Negotiated Commerce

Summary. AI agents are becoming the way economic actors transact. Agents do not displace the published price, which remains the backbone of trade; they collapse the cost of individualised term formation, which for two centuries reserved negotiation for the largest transactions. This paper defines Negotiated Commerce as a category, distinguishes it from the mechanisms it will be mistaken for, grounds it in the economics of transaction costs and bilateral bargaining, sets out its design principles and reference architecture, and states plainly what is demonstrated today, what is specified, and what remains proposed. It accompanies two other documents: a manifesto stating the position, and a technical specification (the Negotiation Extension for Agentic Commerce) defining the protocol. The category, the protocol and the company are distinct throughout: Negotiated Commerce is the category; the Negotiation Extension is an open protocol proposal; Bidit is the reference implementation and operator of the first negotiability index.

Definition. Negotiated Commerce is the formation of the terms of a transaction between two economic agents, in private, under rules each principal controls, executed in the moment and honoured on the rails where the transaction settles.

Every word carries weight, and the sections below exist to justify each: terms rather than price, because terms include value as well as a number; agents because software now represents the principals; in private because publication is what makes flexible terms dangerous; rules each principal controls because neither side's agent is trusted with the other's limits; executed in the moment because agents collapse term formation from days to milliseconds; honoured where the transaction settles because an agreement that the checkout does not recognise is a claim, not an agreement.

Fixed-price commerce publishes a price. Negotiated Commerce allows terms to be formed between buyer and seller. It is not a new way to set a price. It is a new way to form terms.

1. The shift underway

Commerce is being rebuilt around agents, in public, by the largest companies in the world. Shopify has opened its global product catalogue to AI agents and shipped a commerce protocol for them. OpenAI and Google are building agent checkout. Mastercard is building payment credentials that let an agent pay on a principal's behalf. Today these rails carry consumer purchases; the pattern they implement, an authorised agent discovering, deciding and committing, is general.

Two consequences follow. First, anything published is now read by machines continuously: prices, rates, plans, promotions. Second, the historical reason for publishing one fixed set of terms for everyone, that no organisation could afford to negotiate individually at scale, no longer holds. What machines still lack is a lawful, private, rule-bound way to form terms. That gap is the subject of this paper.

2. What Negotiated Commerce is not

The category will be mistaken for its neighbours unless the boundaries are drawn at the outset.

• Dynamic pricing is the seller changing the price, alone, for everyone or for segments. One will, one number.
• Personalisation is the seller predicting what a particular buyer might accept, from data the buyer often never knowingly provided. One will, informed by surveillance.
• Discounting is the seller lowering published terms, publicly, for all potential buyers at once.
• Auctioning is many buyers competing against each other under an allocation mechanism the seller controls.
• Negotiated Commerce is two authorised economic actors forming mutually acceptable terms, each within constraints they independently control, in private, with a committed outcome.

The distinction that matters most is the second will. In every neighbouring mechanism, one party sets and the other takes. In negotiation, both parties shape, and neither party's limits are visible to the other. Descriptions of this model as "AI dynamic pricing", "personalised discounts" or "an offer engine" are not simplifications; they are category errors, because each names a one-willed mechanism.

3. The economics: what agents actually changed

The economic case does not rest on machines being clever. It rests on machines being cheap, and on a body of economics that predates them by a century.

Published terms are a transaction-cost institution. Coase (1937) established that the structure of economic activity is shaped by the cost of transacting, and Williamson's transaction-cost economics extended the point: institutions exist to economise on the costs of forming and enforcing agreements. The posted price is exactly such an institution. Individualised term formation, discovering a counterparty's willingness, bargaining, committing, enforcing, was historically so expensive per transaction that it survived only where transaction sizes justified it: property, wholesale, procurement, employment. Everywhere else, sellers adopted the cheapest available mechanism, one price for all, and accepted its known inefficiency. The published price was never the optimal mechanism. It was the optimal mechanism net of transaction costs.

