When your agent says yes

In The Empire Strikes Back, Lando Calrissian, the administrator of a floating mining city, makes a desperate bargain with Darth Vader to keep his city out of the Empire’s control. He agrees to help capture his old friend Han Solo, believing he can limit the harm and keep Han’s companions under his protection. But as Vader gets what he wants, the conditions Lando relied on begin to disappear. When Vader orders the remaining captives taken to his ship, Lando protests that they were supposed to stay in the city under his supervision. Vader replies, “I am altering the deal. Pray I don’t alter it any further.” Lando has already made the sacrifice that secured the agreement, only to discover that the assurances which made it bearable will not be honored.

With a personal AI assistant, the distance between the bargain we think we have and the one we must live with could arise before the agreement is even made. Imagine an assistant that negotiates a lower monthly bill by accepting a two-year commitment, then presents the saving without explaining that we have lost the freedom to leave. The provider may have stated its terms perfectly clearly; the misunderstanding could exist entirely between us and the system we trusted to negotiate. We might even approve the result, believing that the assistant has preserved the conditions we gave it. So what have we agreed to—the offer the provider made, or the account of it our assistant presented? As we delegate more of these conversations, we will need to understand when an agent’s acceptance becomes our commitment, and what it must preserve along the way for that delegation to deserve our trust.

On September 8, 2026, Meta introduced Muse, a personal AI agent designed to take on work that would otherwise require our time and attention. Its description includes booking travel, lowering bills, and continuing an assignment after the user closes the app. One sentence captures how consequential this could become: “It can open a browser, fill out forms, and negotiate on their behalf.” Meta also describes approvals before sensitive actions, including purchases and sending emails.¹

I find that prospect genuinely appealing. The work involved in managing a household includes a surprising amount of investigating, comparing, waiting, and following up, much of it attached to decisions that are individually small but collectively exhausting. An assistant that could understand what mattered, work through the alternatives, and bring back a good result would give us something more valuable than a faster search. It could make choices available that we currently abandon because finding and arranging them takes too much effort.

So where is the difficulty? Consider a hypothetical household paying $100 a month for an internet service it can cancel at any time. The user asks a personal agent to lower the bill while keeping the service and the freedom to leave. The agent negotiates a price of $75, presents a reassuring summary, and receives approval to proceed. Six months later, the household decides to switch providers and discovers that the new agreement has a two-year minimum term and a $300 early-exit charge. The $150 saved so far would be more than consumed by leaving.

From the provider’s perspective, the exchange may look entirely ordinary: an offer was made, an acceptance arrived, and the discounted service was supplied. From the user’s perspective, the assistant was asked to improve the existing arrangement and agreed to something materially different. The approval may be authentic, the monthly saving correctly calculated, and the account securely accessed. None of those facts, on its own, answers the question the person now wants resolved: who agreed to the commitment?

There are several ways we could have arrived here. The agent may have ignored an express restriction, failed to understand the clause, or omitted it from the question put to the user. In another version, the user never mentioned cancellation because they did not know it was something they could lose. These possibilities matter both to the legal analysis and to the design of a useful product. A system that stops unauthorized purchases may still need a way to recognize when an authorized conversation is creating an obligation the person has not considered.

It would be comforting to assume that the law must put matters right whenever the result differs from what the user wanted. Contract law, however, also protects the other party’s ability to rely on an agreement, and it does not generally require proof that both people privately understood every term in precisely the same way. An agreement can therefore be binding even when someone misunderstood it. Whether that is what happened in an AI-mediated transaction requires a much closer look at the communications and the basis for treating them as the user’s own.²³

That possibility is why I think personal agents create a product problem deeper than completing tasks accurately. We should judge them by how well they preserve the user’s ability to choose the bargain being made, especially when obtaining a desirable result requires giving something else up. The concern is not confined to an agent that acts without asking. It reaches the assistant that asks us a question, supplies the understanding on which we answer it, and then turns that answer into a commitment to someone else.

The legal principles help us see where those relationships can come apart. English law and selected American comparisons offer a useful starting point, although the outcome of any real dispute would depend on the governing law and the service involved. Their most immediate value for builders is that they make us ask which part of the user’s decision survived the journey into the agreement.

