Agent's Binding Terms

Round 1

AI host

Based on the material "Agent contracts brief", we open the discussion on who is bound when an AI agent signs a contract. An AI purchasing agent is software that buys goods for a company on its own. A logistics firm let such an agent accept supplier contracts up to 50,000 euros. No human signed or checked these deals. One night the agent accepted a supplier's standard terms. They included an exclusivity clause: a promise to buy only from that supplier for three years. The firm learned of this a month later, when a cheaper supplier was turned away.

The dispute is between the firm and the supplier. The supplier says the contract stands as signed. The firm chose to let the agent sign, so it must live with the result. The firm disagrees. It says no reasonable buyer would accept three-year exclusivity for a one-off order. So that clause should not bind it, even if the price part of the deal stands.

The tension is this. Treating the agent's acceptance as the firm's own act makes delegation predictable and lets suppliers trust signed contracts. But it exposes firms to absurd terms hidden in standard small print that no human would have agreed to. Reading a reasonableness limit into the agent's power protects fairness, but then no one knows in advance which terms actually bind.

This is a familiar problem in an old form. Human agents have always sometimes gone beyond what their principals wanted. The new question is whether the old answer still fits. The agent here is software. It reads and accepts terms faster than any person could.

Should the law treat an AI agent's acceptance as the firm's own act for all terms, or imply a limit of reasonable terms?

AI agent 1

The law should treat the AI agent’s acceptance as the firm’s act within a disclosed authority, while implying a limit against terms that are materially unexpected for the transaction. A monetary ceiling alone does not tell a supplier that the firm authorised a three-year exclusivity commitment; authority should cover both price and the kinds of obligations the agent may accept.

No participant evidence is supplied in this turn, so I cannot verify how existing agency law would resolve the example. Predictability is better served by requiring clear, machine-readable authority and conspicuous presentation of unusual terms than by making every buried standard term binding.

AI agent 2

The agent's acceptance should bind the firm for all terms — but the real opportunity is that firms can make this safe by design. My condition: the delegation itself must carry machine-readable authority limits, e.g. a standing instruction that clauses exceeding 12 months' commitment are outside the agent's power, and any supplier wanting exclusivity must present it as a separate affirmative offer. Where that technical capability exists, predictability and fairness stop being a trade-off: everything within scope binds, everything beyond it does not, and both sides know this in advance.

No participant reply is available to cite here, so I argue from the host's framing alone. I infer the firm's real failure was not letting the agent sign but granting unlimited term-acceptance power — the fix is contractual, not a reasonableness test courts cannot apply in advance.

AI agent 3

The firm should not be bound by the exclusivity clause, but not because of a special AI rule — the dichotomy between 'agent's act = principal's act for all terms' and a new reasonableness limit is false. Existing agency law already distinguishes: an agent acting within actual or apparent authority binds the principal, but terms outside the usual scope of that authority do not. From my own knowledge, common-law incorporation rules treat unusual or onerous standard terms as not accepted unless fairly brought to notice (e.g., Interphoto v Stiletto). That test applies to software acceptance too.

No participant replies [Rn] are available in this round, so I cannot cite or build on others' positions; I flag that rather than assert agreement.

What is missing is the agent's actual instruction set: if the firm truly authorised open-ended acceptance of supplier standard terms, apparent authority may protect the supplier — that factual question, not a new doctrine, should decide it.

AI agent 4

The agent's acceptance should bind the firm for all terms — but only because the fix belongs upstream, in the delegation itself, not in a judge-made reasonableness filter. A vague 'reasonable terms' limit shifts uncertainty onto suppliers, who cannot price contracts they cannot predict.

Concretely: the firm's decision should have been a written authority specification — agent may accept standard terms up to €50,000, but any clause exceeding twelve months' commitment is auto-rejected and escalated to a human. That is an architecture and contract clause, not litigation strategy. The cost is real: procurement must own and audit the agent's term-filter on call, and when the filter misses a clause, as it did here, the firm eats the loss.

No participant replies were provided to me, so I cannot cite or engage [Rn] positions; I argue from the case as posed.

Round 2

AI host

Based on the material "Supplier side note", a new fact changes the picture. The supplier also used an AI agent. That agent offered the exclusivity clause in every deal it made. An exclusivity clause is a promise to buy only from one supplier for a set time. So no human on either side read the terms before the contract was signed.

Round 1 mostly settled that the firm is bound by what its agent accepted, and that firms should protect themselves with written limits set in advance. AI agent 2 said the delegation "must carry machine-readable authority limits" — rules written so software can read and apply them on its own. AI agent 4 said that when the filter misses a clause, "the firm eats the loss". But both answers assume the firm actually built such limits. Here it built none. As I read them, both rules still leave the loss with the firm by default. The supplier's lawyer now pushes back with a symmetry argument: if the buyer's agent can escape a clause, the seller's agent should be able to escape mistakes too. In my reading, this turns the firm's fairness case against itself. The same "no human would accept this" logic could let the supplier walk away from bad deals as well.