Software agents collapse that cost toward zero. A negotiation that required a salaried human now requires milliseconds of compute. When the cost term that justified an institution disappears, the institution's domain shrinks. That is the entire structural claim of this paper, and it is a more defensible one than "machines can negotiate": the binding constraint was never willingness, it was cost.

A single price forgoes known value. The economics of price discrimination, from Pigou (1920) onward, establishes that a single price excludes buyers whose reservation price, the most they would pay, lies between cost and the posted price, while under-charging buyers whose reservation price is higher. These are deadweight loss and forgone surplus respectively, and both are invisible in a fixed-price ledger. Illustration: a television ticketed at $2,995; a buyer whose reservation price is $2,500 walks, and no record of the lost sale exists. The classical remedies are all one-willed: segment pricing, personalised pricing, public sales. The public sale in particular reaches the marginal buyer only by re-pricing every buyer, including those happy at the ticket, and by signalling the new level to every competitor. Bilateral negotiation within a seller-controlled floor is the two-willed alternative: the marginal buyer is served near their reservation price, the ready buyer pays the ticket, and no public number moves.

Bargaining theory says the surplus is there to split. Nash (1950) formalised what every market trader knew: when a buyer's reservation price exceeds a seller's floor, a zone of agreement exists, and bargaining is the mechanism that locates a point in it. Myerson and Satterthwaite (1983) proved the sobering half: no mechanism can guarantee efficient bilateral trade when both sides hold private valuations, some mutually beneficial trades will always be missed. This is a reason for humility, not despair: the relevant comparison is not negotiation versus a perfect mechanism, but negotiation versus a posted price that forgoes the entire zone of agreement below the ticket by construction.

Information is the by-product. Stigler (1961) framed markets as information problems. Every offer in a negotiated system, won or lost, is a direct observation of a real buyer's willingness to pay for a specific item at a specific moment: demand data that no fixed-price ledger contains, gathered as a side effect of trading. A shelf is silent about everyone who walks past it.

And privacy is the enabling condition. Sellers avoid flexible pricing not because it is unprofitable but because public price moves are permanent, visible and contagious; price-matching policies mechanically propagate one seller's public cut to others, and machine-read markets accelerate whatever propagation exists. Where terms are formed privately, per transaction, none of that machinery engages. Privacy is what makes flexibility safe. (The pace of these dynamics in fully agent-mediated markets is asserted here as a mechanism, not yet as measured fact; the markets are too young. The direction of the mechanism is not in serious doubt.)

4. Why this needs a protocol, not a conversation

Negotiation already happens without any standard: in messages, in email, at counters. If people can form terms in conversation, why can't machines? Because in a human negotiation, everything the words leave out is carried by trust between people. Machines carry none of it, and each missing piece must be engineered or it is absent.

A conversational agreement is a claim, not a commitment. When a chatbot says "we can do $2,600", nothing downstream honours the sentence; the checkout has never heard of it. This is documented, not hypothetical: a tribunal has already ordered an airline to honour terms its chatbot invented in conversation. A protocol replaces the sentence with a commitment object bound to product, quantity, price and expiry, which the settlement rail recognises mechanically.

Language leaks; a protocol seals. A seller's floor is the crown of its position, and a conversational AI holding it can be probed and pressured toward revealing it. In protocol, the floor is not a secret the agent keeps; it is a number the counterparty-facing surface never possesses. The engine holds it and answers only accept, counter or decline. What is structurally absent cannot be extracted.

Machines do not tire, so limits must be laws. An unconstrained agent will ask ten thousand times, shaving cents. Attempt limits, session accounting, the closure of a negotiation on acceptance, and cooling periods after final declines must be enforced by the engine, not left to etiquette.

Ambiguity scales into disputes. "Done, with delivery" has one meaning between two people in context and many meanings across a million machine transactions. Protocol verbs are exact: a value ask is answered per item, yes, or at a stated price, or no; a counter carries one number; acceptance carries one commitment.

Authority and identity must be checkable. How does a seller's engine know a buyer's agent is authorised to commit its principal, and how does a buyer's agent know the endpoint speaks for the seller? Conversations cannot answer; the protocol inherits the answer from the commerce rails it extends: published profiles, declared capabilities, verifiable endpoints.

