20VC: Why AI Cannot Replace Humans in Enterprise | Why Work Processes Not Models Will Be The Most Valuable Asset in AI | Why Europe Has Lost and Building in the US vs EU with Daniel Dines, UiPath

UiPath's Dines Says Work Processes, Not Models, Hold AI's Value

UiPath's Daniel Dines argues that current AI is more limited than the hype suggests. Models do not learn from experience and are probabilistic at every step, so exact enterprise work still needs deterministic software and people. Since models are interchangeable, the lasting value sits in documented work processes.

Takeaways

• AI memory is not the same as learning. Model weights do not change with experience, so models keep notes but are not transformed on the job the way a person is. Dines calls the idea that a self-improvement loop would generate will wishful thinking. Stebbings argues both humans and models improve.
• Exact work belongs on deterministic software. AI is probabilistic at every step, so a 99% success rate across 100 steps leaves roughly 60% overall. Coding agents have made building automations much easier, but deploying agents has not become easier.
• The map of work is the valuable asset. Models are interchangeable, while the workflows, exceptions, procedures and systems that fulfil a process goal carry the value. UiPath's Cartographer Agent interviews experts as they work to document it, and moving to a new base model needs that exact map.
• AI cannot yet replace people wholesale. UiPath has around 4,000 people and does not use AI as a pretext for cuts. Dines notes some firms that replaced hundreds of support staff are rehiring, and urges a ledger of non-core outcomes like customer trust and mentoring before any cuts.
• Enterprise tokens will shift to cheaper models. Dines predicts 90% of enterprise token flow goes to cost-efficient models within twelve months and advises keeping an open source backup against lock-in. He thinks companies may capture only about 10% of a labour opportunity, around $10bn in US legal.
• Enterprises fear IP leakage through frontier labs. Stebbings relays that Alex Karp says the biggest enterprises fear working with frontier labs. Dines agrees, saying the worry is that intellectual property leaks to competitors through models. He expects open source to hold up against a possible duopoly.
• Europe is largely irrelevant in technology. Dines would tell young European founders to build for the US unless they target a specific market, and doubts he would have succeeded in Europe as he did in the US. He still says sovereignty matters and European customers want on-prem software.

Where Dines and Stebbings disagree

• Infrastructure. Dines thinks the buildout is probably overbuilt for the next three years, while Stebbings argues AI is underbuilt given limits on energy, water and regulation.