- Singh said agents have done "like 90% of the work" before staff act. In our reading, the human job then shifts toward judging plans.
- In the excerpts, Singh did not say what approvers must check or how they are trained, so firms have to define that themselves.
If approval becomes the main human task, the skill that counts is judging a plan, not drafting it. Aman Singh, who leads AI at Arca, described a firm where agents prepare the work and call on staff. He spoke on the AI Engineer show. Approving plans is only part of his picture, and in the excerpts reviewed he did not say what an approver must check.
What was said
Arca is a financial services firm that says agents run its business. In Singh's account, agents follow a client goal over days, read emails and draft a plan. The plan cannot run without people, so agents open a shared space and pull in a person when someone must execute work. Singh said staff log in to tasks that agents have prepared. Agents, he said, have done "like 90% of the work" and pull staff in when it is time to act. That is his own account of Arca. Staff also add what they learn in person, reschedule events and mark tasks complete. Each of those actions wakes an agent. He then offered a hypothetical: a new January goal to prepare every client's tax situation before anyone asks. Agents would take it on, and a one-click approval email with a full plan would reach staff for each client. He did not present this as a program Arca runs today. Arca also collects staff reasoning as evals, tests that help the system improve.
Why it matters
Our reading: if one-click approval becomes the main human task, the hard part of the job moves. The person is then the final check on a plan they did not write. Singh's example of a changing plan shows the stakes. In it, agents update the need to $80,000 instead of $100,000 after new signals arrive. An approver must know the plan is current and right. So a firm should write down what a sign-off covers. Is it client facts, amounts, or only fit with the goal? Buyers can ask a vendor what the approval screen shows. People who never draft plans may also get fewer chances to learn how plans fail, which would call for deliberate training.
The other side
In the excerpts reviewed, Singh did not say what approvers check or how Arca trains them. The closest he came was goal alignment and the evals. Neither shows how a wrong plan is caught before the click. The excerpts give no error rates or outcomes. The excerpts skip a stretch between roughly minute 8 and minute 10, so he may have covered more. This was a firm that is hiring, describing its own approach. One-click approval may work if agents are reliable, but the excerpts do not show that.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Inside Arca's Long-Horizon Agents That Run the Business - Aman Singh, Arca (2026-10-23)
