath it classifies that content hours or days earlier. I have sat with enough mid-campaign data to recognize what that drift looks like in practice. The optimization engine works from a signal that has already moved away from what the audience experienced, and the standard reporting does not surface that clearly until the results come in.

THE ISSUE WITH TRAINING DATA

Training data is the second issue, and it rarely comes up as it should. Engineers built most models in this space on English-language content because that is where large, labeled datasets existed and where investment was concentrated.

Multimodal AI performs at the level of the content it was trained on. Brands trying to reach audiences whose primary content lives outside that training universe run into a blind spot that standard campaign reporting rarely catches.

A Spanish-language audience provides a great example. According to Nielsen’s Diverse Intelligence Series, Hispanic consumers represent nearly 20% of the U.S. population, hold more than $4.1 trillion in purchasing power, and spend 55.8% of their total TV time on streaming, well ahead of the national average of 46%. The same report found that less than 1% of digital ad spend from U.S. online retailers was directed to Spanish-language websites in the first quarter of 2025.

We have worked through this directly at Mundial Media in building contextual AI for Spanish-language and multicultural streaming audiences. A model classifying Spanish-language content by genre handles the straightforward part. Understanding the cultural references, emotional layering, and tonal specificity that shape how the audience experiences the content requires training on the content itself. That is where the investment must go. And it is where the category has the most ground still to cover.

WHAT IT TAKES TO GET IT RIGHT

Getting this right means running three things simultaneously inside a live bid cycle. The system must analyze video at the frame level, read emotional and tonal signals from the audio beyond basic transcription, and surface thematic context through a real-time language layer. Engineers need to build it all specifically for long-form streaming.

There is also a privacy dimension that receives insufficient attention. A system deriving its signals from content analysis builds in structural advantages as the industry moves away from identifier-based targeting. The platforms making that engineering investment are building toward something with real staying power.

Brands evaluating partners right now should push directly on millisecond latency during a live bid stream. They should also look at how the system generates the scene-level signal, and at what point in the bid cycle; which specific content environments the team has validated the model against; and how recently they updated the training data. The streaming landscape from two years ago looks meaningfully different from today. And a model calibrated on that landscape is already working from an outdated picture.

Partners who have done this work are easy to spot once you know what to ask. They engage in the specifics of their architecture because they have worked through hard problems and have real answers. The evaluations become more productive, the conversations more substantive, and the gap between what is promised and what is delivered closes considerably. That is the market worth building toward, and based on where the technology is heading, the industry is getting there.

Tony Gonzalez is co-founder and CEO of Mundial Media.