Retool CEO David Hsu: 130M hours of work automated in 12 months, targeting 10% of US labor by 2030

Summary

Retool CEO David Hsu argues that the AI industry's core problem isn't intelligence — it's action. Roughly $1–1.5 trillion has been invested in AI in the US, against an estimated $20–30 billion in revenue across all AI products, with around 80% of that coming from ChatGPT and cloud services. In Hsu's view, the gap exists because LLMs have been deployed almost exclusively as chat interfaces, and chat alone won't grow 30x from here.

Retool's answer is to make LLMs operational inside enterprise workflows rather than just conversational. The company says it automated 130 million hours of work for customers over the past 12 months. Hsu's internal target is to automate the equivalent of 10% of US labor by 2030, and he says current trajectory puts Retool on track.

Determinism vs. improvisation

The core architectural argument Hsu makes is about when you want AI to reason freely versus when you want it to follow a fixed path. OpenAI's Operator works well for consumer tasks — buying socks, booking something online — because it can improvise with a browser as its only tool. Enterprise workflows are different. A company running employee onboarding through Retool doesn't want an agent Googling WikiHow articles; it wants the agent to call pre-built, audited actions — ship a laptop via FedEx under an existing contract, provision a key fob through the right system — and reason only about the variables that genuinely require judgment, like expediting a shipment ahead of a storm. Retool's building-block model gives AI those guardrails: pre-built tools and MCP servers the agent calls rather than reinvents.

Pricing model

Retool prices on agent runtime hours, not outcomes. A DeepSeek-powered agent runs at $3 per hour; an o3-based agent at roughly $120 per hour. Hsu says one hour of o3 runtime is worth approximately 40–50 hours of human labor on knowledge work tasks. Compared to Salesforce's per-ticket outcome pricing, Hsu claims Retool's model works out to roughly 95% cheaper per task completed — framing outcome-based pricing as a margin-extraction strategy rather than a customer-aligned one. The analogy he uses: if AWS charged a percentage of your app's revenue instead of compute and storage, that would be a similar structure to what Benioff is doing with Agentforce.

Foundation model dynamics

Retool is model-agnostic by design, but customer behavior has been more surprising than Hsu expected. Large federal agencies — where on-premises Llama deployments might seem like the obvious choice for data sovereignty reasons — are actually running OpenAI models through existing government contracts. Model selection, it turns out, is driven less by benchmark performance and more by procurement relationships, brand familiarity, and what Hsu calls "spiky intelligence": models hallucinate at similar rates but in different patterns, and once teams build workflows around a model's specific failure modes, switching costs are higher than the "one line of code" narrative suggests. The same dynamic is playing out internationally, with European customers gravitating toward Mistral and Gulf-region customers toward sovereign alternatives.

Retool's customer base spans two-person startups to the US Army and Navy, with the common thread being internal software — a category Hsu estimates accounts for 50–60% of all software built, yet almost no engineering prestige.