Sarah Guo of Conviction: Why LLMs Won't Replace Work That Can't Be Measured

Sarah Guo, founder of the AI fund Conviction, has published an analysis that reframes the central question about AI and labor: not "will AI replace people," but "which work exactly will AI replace, and why not all of it."

Where LLMs hit a ceiling

Language models learn only where there is a clear training signal. The correct answer to a math problem, code that passes tests, a diagnosis that can be verified: all of this can be labeled and scored, and therefore trained on. But Guo points to a structural limit: a huge layer of valuable work cannot be standardized or measured with sufficient precision.

Reinforcement learning by definition requires a reward function, a score for how good a result is. Without such a score, there is no signal for improvement. This is not a temporary problem that the next version of GPT will solve. It is an architectural limit inherent in the nature of trainable models.

What work stays with humans

Guo identifies several categories of work that can't be turned into a dataset:

• Decisions under uncertainty: the right answer is unknown in advance, and the outcome becomes visible years later. A venture investment is the classic example.
• Building trust: between people, companies, investors. Relationships can't be digitized or scored in real time.
• Judgment on one-off situations: no historical analogues, no precedent. Every major business pivot is unique.
• Interpreting implicit signals: tone in negotiations, cultural context, a client's unspoken expectations.
• Creating something new: ideas, product categories, strategies that did not exist before. The dataset for this is empty by definition.

The common thread: no labels, no benchmark, no way to verify the result with enough precision for training.

Why the Application Layer gets venture money

This explains the logic followed by funds such as Conviction (~$300 million under management) and Mento VC: they consistently invest not in builders of foundation models but in products built on top of them.

Foundation models (GPT-5, Claude, Gemini) are gradually becoming commodities. They compete with each other, their capabilities are converging, and the cost of access to frontier quality falls every quarter. Durable value is created not in the model itself but in the context where it is applied, in what models inherently lack.

The Application Layer accumulates exactly that: industry expertise, accumulated behavioral data, workflows that can't be copied without deep domain knowledge, customer trust and, above all, human judgment where automation is structurally impossible.

"There is work that cannot be reduced to a standard or measured.

If it can't be measured, you can't train a model to do it better." This thesis from Guo is the key to understanding which startups get venture money and which don't.

What this means

Guo's argument gives investors and founders a practical filter. Companies built on measurable tasks, such as document processing, template-based content generation and basic data analysis, face maximum pressure from the models themselves and struggle to explain a durable competitive advantage. Those working in the unmeasurable zone gain a structural advantage that won't vanish with the next GPT release.

For investors the question is simple: if a product's value is determined entirely by the quality of the underlying model, it's a weak investment. If the value lies in what a model fundamentally cannot reproduce, it's interesting.