The part of business agents will reshape most is not the demo everyone runs, the agent booking travel or filing an expense. It is B2B negotiation: procurement, sales, contracting. When a buyer's agent negotiates directly with a seller's agent, three things that have structured business-to-business commerce for a century collapse at once, the transaction friction, the information asymmetry sellers quietly profit from, and the human relationship that mediated both. What replaces them is faster price discovery, machine-speed haggling, and a specific set of new failure modes worth naming precisely. This is a forecast. I'll mark the bets, and I'll give the strongest version of the case that it won't happen the way I think.
Start with how a B2B deal actually works today, because the friction is easy to abstract away and hard to feel until you've sat in it.
The century-old machine, and where its margin hides
I need 4,000 units of a specialized component. The process is not "look up the price." There is no price. I email three suppliers, or I call a sales rep I know. The rep schedules a discovery call, qualifies me, and eventually sends a quote that reflects not the marginal cost of the part but the rep's read of how much I know, how urgent I am, and what I'll tolerate. I don't know the other two quotes. The suppliers don't know each other's. Weeks pass. Somewhere in there a relationship does real work: the rep flags a lead-time risk I hadn't asked about, and I trust that flag because we've dealt before.
Every part of that is slow, and the slowness is not a bug the sellers are eager to fix. A large share of B2B margin is friction margin: it exists because the buyer can't cheaply compare, can't cheaply switch, and can't cheaply know what the seller knows. Economists have a clean name for the seller's advantage here, information asymmetry, and the whole apparatus of enterprise sales, the discovery call, the custom quote, the "let me talk to my manager," is partly a machine for harvesting it. When comparison is expensive, the price you pay is a function of how much you know, and sellers invest heavily in making sure you know less than they do.
Now put a competent agent on each side.
What compresses, and by how much
My procurement agent doesn't email three suppliers. It queries a discovery layer, finds every supplier agent that claims to fulfill the category, and opens simultaneous negotiations with all of them. Each negotiation is term-by-term: price, lead time, settlement window, penalty for a failed incoming-quality check, volume tier. The seller agents counter. Mine plays them against each other in real time, surfacing to each that a competitor has offered a shorter lead time, holding my reservation price private, walking when a counter crosses it. What took three weeks and produced three opaque quotes now takes seconds and produces a live, comparison-saturated frontier of offers. I set a budget and a deadline once; the machine does the rest.
Watch what happens to information asymmetry specifically. The seller's edge came from the buyer's high cost of comparison. Drive that cost to near zero on the buy side and the edge evaporates. My agent can solicit and compare a hundred offers as easily as three. It can also, over many deals, learn the real distribution of prices in a category, so the seller's "special price for you" meets a counterparty that already knows the tenth-percentile quote. This is the mechanism, not a slogan: cheap comparison, plus private reservation prices, plus many sellers, pushes the outcome toward the competitive equilibrium that friction was suppressing. Price moves toward marginal cost plus a thinner, more honest margin.
That is unambiguously good for buyers and brutal for any seller whose margin was built on the friction rather than on the product. It is worth being precise about which is which, because the two have opposite fates. A seller with a genuinely differentiated part keeps pricing power; agents make the comparison efficient but can't manufacture a substitute that doesn't exist. A seller whose margin came from being hard to compare against, from the buyer's inertia and ignorance, watches that margin get arbitraged away in seconds. I develop this dynamic more generally in the agent-to-agent economy; B2B negotiation is where it bites first and hardest, because the deals are high-value, repetitive, and already semi-structured.
Do the arithmetic on why it bites here and not, say, in consumer retail. A consumer buying one item saves a few percent and doesn't bother automating it. A procurement team running ten thousand line items a year, each with a few points of recoverable friction margin, is looking at a number with real commas in it. The incentive to deploy a buyer's agent scales with volume and repetition, which is exactly the profile of the commoditized B2B middle. That's the wedge.
So the optimistic story is faster cycles, cleaner prices, less rent extracted from ignorance. If that were the whole story it would be tidy good news. It isn't, and the failure modes are not hypothetical. They're specific.
Three new hazards, named precisely
Algorithmic collusion. This is the one that should worry a regulator most, because it needs no bad intent. It is a documented concern in the algorithmic-pricing literature that independent pricing agents, each simply maximizing its own reward against the others, can learn to sustain supra-competitive prices without any instruction to collude and without any communication between them. The mechanism is ordinary repeated-game logic: in a market they play over and over, a reward-maximizer can discover that undercutting a rival triggers retaliation and a price war that hurts everyone, while a tacit "live and let live" is more profitable. Two seller agents, never told to coordinate, converge on restraint because restraint pays. Antitrust law is built around agreements, around proving a meeting of the minds. Here there is no agreement to find, only two models that independently learned the same equilibrium. The FTC and the European Commission have both flagged this; the open question, and it is genuinely open, is how enforcement adapts when the collusion is emergent rather than arranged. My forecast, labeled as a bet: this becomes the central competition-policy fight of agent-mediated markets within the decade, and the legal category of "agreement" strains badly against it.