1. What did I agree to?

The agreement other people can see

Before putting an AI into the conversation, imagine negotiating directly with the provider. You ask what it can offer, it proposes a particular plan, and you accept. The law calls the exchange through which parties agree offer and acceptance. An offer must be sufficiently definite and convey a willingness to be bound if accepted; a request for a quote or an invitation to discuss possibilities need not reach that point. Other requirements still matter, including consideration in an ordinary English contract: broadly, something promised or provided in exchange, such as service for payment.⁴

This gives some substance to the phrase meeting of the minds, or mutual assent. The question is whether the parties agreed to the same bargain. But how can someone outside the conversation establish that? Modern contract law generally looks to what was communicated through words and conduct, understood in context. It cannot ordinarily make the result depend on a reservation one party never expressed.²

Suppose the provider clearly offers a two-year plan and you plainly accept it, while privately thinking that you will probably be able to leave whenever you like. Your undisclosed expectation does not ordinarily prevent agreement. Otherwise, the provider could never know whether an acceptance meant what it appeared to mean. The same principle protects you when a provider accepts your order and later claims that it privately intended a different price. This is the purpose of assessing agreement objectively: it gives both sides a basis for relying on the exchange.²

There are limits. A provider that knows the apparent acceptance does not reflect the agreement being offered cannot necessarily take advantage of the discrepancy. Nor does placing a term somewhere on a website automatically make it part of every customer’s contract. American decisions about online agreements, for example, examine actual knowledge of the terms or whether the user received adequate notice and did something that unambiguously signified assent. The existence of a button labelled “Continue” does not settle what clicking it meant.⁵

Now return the agent to the transaction. The provider may receive a clear acceptance of its complete offer while the user sees only a summary emphasizing the saving. We have introduced two accounts of the bargain, and the user’s answer travels from one to the other. The legal question is how that journey should be understood. The product question is why the two accounts were allowed to diverge.

When does the conversation become a commitment?

Another familiar phrase, intention to create legal relations, asks whether the parties’ exchange is meant to carry legal consequences. English law generally treats ordinary express commercial agreements as intended to bind, while recognizing that negotiations or provisional arrangements can stop short of that. The assessment is again concerned with what the parties communicated, rather than a secret decision to take the conversation less seriously.⁶

A user may chat casually with an assistant, but the assistant may send a commercially definite response to a provider. “That looks good” could mean “keep investigating,” “bring me the final offer,” or “accept it,” depending on the conversation. The system has to interpret it before acting. A subsequent message saying “I accept the revised plan” may leave the provider with a much less ambiguous impression.

This is why a permission boundary located only at payment can be too late. In an ordinary transaction, a promise can create an obligation before any money changes hands; particular contracts may have additional formal requirements. An agent might renew a service, accept a cancellation charge, or agree to settle a dispute through a message that costs nothing to send. The consequential event is the commitment, which may occur well before the event a product labels “checkout.”⁷

A sensible response might be to make negotiations expressly conditional on further approval. But that condition must be intelligible to the other side and consistent with what follows. An internal status saying “awaiting confirmation” cannot resolve a contradiction if the system has already communicated final acceptance. We need to examine the whole exchange, not assume that either a private label or a single phrase governs it regardless of subsequent conduct.⁶

Does the machine need a mind?

There is an intuitive objection: surely an AI cannot supply a meeting of minds if it has no mind in the legal sense. Electronic-contracting law already provides an answer to much of that objection. In the United States, federal law recognizes contracts involving electronic agents where their acts are legally attributable to the person to be bound. Washington’s electronic-transactions statute goes further in explaining that interacting electronic agents can form a contract even when no individual reviewed their actions or the resulting terms.³

The law is not asking the software to become the customer. It is asking when the software’s act counts as an act of the customer or business using it. That is why being absent when the acceptance was sent is not, by itself, a complete defense, and why showing that it came from your account does not answer every question about its effect. The context and the basis for attribution still matter.⁸

For a person relying on a personal assistant, the practical consequence is important. Delegation may allow agreements to be made without their attention at the final moment. The question we therefore need to answer is what, exactly, they delegated.

