Multimodal models-combining text, image, video, and audio-are moving from research to product. For VPs of Engineering and Product, this opens new UX and capability vectors. The question is no longer "can we do it?" but "where does it create durable value and how do we ship it responsibly?"
Where multimodal matters
Use cases range from assistive interfaces and content understanding to creative tools and search. Identify where "see and reason" or "hear and act" creates a step change in user value versus text-only flows. Not every product needs vision or audio; focus on use cases where the modality is essential to the task (e.g., describing an image, transcribing and summarizing a call) rather than a nice-to-have.
Run small experiments before committing. Prototype with existing APIs, gather feedback from real users, and measure quality and latency. Multimodal models are heavier and often slower than text-only; ensure your UX and infrastructure can handle the tradeoffs. Define success metrics up front so you know when to scale, when to iterate, and when to kill the experiment.
Build vs. partner
Foundation model providers are racing on multimodal. Evaluate APIs and fine-tuning options before building custom stacks. Invest in evaluation, safety, and UX; consider partnership or acquisition for core model capability. In most cases, partnering with a provider and focusing your build on integration, evaluation, and safety is the right call. Only build custom when you have a clear differentiator (e.g., domain-specific data, latency or cost requirements) that off-the-shelf models can't meet.
Lock in evaluation and safety early. Multimodal outputs are harder to evaluate than text; define what "good" looks like for your use case (accuracy, safety, fairness) and build pipelines to measure it. Plan for content moderation and PII handling; image and audio can contain sensitive information that needs to be filtered or redacted.
Roadmap and risk
Place multimodal bets in the context of your product strategy and competitive landscape. Plan for latency, cost, and quality variance; have fallbacks and clear success metrics before broad rollout. Communicate to stakeholders that multimodal is evolving fast: models and APIs will improve, and your architecture should allow you to swap providers or models without rewriting the product.
Manage risk around bias, safety, and compliance. Multimodal models can inherit and amplify biases from training data; test across diverse inputs and user segments. Align with legal and security on data handling, especially if you're processing images or audio that may contain personal or sensitive information. Document your assumptions and limits so the rest of the org understands what multimodal can and can't do today.
Multimodal is a capability layer, not a strategy. Tie every initiative to a concrete user outcome and business metric.