Can AI actually replace an employee?
The headlines say AI is replacing workers. Eric and John dig into what's actually working, what isn't, and where the real ceiling is right now.
Show content
Listen to the audio, watch the video, or check out the Show Notes for a summary, key takeaways, and links to people, content, and tools we mention.
Audio
Listen on Spotify or Apple Podcasts, or here on Substack:
Video
Summary
Eric opens with a viral post from David Cramer, founder of Sentry, pushing back on the idea that people are running fleets of AI agents doing real work overnight. John responds from firsthand experience, explaining that his company has run dozens of internal experiments, and the honest answer is that almost none of them are used to do real client work.
They map the landscape by use case, from personal productivity tools to team-wide deployments, and find that the team tier is where almost everyone stalls. The tools are developer-focused, the adoption problem is real, and getting AI to work reliably across a group requires far more investment in guardrails and oversight than the demos suggest.
The episode ends with guidance on what’s practical today. The most compelling near-term model is not a zero-person company but a “co”: a single AI assistant that one person owns, trains over time, and stays responsible for.
Key takeaways
Impressive demos and production deployments are two different things: most agent experiments stay internal, and the gap between “kind of works” and “works with real clients” is larger than most AI coverage admits.
What works at home does not automatically work at work: personal AI tools, team tools, and company-wide deployments each have different friction points, and almost everyone has figured out the personal tier and almost no one has figured out the team tier.
AI tools are built by developers, for developers, and it shows: most frameworks default toward building and generating, with not enough support for planning, quality checks, and oversight, which limits what they can reliably do.
AI will try to answer even when it shouldn’t: agents respond by default even without enough context to be accurate, and building the guardrails to prevent that is harder and more expensive than it looks.
Owning a single AI assistant beats managing a fleet: a one-to-one “co” that you prompt carefully, iterate over time, and stay responsible for is more practical and more trustworthy right now than trying to orchestrate autonomous teams of agents.
AI helps analysts work faster, but it cannot replace what they know: giving AI access to data and asking it to run queries works well when a skilled human with domain knowledge is in the loop; without that, the answers are unreliable.
Notable mentions and links
David Cramer’s post on X is the episode’s opening provocation, in which the founder of Sentry argues that nobody doing serious work is running 20 agents overnight, and that the real benchmark is whether you can reliably ship one production-quality fix at a time.
Block, Inc. is the financial services company behind Square and Cash App, and its high-profile layoff of over 4,000 employees in February 2026 became a recurring example in the AI-is-replacing-workers news cycle that frames the episode.
OpenClaw is an open-source personal AI assistant that runs on your own hardware, connects to messaging channels like iMessage and Telegram, and can be given broad access to your computer, including, for those who push it furthest, credit cards and prediction markets.
Zo Computer is described as a middle ground between OpenClaw and a consumer app: AI running inside a secure cloud computer with built-in limits, more powerful than a chat interface but without the security exposure of a fully local setup.
Poke is a consumer-facing personal agent that works entirely through existing messaging apps like iMessage or Telegram, with no separate interface of its own.
Obsidian is a note-taking app John connected to OpenClaw and GPT for a personal productivity experiment, running a carefully iterated prompt on a schedule to filter ads from his email inbox multiple times a day.
Claude Code is cited as the most technically capable individual work tool available and as a clear example of how current AI products skew heavily toward software development use cases.
Polymarket comes up as an example of how far some OpenClaw users push autonomous access, with agents given permission to trade on prediction markets on their behalf.
The “Mad Lib problem” is John’s phrase for why AI adoption stalls on teams: the interface is not fully blank, but knowing which pieces to fill in requires judgment most non-technical people have not built yet.
The “co” model is John’s preferred near-term frame for working with AI: a single assistant that one person owns, instructs carefully, and stays accountable for, rather than an autonomous system operating on its own.