If you ask a warehouse operations manager where their team’s time goes, you’ll usually hear about picking, packing, receiving, exception handling, and “putting out fires.” All true. What you won’t hear about, because it’s invisible until you go count it, is the swivel chair tax.
The swivel chair tax is what happens when a piece of information has to leave one system, ride along on a human, and get re-typed into the next system. Someone reads a number off a carrier portal and types it into the WMS. Someone copies an ASN out of an email, opens the TMS, pastes it into a new shipment record. Someone reconciles a paper BOL against a screen, then keys exceptions into the ERP. Each move is small. Each move requires a brain. Each move loses 30 seconds to 5 minutes. None of them are anyone’s job title.
We’ve audited warehouse operations across half a dozen industries now. The pattern is so consistent it’s almost boring: 20-25 hours a week, per operations clerk, go to swivel-chair work. In a small warehouse with three clerks, that’s most of a full-time employee, every week, doing nothing but moving data between systems by hand.
The swivel chair is exactly the kind of work a coordinated set of agents can do well — bounded, repetitive, traceable, low judgment per action but high judgment in aggregate.
Where the tax accumulates
Three places, in order of impact:
1. ASN reconciliation. The supplier sends an Advance Ship Notice. Sometimes it lands in the WMS via EDI. Sometimes it lands in someone’s inbox as a PDF. Sometimes it lands in a portal you log into once a day. The data is the same. The pipes are not. So a clerk pulls the ASN out of whichever channel it came in, compares it to the open PO, flags discrepancies, and updates the receiving plan. If you’re a 3PL handling multiple shippers, multiply that by every customer’s preferred channel. The variance is the work.
2. Exception handling. A pallet comes in damaged. The carrier flagged a delivery exception. The temperature log shows a deviation. The supplier shorted a line item by 12 units. Every one of these requires a coordinated dance: photo to file, claim to carrier, note to customer, adjustment to ERP, update to WMS, follow-up reminder to the dock. The actual judgment — “is this a real exception or a documentation issue?” — takes thirty seconds. The dance takes forty-five minutes.
3. BOL matching. Bills of lading come in from carriers in formats that range from “clean EDI” to “scan of a fax of a handwritten note.” Matching them against open shipments, flagging discrepancies, and updating the TMS is mostly transcription work — but transcription work with high-stakes consequences if you get it wrong. So a senior clerk does it, slowly, because the cost of an error is much higher than the cost of going slow.
There are a dozen more — return processing, dock scheduling notes, supplier comms, inventory adjustments, dimensional weight reconciliation. The list is long. The common thread is that all of them sit between systems that don’t talk to each other, and the bridge is a human’s clipboard and keyboard.
Where an agent fleet earns its keep first
This is the work that pays back fastest, in our experience, when you point a coordinated set of agents at it.
The pattern we use:
- An intake agent that watches every channel an ASN can land in — EDI feed, shared inbox, supplier portal — and normalizes them into a single internal representation.
- A matching agent that compares the normalized ASN to the open PO and flags real discrepancies (vs. format noise) using rules tuned with the customer.
- An exception agent that, when a discrepancy or damaged pallet shows up, kicks off the documented dance — photo intake, claim drafting, customer notification, ERP adjustment — and only escalates to a human when the case is genuinely ambiguous.
- A reporting agent that rolls up the day’s exceptions, throughput, and accuracy into a single morning summary so the operations manager isn’t piecing it together from four dashboards.
These aren’t one giant model with a thousand-line prompt. They’re small, focused agents that pass context through an orchestration layer. Each one is easy to evaluate, easy to swap out, easy to monitor.
The numbers we see, conservatively, after three months of running this kind of fleet in a mid-sized warehouse:
- 60-75% of routine ASN reconciliation handled end-to-end without human touch.
- 40-55% of exception cases resolved through the agent dance, with humans only on the genuinely ambiguous tail.
- 10-15 hours/week back per operations clerk, redirected to the work that actually requires judgment — vendor relationships, layout improvements, customer service.
Note the conservatism. We don’t promise 100% automation, and any vendor who does is selling you a demo. The tail of weird, ambiguous cases stays human. The point isn’t to fire the clerks. The point is to stop wasting their time on transcription so they can do the work that needs a brain.
What agents still can’t do on the floor
Honest list, because this matters:
- Physical judgment. Is this pallet actually damaged or is the wrap just torn? Is that box mis-labeled or is the SKU correct and the label is from a previous job? An agent can flag the question. A human still answers it.
- Relationship work. Supplier escalations that go beyond a templated email. Carrier renegotiations. The phone calls that smooth out a 90-day rough patch. These stay human.
- Local context that lives in someone’s head. “Customer X always wants the wood crates on top.” “Driver Y always has the BOL in the cab, not the envelope.” That tacit knowledge has to be captured before an agent can act on it, and capturing it is its own project.
- Anything safety-critical without a human in the loop. Forklift routing. Hazmat handling decisions. Lockout-tagout. Not the place for an agent to be making solo calls.
If a vendor pitches you a warehouse AI that handles all four of these without humans, push back hard. The right design keeps humans on the work where humans add the most value and takes the swivel-chair work off their plate.
How we’d start
If you’re an operations leader and any of this sounds like your week, the cheapest way to find out whether agents can help is to spend two weeks just counting. Have your clerks log, in 15-minute increments, what they’re doing. Don’t change anything. Don’t add tooling. Just count.
Two patterns will jump out almost immediately:
- The percentage of the day spent on transcription will surprise you.
- The number of systems involved per transaction will surprise you more.
That’s the input for the conversation about agents. Not “should we add AI to the warehouse,” but “we just measured that we lose 22 hours a week to swivel chair work — which of these three workflows should we pilot an agent fleet on first?”
Foundation AI designs and deploys the orchestration layer that lets small, specialized agents coordinate this kind of work. If you can tell us, in two or three sentences, where the swivel-chair time goes in your warehouse, we can usually tell you, in the first conversation, whether an agent fleet is the right answer or whether you’d be better off with simpler integration work first. Start there — or read more about our approach to AI automation for warehousing.