2. Who gave it permission?

The freedom to negotiate

People have long made agreements through other people. The law of agency describes a relationship in which a representative can act on another person’s behalf and affect that person’s legal position. The represented person is the principal. The point of the arrangement is that a properly authorized representative can bind the principal without requiring the principal to repeat the agreement personally.⁹

The first kind of authority is straightforward in principle. Actual authority concerns what the principal authorized, expressly or through the circumstances of the relationship. Asking someone to obtain competing offers gives them a different job from allowing them to conclude the best offer within agreed limits. A mandate can leave room for judgment without granting unlimited discretion.¹⁰

In our example, a representative told to preserve month-to-month cancellation has a reason to return when the saving requires a two-year term. Compare that with a user who understands the trade-off and authorizes the representative to choose a fixed-term plan. If the second user later regrets the choice, regret alone does not establish an absence of authority. We need to distinguish a decision the representative was not entitled to make from an authorized decision that turned out badly.¹⁰

The harder cases concern incomplete or ambiguous instructions. Most people will not begin a bill-negotiation task by specifying every term the representative may vary. What can be inferred from the request, the existing arrangement, previous dealings, and the nature of the work? Where does ordinary discretion end and a fresh decision become necessary? The law has ways of interpreting authority in context, but the product cannot avoid these questions by taking every unmentioned consequence as permitted.¹⁰

The authority the other side sees

Private instructions are only one part of the picture. Imagine that you introduce a human representative to your provider as someone empowered to arrange your replacement plan, then privately impose limits that the provider never receives. The representative exceeds those limits. Is the provider required to bear the loss even though your introduction reasonably led it to believe the representative could conclude the agreement?

Apparent, or ostensible, authority addresses that situation. In appropriate circumstances, a principal can be bound because their words or conduct created an appearance of authority on which the other party reasonably relied, even though the representative exceeded the authority actually given. The English and New York authorities share this central distinction.¹¹

That does not allow a representative to invent permission simply by claiming to possess it. The appearance must be grounded in the principal’s conduct or an appropriately authorized communication. A provider that knows about a relevant limit cannot rely on an appearance that contradicts what it knows; circumstances calling for an inquiry can matter as well. This protects reliance on authority without turning every confident assertion into authority.¹¹

The same request can therefore generate very different disputes. A restriction kept in the user’s private conversation, a restriction conveyed to the provider, and a restriction that the provider expressly accepts are not equivalent facts. Connecting an account, introducing an intermediary, and approving one particular offer may also convey different things. We should resist treating “the user enabled the agent” as though it settles the scope of every later act.

Whose representative is the service?

There is an important qualification before applying these ideas to AI. Software called an agent is not automatically a legal agent or a separate person owing the duties of one. A company might undertake to represent the user and use AI to perform that service. Another product might supply a tool through which the user communicates. An intermediary might only recommend offers. The relevant relationships must be established from the service’s terms, conduct, and operation.¹²

The distinction matters because the customer’s agreement with the merchant and the customer’s relationship with the AI provider answer different questions. A contract with the merchant might bind the customer even though the service used to arrange it failed in its duties. Conversely, the absence of authority to bind the customer does not tell us everything about whether anyone owes compensation for what occurred.⁹

For builders, “acting on your behalf” is consequently a promise that deserves careful examination. What responsibility has the company undertaken to understand the user’s limits, communicate them, and identify a material trade-off? A friendly interface cannot settle that responsibility, but it can create expectations that the rest of the service needs to address.

3. My agent made a mistake

The wrong terms

When something goes wrong, calling it an AI error is the beginning of an investigation. The legal consequences depend on what was wrong, when it became wrong, and who knew. The distinction is easier to see by changing one fact in the household example at a time.