Records resolve; transcripts argue. A protocol negotiation leaves a precise history of offer, counter, agreement and expiry. A conversational one leaves prose to interpret.

Without a standard, none of it connects. A thousand sellers solving these problems privately means a thousand bespoke integrations per agent. Payments moved at scale only when they became protocol; documents became an economy only when requests became standard. One open grammar for term formation, and every agent can negotiate with every willing seller.

The conversation is the right model for the negotiation. The protocol is what makes the conversation executable between parties who have never met and are made of software.

5. Design principles

Six principles define a conforming Negotiated Commerce system. They are stated as canon because every architectural decision in section 6 descends from them.

1. Private. Agreed terms are known to the parties and no one else: never published, never indexed, never fed to comparison.
2. Floor-protected. Each principal's limits, on price and on value concessions, are held by their own engine and never exposed to the counterparty in any response, error or hint.
3. Bounded. Attempts are limited and enforced by the engine, never trusted to the counterparty's agent; agreement closes the negotiation.
4. Instant. Every proposal receives an immediate, definitive answer: accepted, countered, declined, or blocked.
5. Honoured. Agreement produces a commitment redeemed where the transaction natively settles; the seller remains seller of record and the negotiation layer never holds funds.
6. Honest. Negotiability is never overstated at discovery, and agents relay outcomes truthfully.

6. Reference architecture

The architecture has six stations, and the paper defines what crosses each boundary. The full wire-level detail belongs to the technical specification; this section fixes the concepts.

Principal → Agent. A principal (a shopper, a company, a machine's owner) grants its agent authority: the scope within which the agent may commit, a budget ceiling, a product scope, an expiry. Authority is declared, not assumed; a conforming engine may verify it before treating an offer as binding.

Agent → Discovery. The agent resolves the counterparty's negotiability answer before promising anything: from the counterparty's own published capability profile where one exists, else from a negotiability index. The answer has exactly two values, yes with the applicable rules (scope, attempt limits, counter behaviour), or not yet, optionally with a consent-gated intent-registration recipe.

Agent → Negotiation Engine. Within a session, a stable identifier binding one principal's conversation so attempts are counted honestly, the agent submits an offer: item, quantity, consideration. The engine resolves it under the seller's policy into one of four outcomes; a counter carries exactly one number; after agreement, value asks are answered item by item. All limits live server-side with the engine.

Engine → Commitment. Agreement mints the central object of the architecture:

The Negotiated Commitment — conceptually: issuer (the engine, on behalf of the seller); counterparty (the principal, via its agent's authority); item and variant; quantity; consideration (the agreed price and currency); negotiated inclusions (the value terms: delivery, warranty, extras); expiry; redemption conditions (where and how it may be settled, and that it is single-use); a session and nonce binding it to one negotiation to prevent replay; and integrity protection, so the settlement rail can verify it was issued by the engine and has not been altered. The specification, not this paper, fixes the exact fields and their encoding.

The commitment answers the questions a technical reader should ask. What if inventory changes before redemption? The commitment binds a variant and fails safe: if the item is gone, the commitment is unredeemable and the negotiation's closure stands; it is a right to a price, not a reservation of stock, unless the seller's policy says otherwise. What if it expires? It lapses; nothing renews silently. What prevents replay or transfer? The nonce, the single-use redemption condition, and the binding to the issuing session. Who is trusted? The settlement rail trusts the engine's integrity protection; the buyer trusts nothing but the outcome, since the transaction completes at the seller's own checkout.

Commitment → Settlement rail. Redemption happens where the transaction natively completes, a retail checkout, a booking system, a procurement or settlement platform, on existing payment rails. In today's retail implementation the commitment travels as a discount application scoped to the variant; in agent-checkout protocols it can travel natively. The negotiation layer's involvement ends at handoff.

7. What is demonstrated, what is specified, what is proposed

Credibility requires this section to be exact.

