This week's Autonomy Exchange gathering in Malmö included a session on a question that does not usually get its own slot at a robotics conference: what does the back office of a physical AI company look like once agents are doing a share of the daily work. It is worth writing down plainly, because it says something about the same bet the product makes on the warehouse floor — give a system continuous, structured visibility into what is actually happening, and the work around it changes shape.
The story hardware companies don't usually tell
Most companies selling autonomous mobile robots and forklift automation into warehouses are organized, on the back end, the way enterprise software companies have been organized for two decades: people doing account research, people drafting outreach, people scanning for competitor and customer signals on a weekly or monthly cadence. There is rarely anything structurally unusual to report, because the operating model looks the same everywhere.
Staer's does not. A meaningful share of its day-to-day commercial and marketing motion — account research, dossier-building on prospective customers, inbound triage, first drafts of outreach and content, monitoring for competitor and category news — runs through AI agents rather than a person opening a CRM each morning. That work happens continuously instead of in the bursts a human calendar imposes: a dossier on a prospective account gets written the day a public signal appears, whether that is a funding round, an automation announcement, or a leadership change, not whenever someone next has a free afternoon.
What deliberately stays human
None of this is autonomy for its own sake, and none of it is unsupervised. Every piece of content an agent drafts is reviewed by a person before it reaches a customer, a prospect, or the public — nothing publishes itself. Negotiation, the judgment calls on which deals to prioritize and how hard to push a stalled one, the physical work of getting cameras and hardware installed on-site, and the relationships that carry a long enterprise sales cycle from first meeting to signed contract all stay squarely human.
That division is a deliberate design choice, not a gap waiting to be closed. The calls that require trust, context a system does not have, and personal accountability are exactly the calls a company should not want to hand off, no matter how capable the tooling gets. The agents are given the parts of the job that are repetitive, well-scoped, and benefit from never sleeping — not the parts that require someone to be in the room.
In practice that looks less like replacing a marketing or sales team and more like changing what the people on it spend their time doing. Less time assembling a research brief on a prospective account from scratch, more time deciding what to do with it once it exists. Less time noticing that something needs a reply, more time deciding what the reply should say.
Why it matters to a company evaluating a vendor
For a warehouse operator or a 3PL comparing spatial-intelligence vendors, how a company runs its own operations is a reasonable proxy for how it will behave as a supplier. A company that has already built continuous monitoring and structured human review into its own commercial process is more likely to apply the same discipline to a customer's fleet data — flagging a stalled sensor or a coverage gap quickly rather than at the next scheduled check-in, and doing it consistently across every site rather than depending on who happens to be paying attention that week.
It is also a small preview of where physical AI is heading more broadly. The companies building spatial intelligence for warehouses are, in parallel, becoming a proof point for what continuous perception does to an organization's own tempo — the same argument, running in both directions at once.
An open question, not a finished answer
None of this means agent-native operations are a solved problem, or that every function should run this way. It is closer to a running experiment: which parts of a physical AI company's own work genuinely benefit from continuous, agent-driven attention, and which parts still need a person in the room no matter how good the tooling becomes. That line will likely keep moving as the tools improve.
Anyone building a physical-AI company and wrestling with the same question — which of your own decisions should stay human, and why — is welcome to compare notes.