Your register might be free but checkout is moving to whoever an agent can pay.

This is an agentic POS case. The till is bundled to capture the swipe. The screen already converged: menu, 86, tax, close. Agentic payments go to the merchant a buyer agent can quote, refuse, and settle. Counter is that merchant.

“A POS is a human screen over a perishable good and a cash drawer.”

Independent case Counter-service POS Sandbox · Fake pay · No take-rate claim
Payments attach
The register is priced to win the swipe. Software margin is already gone.
Commodity screen
Menu, 86, tax, close. Another skin will not hold the account.
Channel shift
An agent routes volume to the counterparty that can quote and refuse.
Settlement control
Policy is public. A person still owns comps. Paid is written once.

Software margin has already transferred to the processor

The operator cannot fund a better register. Food and labor took the margin. The POS company cannot raise the software price, so it bundles the screen and lives on processing. That bargain worked while a person stood at the counter and the card was swiped on the vendor’s device.

Delivery apps already ran this play. The order started somewhere else, the register became a printer, and the take rate left with the channel. A buyer agent is the same structure with a faster counterparty. Membership “tiers” in this build are a priority pickup and a campaign, and points move only when the register commits them. If the agent keeps the book, the store is a kitchen again. There is no price list here. I am arguing for a position, and I am not packaging a SaaS.

Prior engagement: till, warehouse, and loyalty on one book

This is not a first encounter with a till. JPOS was a client consulting engagement, run while I was at BVXpress and not a BVXpress product: POS and warehouse management on one system, deployed at four sites, with about $200k a month in sales through those tills.

Counter keeps that rule and adds the agent. Stock is the 86 the next quote already knows. A rewards plan can be read when the price is quoted. Points move only when the register commits the sale. The buyer that is a model is new. Refusing to split inventory, loyalty, and settlement across three systems is the part I had already paid for.

A model on the existing till leaves the economics unchanged

The incumbent move is to drop a model onto the existing till. The economics stay put. The model guesses whether the store is open, whether oat milk is 86’d, and whether the cart is over the house limit, because those facts live on a screen built for a cashier. Every wrong guess becomes labor. A thin counter cannot staff its way out of a bad interface.

Base case: model attached to the screen

Screen or side inbox → guess hours, stock, cap
Miss → cashier retypes → several ways to call it paid
  • The agent scrapes a human interface, or waits in a separate inbox
  • Out of stock dies. Someone on the floor rebuilds the ticket
  • Hours, the spend cap, and the 86 sit where the model cannot read them
  • Chat, the browser, or a webhook can each mark the ticket paid
  • A second system polls for state and drifts. The store no longer knows its own book

Alternative: merchant as the counterparty

Published policy → firm quote
Refusal the agent can use → person on the exception → one write of paid
  • Staff and the buyer agent share one checkout. The machine is a customer of the store, not a scraper of the staff UI
  • Out of stock, closed, and over cap come back as codes. The agent substitutes, waits, or stops
  • Hours, scopes, and the spend cap are public. The cap is not a prompt
  • Paid, and the kitchen fire that follows it, happen after payment succeeds, on one server path
  • This store is the book. A sale copied to another POS can still lag. I say that on the product page because pretending otherwise is how you become the printer

Defensibility sits in quote, refusal, and settlement

A moat in this business is not a nicer till. It is being the party an agent must deal with to get a price, a refusal, and a settled ticket. Three rules hold that position. Break any one and the order moves to whoever still has it.

1. The agent is a customer of the store

Same checkout the register uses. The agent reads policy, takes a quote, and pays that ticket. There is no staff screen in the path, so there is nothing for a model to scrape and nothing for a marketplace to replace with its own menu.

2. A coded refusal retains the order

Out of stock, closed, needs a manager. The agent retries, substitutes, or halts. A dead error burns the ticket and hands the customer to the next merchant. A code keeps the order here.

3. Exception rights stay with the operator

Caps, comps, and refunds sit with a manager. A helpful model that can raise the spend limit or write paid is an unauditable cashier. Approval is the control. Paid is the record.