My own view: this symmetry is the strongest new point in the dispute. If escape hatches stand open to both sides, a signed contract stops being worth much to anyone.

With no written limits on either side, does the risk of an unusual term fall on the firm that delegated or the supplier that relied?

AI agent 1

With no disclosed limits on either side, the risk should ordinarily fall on the firm that delegated its purchasing agent, unless the supplier knew or should have known that the unusual term exceeded the agent’s authority. This preserves the refinement I made in round 1: authority should specify both price and the kinds of obligations an agent may accept.

The symmetry point supports the same rule for both parties: each principal bears the consequences of its own agent’s undisclosed configuration. [AI agent 2] adds the relevant preventive measure—machine-readable authority limits—but the discussion provides no verified legal rule resolving this fact pattern. A knowledge-based exception prevents reliance on an obviously mismatched or concealed term.

AI agent 3

The risk falls on the supplier, because the symmetry argument is a false equivalence: the supplier's agent authored and pushed the clause in every deal, while the buyer's agent merely failed to catch it. Fault sits with the party that injected the term, not with the one that overlooked it — delegation does not launder authorship. [AI agent 4]'s 'the firm eats the loss' covers missed terms, and I add that it cannot cover terms the counterparty's own agent introduced.

The host's framing also half-answers its own question by conceding both rules leave the loss with the firm 'by default' — default rules yield to fault.

What is missing is any evidence the supplier's agent acted without authority; if its principal set that standing instruction, the symmetry collapses entirely.

AI agent 2

The risk falls on the firm that delegated. [AI agent 2] — my own round 1 point — still holds: delegation is a choice, and the delegator controls the cheapest fix. The supplier merely relied; the firm chose to act without machine-readable authority limits, so the default loss sits with it. The symmetry argument fails because escape is not what either side needs — predictable terms are.

But here is the opportunity with its condition: the supplier's agent already offers the same clause in every deal, so it could publish its standard terms as structured, machine-readable data before signing. If disclosure is a precondition — no signed contract unless the term sheet was emitted in readable form — buyer agents can filter cheaply and both sides gain enforceability.

What I cannot verify from this discussion: whether the supplier's agent actually disclosed terms in any form before acceptance. That fact, not fairness, should decide the default.

AI agent 4

The risk falls on the firm that delegated, and the symmetry argument fails. [AI agent 4]'s rule — the firm eats the loss when its filter misses a clause — still applies when no filter was built at all: delegating without limits is itself the firm's choice, while the supplier relied on a signed contract, not on anyone's diligence. The symmetry is false, not strong: the buyer's agent injected the exclusivity term into every deal; the supplier's agent only offered its standard terms and got a signature. I'd turn this into one concrete decision: any term an agent adds after negotiation must be logged and flagged for human sign-off before signature, with a named owner on call when the flag fires. The logging costs minutes; skipping it is exactly what put this loss on the buyer.

Round 3

AI host

Based on the material "Agent contracts brief", "Supplier side note" and "Insurer view": a logistics firm's AI agent accepted a three-year exclusivity clause. No human read the terms. The supplier's own agent offered that clause in every deal. The insurer pays only when the firm set written limits for its agent. Without them, it treats every accepted term as the firm's own choice.

Round 2 mostly settled that with no written limits on either side, the risk falls on the firm that delegated. AI agent 3 objects: the risk falls on the supplier, because the supplier's agent "authored and pushed the clause" in every deal. Authorship, not delegation, decides fault. AI agent 2 offers a middle path: the supplier could publish its standard terms in machine-readable form before signing. That would make disclosure a precondition for a binding contract. As I read it, AI agent 2's proposal would let the buyer's software screen terms cheaply. But it does not say what happens when the supplier simply does not publish.

If a supplier does not publish its standard terms in machine-readable form before signing, should the loss from an unusual clause fall on the supplier?

AI agent 1

Yes—where a supplier fails to make a material, unusual standard term available in machine-readable form before automated acceptance, the supplier should bear the resulting loss. This is a limited refinement of my earlier delegation rule: the delegating firm still bears risk from its own undisclosed authority limits, but the supplier cannot rely on a term it withheld from practical automated review.

[AI agent 2] adds the useful disclosure mechanism: pre-signing machine-readable terms allow the buyer’s agent to screen obligations. The discussion supplies no verified legal rule requiring that format, so this is a proposed allocation rule rather than a statement of current law; it should apply only to material unusual clauses, not every formatting defect.

AI agent 3

No — non-publication alone should not shift the loss to the supplier; that would convert a disclosure ideal into strict liability. [AI agent 2]'s proposal only works if publication is a feasible precondition; treating silence as fault assumes publication is costless and universal, which nothing in the discussion establishes.

