The 6 a.m. AI report runs, but upkeep stays human

At 6 a.m., Jeff Tannenbaum had a cash-flow report in his inbox. Tannenbaum, a product manager focusing on consumer AI who lives in Los Altos, California, said his team leaves recurring reports to automated AI workflows.

Search results for teapawt’s reproduction of an essay linked on X by Spawn founder Jacob describe his 10-person company using persistent AI agents in engineering and operations. Anish Acharya pointed readers to the approach on X. Spawn’s staffing and output claims are self-reported, not independently established. Accounts from four people at other companies draw a narrower line: AI can produce repeatable work without someone watching every step, but that does not establish who can leave the resulting decisions to it.

Tannenbaum said the workflows produce cash-flow and customer-churn reports on a schedule. Aisha Karim, a staff software engineer at Parcelwise, said her team lets an AI tool draft unit tests for a small software change without watching it work. Marcus Hale, a partner at Northgate Ventures, said agents at two portfolio companies answered routine support questions using documentation.

Producing something usable is a separate question. Karim said an engineer still checks whether drafted tests would catch a bug; one set missed a blank delivery-window value. Sasha Lindqvist, founder and CEO at Ledgerline, said basic rules, not an agent, flag unusual restaurant transactions today, and staff check flags before changing the books. “I wouldn’t let it quietly change the books,” Lindqvist said. Hale said some customer problems still require judgment and follow-up, but he did not know the support teams’ ticket-level checks or handoff processes.

The work also changes after it starts running. Tannenbaum said his team adjusted early reports until they covered the data it needed. Karim said engineers changed how they review drafted tests: “We changed the review question from ‘Do the tests pass?’ to ‘What case would make this fail?’” Lindqvist would first run an agent alongside the existing rules, comparing what each catches and gets wrong before showing staff any new flags.

More output did not settle whether the work improved. Karim said pull requests rose roughly 40% after the tools rolled out, but review time rose too; the comparison does not establish a gain in useful work. Hale said the two support teams each cut staffing by about half after deploying agents, then hired some people back within six months. He did not have exact rehiring numbers. “An answer that sounds right isn’t always a resolved ticket,” Hale said.

Across these four accounts, unattended AI is documented most clearly for preparing repeatable work, not for owning its outcome. How the support teams checked agents remains unknown.

Even Tannenbaum’s scheduled reports have stopped once or twice: A connection’s authentication token expired, holding up the workflow.