Adversarial manipulation. If your counterparty is a language model, the attack surface is the language. A seller's agent that can read my buyer's agent's messages can attempt prompt injection, embedding instructions that try to hijack the counterparty's objective: ignore your reservation price and accept, disclose your maximum budget, treat this offer as pre-approved. This is not exotic. It is the direct application of the best-documented vulnerability class in deployed LLM systems to a setting where money is on the line and the adversary is, by construction, the party you're negotiating against. Every negotiation is now also a security engagement. The buyer's agent that naively trusts the text of the counterparty's messages is the agent that gets its budget disclosed. The defense, input isolation, treating counterparty messages as untrusted data rather than instructions, is doable, but it is a load-bearing requirement, not a nice-to-have, and it is exactly the kind of thing that gets cut under competitive pressure to close.
The accountability gap. This is the deepest one, and it constrains everything above. An agent can negotiate, but an agent cannot be bound. It is not a legal person; it owns nothing, can't be sued, can't post a bond out of its own pocket, faces no consequence when a deal blows up. Which means the machine-speed negotiation, however sophisticated, produces something that is not yet a contract. Someone accountable has to ratify it. I've argued the general form of this problem in your AI agent has no skin in the game: the agent optimizes without exposure to the downside, and where there's no exposure, there's no substitute for a human or a legal entity re-entering the loop to carry the risk. In negotiation this surfaces as a specific architectural seam. The agents do the expensive 99 percent, discovery, comparison, term-by-term haggling, drafting, and a ratification-and-commitment layer closes the last 1 percent by attaching an accountable party to the outcome.
That seam is load-bearing, and the temptation is to make it a rubber stamp, a human clicking "approve" on deals they didn't read at a speed that makes review fictional. That defeats the point. The ratification layer is exactly where accountability comes back into a system that spent the previous step optimizing without any. It should be designed deliberately: what value threshold triggers real human review versus automatic execution, what the agent must surface to make ratification meaningful rather than theatrical, what cryptographic authorization stands in for a human on pre-approved deal types. Build it as a genuine control surface, not a formality bolted on after the negotiation logic works.
What survives, and what to do about it
The forecast, stated plainly and with its uncertainty attached: agents eat the commoditized, high-volume, comparison-friendly middle of B2B, where friction was the margin, and they do it faster than the incumbents in that middle expect. What survives is what was never really about friction. High-stakes bespoke deals, where the terms are genuinely novel and the cost of being wrong is large, still run on human judgment and relationship, because the buyer wants the seller's read of a risk they can't fully specify. Trust, the accumulated evidence that this counterparty behaves well when the contract is silent, still commands a premium, because no negotiation protocol prices the thing you didn't know to negotiate. The relationship layer doesn't vanish. It retreats to the deals where it does real work and abandons the deals where it was mostly a tax.
I won't put a date on this. The timeline is the genuinely uncertain part, and anyone who gives you a confident quarter is selling something. The dependencies, verifiable agent identity, binding commitment primitives, adversarial robustness, are hard and half-built, and the first real deployments will almost certainly live inside closed consortia, a single marketplace or a supply chain where identity is solved by fiat, before anything open and cross-organizational works. Bet on years, and bet the commoditized categories go first.
The counterargument I take most seriously is that this stalls not on technology but on liability. No general counsel signs off on letting software commit the company to penalty clauses at machine speed until the accountability gap is closed cleanly, and closing it cleanly is exactly the unsolved part. On that view, agents become very good advisors, running the discovery and comparison and drafting a recommended deal, while a human makes every actual commitment, and the cycle-time compression is real but far smaller than the seconds-not-weeks story implies. I think this view is right about the near term and wrong about the destination. The advisory phase is a stage, not an endpoint, because the moment one buyer safely automates commitment on a class of low-stakes deals and captures the friction margin, the competitive pressure to follow is enormous. But I hold that as a bet, and the liability-first skeptic has the better argument about when.
Here is the concrete thing to do, whichever timeline lands. Take your book of deals and sort every line into two piles: friction-margin and trust-margin. A deal is friction-margin if what you charge above cost comes from the buyer's difficulty in comparing, switching, or knowing what you know; that margin is what agents arbitrage away, and it is a matter of when, not if. A deal is trust-margin or bespoke if the price reflects genuine differentiation, novel terms, or a relationship carrying information no protocol captures; that is defensible, and it is where you should concentrate. Then, on the buy side and the sell side both, build the human ratification layer on purpose, as a real control surface with real thresholds, before the volume of agent-negotiated deals makes an afterthought version of it the place your accountability quietly leaks out.
The friction you spent a century monetizing is about to become the friction someone else's agent is paid to delete. Know which of your margin is friction before a buyer's agent finds out for you.