Demonstrated. In retail: a policy-controlled engine negotiating real products at real retailers, in production with 23 merchants at publication. Negotiation is on price today; value asks (delivery, warranty, inclusions) are specified in the protocol and ship in the engine's next release. A negotiability index covering more than one thousand retailers across twenty countries, honouring the two-answer rule; consent-gated intent registration with aggregate-only reporting. This is evidence that the architecture operates in retail. It is claimed as nothing more.

Specified. The Negotiation Extension for Agentic Commerce, an open technical specification of the verbs, outcomes, commitment and conformance rules, expressed in the capability grammar of the emerging agentic commerce protocols. It publishes the day its hosted reference endpoint is live, so that the first developer who calls it gets a real answer. It is a proposal. The market, not the publisher, determines whether a proposal becomes a standard.

Proposed. The extension of the same primitives to travel and hospitality (perishable inventory offered privately), business buying (volume, timing and service under policy), home services and big-ticket categories, energy (small frequent trades between distributed resources and the grid), compute (capacity cleared between agents under provider floors), and ultimately machine-to-machine commerce. The primitives, discovery, proposal, resolution, commitment, settlement, are domain-agnostic by construction, and that is precisely why they are proposed for these domains; they are not demonstrated there, and this paper does not claim otherwise.

8. Relationship to the emerging infrastructure

Negotiated Commerce is deliberately complementary to the agentic commerce stack now being built. Catalogue and discovery protocols answer what exists; checkout and payment protocols answer how it settles; agent-payment credentials answer who may pay. None answers how terms are formed when the published terms are only an opening position. The Negotiation Extension is written as an extension to that stack, not a rival to any layer of it, and the negotiation layer's refusal to hold funds is what keeps it compatible with every payment network simultaneously. The negotiability index, likewise, is neutral infrastructure: its value depends on every platform being able to trust it, which is why it is exclusive to none and why its honesty rule is absolute.

9. The rules that keep it honest

Stated plainly because a conforming implementation enforces them in code: the floor is never disclosed; agreed terms are never published; discovery never overstates; attempt limits bind the agent; consent is collected before contact and enforced at storage; registered contact details are never publicly readable; counterparties see aggregates only and can correct their own index entry. A negotiation layer that abandons these becomes surveillance or becomes discounting. The category is defined by being neither.

10. Limitations and the adoption path

Honest limits, so the claims above keep their weight. Bilateral negotiation cannot capture every efficient trade; Myerson and Satterthwaite guarantee that. Sellers must set floors well, a policy skill the tooling can assist but not replace. Buyers' agents must adopt discovery-before-promise as a norm, which no publisher can force. The economics of agent-mediated market dynamics are argued here from mechanism, not yet from longitudinal data, because the markets are months old. And a specification with one reference implementation is a proposal: it earns standardhood only through independent implementations, critique and revision.

The adoption path follows from the limits: publish openly; operate the reference implementation and index in the proof market; invite implementations, criticism and extension from platforms, merchants, agent developers, payment networks and standards bodies; and let counterparties progressively declare negotiability in their own published profiles, with the index as the market-wide answer and archive of record.

An invitation. Sellers: your index entry is yours to correct, and participation takes a day. Platforms and networks: the specification is open and the reference implementation is live. Builders and critics: implement it, break it, improve it. Correspondence, corrections and implementations: registry@bidit.com.au.

Negotiated Commerce is not a company and not a feature. It is the term-formation layer the agentic economy is missing, proposed openly, demonstrated where agents trade today, and offered for the market to make its own. The protocol is open. Bidit enables it.

References. Coase, R. (1937), The Nature of the Firm. Pigou, A.C. (1920), The Economics of Welfare. Nash, J. (1950), The Bargaining Problem. Stigler, G. (1961), The Economics of Information. Myerson, R. & Satterthwaite, M. (1983), Efficient Mechanisms for Bilateral Trading. Williamson, O. (1979), Transaction-Cost Economics: The Governance of Contractual Relations.

Bidit — registry.bidit.com.au · registry@bidit.com.au