Dan Shipper, founder of Every, says his roughly 30-person company has rolled out Claude Code, AI-powered customer support, and automated email tools across daily operations, yet no one was laid off because of AI. The opposite happened. Human work expanded, and job boundaries shifted: managers now submit code, engineers spend more time with customers, and Shipper says 95% of his work email is drafted by AI, though he still reviews messages one by one.
In a new essay published on every.to, Shipper pushes back on the common assumption that stronger models automatically mean fewer human jobs. His argument is narrower and grounded in practice: as long as AI is answering questions framed by humans, it remains behind the people deciding what the real problem is, whether the output is good enough, and what should happen next.
Agents take repetitive work, while people handle direction and review
Shipper describes two main patterns inside Every. One is the “coworker agent,” software embedded in collaboration tools such as Slack and called on like a teammate for specific tasks. The other is the embedded agent, placed directly inside a workflow such as customer support to process routine tasks with clearer boundaries.
One example in the article is Fin, an AI support agent. During one week in May, Fin participated in 65% of Every’s 202 customer support conversations and closed 81 tickets without human intervention, equal to 40.1% of all addressable conversations. That changed the support manager’s role: less time spent replying to basic tickets, more time spent building systems and handling cases that need closer judgment. Shipper’s point is simple. Agents can absorb stable, repetitive layers of work, but once complexity rises, quality still depends on humans staying actively involved.
Automation creates another layer of work: maintaining the automation
He calls this setup the “human sandwich.” AI handles the middle of the task, while people sit at the beginning and the end, setting direction, correcting mistakes, and approving results. The catch is that the system itself needs constant care. Every initially tried assigning each employee a personal agent, then moved back toward team-level and company-level agents because neglected agents quickly became stale and unreliable.
The company now has a dedicated AI engineering team responsible for keeping those systems working. Even tasks that sound simple can turn into large operational builds. Shipper writes that one PowerPoint automation workflow at Every includes 24 skills and 18 scripts, with token costs reaching $62 per generated deck. In his framing, automation does not remove people from the process. It shifts people into design, supervision, maintenance, and quality control.
As model scores rise, judgment becomes the scarce layer
Shipper also cites benchmark gains to show that model capabilities are climbing quickly. In Humanity’s Last Exam, top model scores rose from the low single digits a year ago to around 44%. In GDPval, which measures performance on real economic work against human output, scores climbed to roughly 85%. METR’s early May testing of Claude Mythos found an 80% success rate on tasks that would usually take human experts about 4 hours.
Even so, he argues that benchmark progress does not translate neatly into disappearing work. AI turns “yesterday’s human capability” into a commodity: the parts of expertise that can be written down, trained, and repeated. Once the same models are available to everyone, output scales fast and sameness spreads with it. What remains scarce is not the ability to produce a first draft, but the ability to spot what is wrong, decide what should be cut, reshape the result, and choose the problem that matters right now.
That, in Shipper’s view, explains why Every did not shrink after adopting AI at scale. The technology is reshaping expert knowledge work, but in the cases he is seeing inside the company, it is not eliminating the need for people. It is producing more work that depends on human judgment, review, and system-level oversight.