Interesting question. As the founder of an AI collaboration platform (Markhub), we live and breathe this problem every day. My take is that the best approach isn't a simple choice between in-house vs. outsourced, but a hybrid model focused on the quality and context of the data.

For our foundational models (e.g., text summarization), we start with powerful base models like Gemini and fine-tune them. But the real magic happens with our proprietary data, and for that, outsourcing is not an option.

Here's our approach: Our own product, Markhub, is our primary annotation tool.

When our early users give feedback—like circling a button on a screenshot and commenting "This color is wrong"—they are, in effect, creating a perfect piece of labeled data: [Image] + [Area of Interest] + [Instruction].

We call this "Collaborative Annotation" or "In-Workflow Labeling." The data quality is incredibly high because it's generated by domain experts (our users) as a natural byproduct of their daily work, full of real-world context. This is something an external annotation firm can never replicate.

So, to answer your question on how the trend will evolve: I believe the future isn't a binary choice between in-house and outsourced. The next wave will be tools that allow teams to create their own high-context training data simply by doing their work. The annotation process will become invisible, seamlessly integrated into the collaboration flow itself.

That's a great insight, Paul. As someone who has been researching the data annotation space, your perspective really resonates.

I completely agree that the first-hand, contextual information you get from actual users is something an external firm can never replicate. It seems like the most effective and efficient way to spin the data flywheel at high velocity.

This leads me to a question I've been struggling to understand: If this approach is so powerful, why do you think even companies with the vast resources of Big Tech still rely on what seems to be a riskier path—using external human evaluators—instead of fully building this feedback loop in-house?

I feel like I'm missing a key piece of the puzzle. I would be very interested to hear if you have any thoughts on this.