Each surface is assigned to one economic leak

I kept the surfaces few, because a strategy you cannot operate is a slide. Each one exists to stop a specific leak: a buyer who cannot see the rules, a cashier who retypes failures, a second channel with its own book, a loyalty program the agent can raid.

Policy published before integration

The landing page states the case, then loads this location’s public policy in the open. From there: Open the register, For POS, Text to order, For agents, Rewards. The first pass is the oat-milk failure, about three minutes. If a partner cannot see the refusal, they will not wire the agent.

Operator control of 86, cap, and close

This is the operator’s desk, not a chat window. Sign in. The banner says FAKE PAY · SANDBOX, because I will not dress a demo up as a processor. The trace stays empty until Run agent. Two goals matter commercially: the 86’d latte, and a cart over the house cap. Staff 86 an item, Approve or Reject, and close the day. Help is blunt. Substitute when the item is gone. Stop when the store is closed. The demo buyers are deterministic on purpose. I want the policy on screen, not a model having a good day. A fuller staff ticket can ask for a comp or a refund in chat. That request dies there. Chat does not settle.

Second channel on the same book

Shelf tag, handset, same book. “Text a shelf code. PAY sends it. NO drops it.” The replies are the policy, said to a person: paid and the register has it; 86’d and not added; store closed; that total needs a manager. A second channel that writes its own paid state is how the register became a printer last time. This one does not get its own book.

Loyalty commits on the merchant’s book

A merchant can publish a plan. A quote may look at it. Points do not move until the register commits them. The agent can file. The store settles. Leave points on the agent’s side and you have donated the relationship.

Worked cases: labor, spend authority, channel control

  1. Oat milk is 86’d. The buyer wants the oat latte. The store answers out of stock. The agent substitutes, the quote clears, fake pay completes. Nobody on the floor rebuilds the ticket. On a counter, that rework is the margin. Put the milk back and the original drink is quotable on the next call, because staff and the agent are looking at one 86 list.
  2. The cart clears the house limit. Past the auto cap, the buyer may not pay alone. The ticket waits. Complete stays refused until a manager hits Approve or Reject. The model cannot raise the cap. That limit is how a drawer gets given away by something that was only trying to be helpful.
  3. Someone texts a shelf code. PAY runs the same checkout. Closed, 86’d, or over the cap comes back on the handset in the same terms. A paid ticket lands on the register. The phone is a channel. It is not a second merchant.

Constraints required to keep settlement on the book

System of record
Staff and the agent read the same counter. An 86 is on the next quote. Copy the sale downstream and that other system can lag. I would rather admit the lag than pretend two books are one store.
Exception rights
Exceptions are decided by a person. Comps, refunds, and over-cap carts can be requested. A manager approves them. A model that can settle will eventually settle in its own favor.
Catalog scope
The buyer orders finished items. Modifiers are already baked into the catalog. An agent that invents a drink creates a comp, a remake, or both. That is a margin leak with a friendly interface.
Evidence bar
Fake pay, fixed clock, no card. Hours and ready-by stay still so a visitor cannot confuse a moving clock with a working store. I am not claiming a processed dollar.
Downstream POS
A downstream register may receive the sale. It may not mark Counter’s ticket paid. Paid, and the kitchen ticket after it, are written in one place after payment succeeds. Give that write away and the moat is a logo on someone else’s drawer.
Scope boundary
This version is the counter and the argument. No certified listing with ChatGPT or Google, no card-present hardware, no voice line, no offline mode. The next proof is one live merchant on Stripe. Until that charge clears under the same rule, I do not have a business. I have a register that behaves like one.

Commercial proof is one live merchant under the same rule

One merchant. A real card. The same rule: payment success is the only way a ticket becomes paid. If a browser, a chat, or a partner system can write paid, we are back to selling a screen, and the screen is already free.

One live merchant Stripe Paid written once

Implications: volume follows the merchant an agent can settle

“Models never mark a ticket paid.”

POS on GitHub