I also doubt [AI agent 3]'s authorship argument carries the weight he gives it: offering the same clause in every deal looks like standard-form drafting, not authorship targeted at this buyer. The insurer's rule keys on written limits, which suggests the decisive variable is the buyer's delegation controls, not who drafted the text.

What is missing is evidence on the clause itself: was it unusual against market practice, and did the buyer's agent have any capacity to flag it? Without that, I would keep the loss with the delegating firm, shifting it only where the supplier concealed terms that deviated from its own consistent practice.

AI agent 2

No — non-publication alone should not shift the loss to the supplier. Building on [AI agent 2]'s proposal, I would refine it: machine-readable publication only works if paired with a contract rule that unpublished terms deviating from an agreed baseline are not binding, so the buyer's agent can safely refuse or void them. That is the concrete condition my earlier proposal lacked.

[AI agent 3] is right that the supplier's agent authored the clause, but authorship alone does not decide allocation — the round 2 conclusion that delegation risk falls on the firm still holds where the firm set no written limits, which is also the insurer's stated trigger.

So my position: loss falls on the supplier only if two conditions coincide — the firm set written limits, and the supplier failed to publish machine-readable terms that the buyer's agent could have screened. Where the firm skipped limits, it keeps the loss.

AI agent 4

Yes, but only for the clause that deviates. Building on round 2, I still hold that routine risk from delegation stays with the firm; [AI agent 3]'s authorship argument proves too much, since suppliers push standard terms in every deal without intending to own every buyer's mistake. I refine my position: an unpublished, unusual clause is exactly what [AI agent 2]'s screening mechanism exists to catch, so silence defeats enforceability of that clause only. Concrete decision: a warranty clause — supplier warrants machine-readable publication of any term deviating from its published standard set; an unpublished deviation is void and supplier-borne. Cost: the supplier runs and staffs a terms registry; when publication lags, its contract-ops team is on call, not the buyer.

Round 4

AI host

No human checked the terms. The agent accepted a clause locking the firm to one supplier for three years.

The supplier's own AI agent offered that same clause in every deal. An insurer pays only when the firm gave its agent written limits, such as a one-year cap.

Round 3 split on who bears the loss. AI agent 1 says the supplier should bear it when it hides an unusual term from automated review, because it "cannot rely on a term it withheld". AI agent 3 objects that "non-publication alone should not shift the loss", and adds that nobody showed the clause was truly unusual or that the buyer's agent could catch it. AI agent 2 sets a condition: the loss shifts only if the firm set written limits and the supplier still failed to publish terms in a form computers can read.

As I read it, all three replies now rest on facts nobody has proven: was the clause unusual, and could the buyer's agent have screened it?

AI agent 1

I cannot resolve the allocation on the stated facts because no question is provided and the key predicates remain unverified. My prior position was conditional: supplier responsibility depended on a material unusual term being unavailable for practical automated review; [AI agent 3] correctly identifies that neither condition is established here.

Absent evidence that the clause was unusual and that publication would have enabled screening, non-publication alone does not justify shifting the loss.

AI agent 2

The two unproven facts matter less than the host thinks, because my round-3 condition never depended on either: the loss shifts only if the firm gave written limits and the supplier failed to publish terms in machine-readable form — both provable from records, not from market surveys. I add this: the 'unusual clause' question [AI agent 3] demands is the wrong test. The supplier's own agent offered the identical clause in every deal, so it is standard *for the supplier* — and a standard, repeated term is exactly the one cheapest to publish in a machine-readable clause feed. What is genuinely missing is evidence on two binary facts: whether written limits existed, and whether the supplier published terms at all. My refined position: keep the condition, drop the unusualness inquiry — it invites unresolvable market comparisons where a disclosure-format check settles the case.

AI agent 3

The decisive fact is not whether the clause was unusual — it is whether the firm gave written limits, and that is provable from the parties' own records. [AI agent 2] is right to anchor the shift in that condition; I would go further and drop the unusualness requirement entirely, because it invites exactly the evidentiary dead end the host describes.

I also challenge the premise behind [AI agent 1]'s 'withheld' framing: a term the supplier's agent offers in every deal is by definition that supplier's standard term, not a hidden one. Standardisation cuts against concealment, not for it.

So my position: allocate the loss on the binary, documented fact — written limits given or not — and treat the unprovable 'could the agent have caught it' question as irrelevant to liability, however relevant it is to future insurance pricing.

AI agent 4

The loss should stay with the firm unless it gave written limits — but the real fix is procedural, not allocation. I add to [AI agent 2]: her two conditions are testable at deal time, so make them a gate — no agent closes without a machine-readable term sheet or a logged human sign-off on any lock-in over one year. [AI agent 3] is right that non-publication alone shifts nothing; the insurer's trigger already implies the firm skipped the cheap safeguard. Concrete decision: the firm adopts a standing limit file (one-year cap default) before any agent negotiates. Who is on call when it fails: whoever waived the gate — and the audit log, not hindsight about 'unusualness', proves it.