Let’s talk about the quote request that went out to five suppliers at once, where every one of them thought they were the only one in the room.
Picture a print shop on the East Rand. Owner’s name above the door. Friday afternoon, a request for pricing lands from a procurement address. It’s polite. It’s specific. It knows his lead time, it knows the paper stock he runs, and it asks for fifteen percent off and sixty days to pay.
He says yes. Not because the numbers work. Because the pipeline was thin and the email was polite.
Three weeks later, four more. Different companies. Same tone. Same structure. Same fifteen percent.
He thinks he has a pricing problem. He has a search problem. Somebody else’s.
On the fifth of October this year, two researchers put a paper on arXiv called Parallelism or Concession. Xu and Zhu. The subject is a buyer’s agent that takes one purchase and splits it into many negotiations running at the same time. Not one long conversation with many suppliers. One decision, many parallel threads, each carrying its own price cap.
Their result isn’t a metaphor. It’s arithmetic. The value of each extra thread falls away quickly — but with every thread you add, the price the buyer can accept on each one goes down. They write it plainly: parallelism substitutes for concession.
Which means the buyer’s agent never has to give ground. It doesn’t need to be a better negotiator than you. It needs to ask in more places at once and take the first yes that lands under a number it decided before you replied.
This isn’t a lab curiosity. Deloitte’s own write-up of agentic commerce describes the mature version: a hospital’s procurement agent and its suppliers’ sales agents configuring price, quoting and terms in real time. McKinsey describes agents assembling and negotiating bundled offers across retailers. The people building this measure success in basis points, and basis points come out of somebody’s margin.
Here’s the pattern, and it’s the third time we’ve watched it.
First the feed owned attention. Then the checkout owned the transaction. Now the negotiation owns the price. Every time, the layer that looked like plumbing turned out to be the layer where the money gets decided, and the party doing the deciding was never the party doing the work.
Two things nobody is pricing.
One. A concession doesn’t expire. It becomes a row.
A human buyer you discounted for in March has forgotten by June. A table hasn’t. Every fifteen percent you gave away is an example of a seller who moves when asked politely. The next agent arrives already knowing that, and the one after arrives knowing it cheaper. You’re not negotiating with a person who might forget. You’re negotiating with a record that gets better every time you say yes.
Two. Your inconsistency is the invitation.
Same paper, different finding: when prices are spread out, the buyer’s agent gains more by searching harder. Read that from the seller’s side. The wider your quotes range, the more money there is in shopping you. Every friend price, every Friday-afternoon number, every “don’t tell anyone” — that’s not goodwill. That’s variance somebody’s model is going to harvest.
And this runs in both directions, which is what makes it a market story and not just a buyer story. Calvano and three co-authors, American Economic Review, 2020: pricing algorithms given nothing but a profit target learned to hold prices high, without being told to and without talking to each other. Assad and co-authors, looking at German retail petrol stations: the ones running pricing software earned margins around fifteen percent higher than the ones pricing by hand.
Put the two halves together. Machines on the sell side learn to hold the line. Machines on the buy side learn to push. And the only party in the room with no memory and no model is the human being — who is also the only party who has to sign.
There’s a second cost, quieter than the margin.
The record of that negotiation lives in their system. If an order gets disputed — by a bank, by a client, by a lawyer — the buyer’s agent can produce a transcript with timestamps, a price cap, and the exact words that closed it. What do you have? An email thread and a memory of how the call felt.
That’s not a fair fight. That’s a filing cabinet against a database.
So what do you actually do, starting with this week’s quotes?
Write the mandate before you write the price. Floor. What you’ll trade. What you’ll never trade. Then hand the mandate to whatever is quoting on your behalf, human or not. An agent given a goal will invent a discount. An agent given a mandate has to come back and ask.
Quote a rule, not a mood. A rate card with the conditions for movement written on it. Machines can’t charm a rule. They can only test it, and a test has an answer.
One price, or a documented reason it isn’t. If your number changes, the reason has to be written down and defensible — volume, term, scope, risk. Variance you can’t explain is variance someone else will use.
Trade the discount. Never donate it. Fifteen percent buys ninety days of committed volume or a fifty percent deposit. In writing, both sides. A discount with nothing attached is a gift, and gifts get catalogued.
Own the record. The confirmation comes from your system, in your format, with your timestamp. If you can’t produce the agreed terms in thirty seconds, you don’t have an agreement. You have a recollection.
And audit your own buying agent. Whatever it gives away, it gives away in your name, all day, with no note attached.
One last thing, because it’s bigger than any single deal.
A market can lose its floor without anyone deciding to cut. No meeting. No memo. No price war. Just cheaper search. When the cost of asking a hundred sellers falls to almost nothing, the floor doesn’t get defended — it gets discovered, downward, one polite email at a time.
Which means the businesses that hold aren’t the cheapest ones. They’re the ones whose price has a reason attached that a machine can read and a customer can repeat back to you.
The principle that has to survive isn’t romantic. It’s the right to a price you can explain.
Not the right to win, or the right to be chosen. The right to know what your yes cost, and to be the one who decided it — instead of the last person in the chain to find out.
Set the rule first. The machine will meet you at it.