Should enterprise AI be adopted bottom-up through tools individuals use, or top-down through designed workflow automation?

Alex Lebrun · Jun 19, 2023

A bottom-up, self-serve motion where individual physicians adopt the product without any approval and then advocate to their bosses is an effective way into healthcare institutions

Doctors can start using the web app or Chrome extension tomorrow, fall in love with it, and then push for clinic-wide use

40:29 20VC: Why No Models Today Will Be Used in a Year, Why Open Will Always Beat Closed in AI, Why Proprietary Data is Less Important Than Ever And Why EU AI Regulation is a Disaster with Alex Lebrun, Founder & CEO @ Nabla

Alexander Embiricos · Feb 21, 2026

Top-down AI deployment massively underleverages AI's potential inside a company compared with also putting AI directly in the hands of the people doing the work

Workers who use ChatGPT themselves build intuition for how AI can help, pull it into their own workflows, and feel empowered to steer automation rather than having it done to them

Scope: the two approaches can run in parallel

14:17 20VC: Codex vs Claude Code vs Cursor: Who Wins, Who Loses | Will All Coding Be Automated - Do We Need PMs | The Real Bottleneck to AGI | The Three Phases of Agents and What You Need to Know with Alex Embiricos, Head of Codex at OpenAI

Alexander Embiricos · Feb 21, 2026

Building for individuals first, letting them become fluent, and then pulling in automation will produce far more impact than approaching agents purely through the enterprise workflow lens

Fluency at the individual level is what makes automation land; enterprise-first workflow design misses that

Scope: acknowledges it may be counterintuitive; enterprise GTM education and configuration still required on top

28:59 20VC: Codex vs Claude Code vs Cursor: Who Wins, Who Loses | Will All Coding Be Automated - Do We Need PMs | The Real Bottleneck to AGI | The Three Phases of Agents and What You Need to Know with Alex Embiricos, Head of Codex at OpenAI

Scott Belsky · Dec 20, 2023

You should treat yourself as a business and adopt new AI tools personally before introducing them at work, because novelty precedes utility

Playing with something gives you ideas for how to pilot it in your day job, and a pilot is how you learn whether it's relevant to actual practice — the pattern by which Google Apps, Evernote and Notion entered companies

41:38 20VC Roundtable: Spotify, Adobe & Linkedin CPOs on How AI Changes The Future of Product, Why AI is Now the Product, How TikTok Changed Product, Why Cost is the Biggest Barrier to LLM Usage & Why Incumbents Can Adopt AI Faster Than Any Prior Innovation Cyc

Alexander Embiricos · Feb 21, 2026

Pushing workflow automation before end users are fluent with the tooling produces a pipe dream that only works for the most extreme power users; you first need a base of people using the tool daily and configuring it, after which letting it run independently in the cloud is a small step.

Once daily use makes the tool better with every interaction, the jump to autonomous cloud execution is much smaller — the same logic as their forward-deployed engineer approach.

38:51 20VC: Codex vs Claude Code vs Cursor: Who Wins, Who Loses | Will All Coding Be Automated - Do We Need PMs | The Real Bottleneck to AGI | The Three Phases of Agents and What You Need to Know with Alex Embiricos, Head of Codex at OpenAI

Alexander Embiricos · Feb 21, 2026

Agent tooling should be built as a system of engagement people love using, with automation laddered in later, rather than as workflow automation first

If a tool isn't the naturally easiest way to get something done, people won't use it — automation-first products require FDEs and deployment consultants to force adoption, whereas a loved tool gets used for partial tasks and improves over time

51:34 System of engagement first automation laddered in later

Alexander Embiricos · Feb 21, 2026

Desktop software is back, and the fastest productivity gains from agents will come from meeting users on their own computers first, connecting to systems of record only later

Everything in knowledge work ultimately happens on the user's computer, in the browser or in local apps, and this requires no deployment engineers to set up

52:29 Meet users on their own computers before wiring into systems of record