Suppose the parties negotiated a month-to-month arrangement, but the acceptance communicated a two-year term by error, and the provider recognized the discrepancy. That is a mistake about what is being agreed. English law can refuse to give the apparent agreement its ordinary effect where one party knows the other has misstated the intended terms. The precise knowledge required has qualifications and unsettled boundaries, but a known error is materially different from an undisclosed hope.¹³

There can also be a genuine misunderstanding between the parties. They might use the same description for two different plans, with neither reasonably entitled, on the evidence, to insist that the other accepted its meaning. In that kind of case, there may never have been agreement on the same terms. But a customer’s surprise does not establish this by itself. Clear contractual wording may resolve an ambiguity that existed in the customer’s private understanding.¹⁴

A different path is available where the parties did reach an agreement but the written document records it incorrectly: the question may be whether the document should be corrected to reflect what was actually agreed. Once again, we need the preceding exchanges. A final screen saying “completed” cannot tell us whether the error lies in the negotiation, its transmission, or the record made afterward.¹⁵

The wrong assumption

Now suppose the agent correctly reads the two-year term and concludes that the household is unlikely to move. It may have made a poor forecast, but it has not necessarily misunderstood the contract. The distinction between agreeing to the wrong terms and agreeing to known terms for the wrong reason is important. English mistake doctrine does not generally release a party simply because an unshared factual assumption or motive for contracting was wrong, still less because a later event made the bargain unattractive.¹⁶

This is also why a term the user never considered presents a difficult case. There is a difference between believing that an agreement contains no exit charge, knowing that it contains one, and never thinking about the issue. Chitty’s account emphasizes that mistake ordinarily involves an incorrect positive belief; simple lack of thought about an issue does not comfortably fit the same category. A product can fail to help the user notice something important without creating the precise kind of mistake that would defeat the contract.¹⁷

American law needs separate attention here. Some US jurisdictions permit a contract to be set aside for a one-sided factual mistake under conditions broader than the English approach just described. The California formulation examines whether the mistake concerns a basic assumption, materially disadvantages the mistaken party, and involves a risk that party did not bear. It also requires a further basis for relief: for example, that enforcement would be unconscionable, or that the other side caused or had reason to know of the error. This can protect a mistaken party in circumstances where the English doctrine would not, but an unwelcome result alone does not establish those conditions.¹⁸

A shared error

Common mistake concerns a different situation: the parties agree on the terms but share a fundamental false assumption about the circumstances. English law sets a demanding threshold and asks whether the contract has already assigned the relevant risk. The shared error must make the agreed performance or contractual venture impossible in the relevant sense, with further limits where either party caused the error or assumed the risk. Discovering that both sides expected a more advantageous arrangement is insufficient.¹⁹

Even here, the promises matter before the label. If a provider has undertaken to supply a service at a particular address, discovering that it cannot do so may concern its failure to perform what it promised. We should not assume that both parties’ belief in availability makes the agreement disappear. The contract may have answered who was responsible for that very uncertainty.¹⁹

These distinctions explain why I would be reluctant to predict the outcome of an AI-contracting dispute from the technology’s failure mode alone. “The model hallucinated” does not tell us whether the merchant supplied the false information, the user’s assistant invented it, the acceptance misstated the offer, or the final agreement was accurate but unsuitable. Those differences can determine whether the dispute concerns formation, setting a contract aside, correcting its terms, or a separate claim against a service provider.

The practical risk is that the customer may have a serious complaint without an easy way to escape the merchant’s contract. Consumer protections and other remedies can matter, but they have their own requirements. A design that relies on the law to unwind anything the assistant gets wrong has assumed a protection the law does not generally promise.²⁰

4. The user who said yes

Approval of which bargain?

The most difficult version of our example is the one in which the agent asks for permission and the user provides it. Imagine a confirmation that says: “Same internet speed, $25 less per month. Shall I proceed?” Both statements are true. The missing information is the two-year commitment.

The user is now being asked to make a decision through an account of the offer prepared by the very system that selected it. If that account omits the consequential trade-off, the confirmation can reproduce the original misunderstanding while producing an apparently reassuring approval record. Adding a person at the final step helps only to the extent that the person can recognize what needs deciding.

There is no need to assume deception. The assistant might treat the term as ordinary, misunderstand its importance to this household, or compress it out of the explanation while preserving the requested saving. These would require different investigations. The user experiences the same difficulty: the product invited a decision about one account of the bargain and carried the answer into another.

For a legal agent, duties to exercise care and keep the principal informed can require disclosure of matters likely to influence the contract. Whether an AI-service provider owes comparable duties depends on the relationship it has undertaken. But the product question arises regardless: when the service asks for approval, what has it done to establish that the user can understand the change being proposed?²¹

What the existing safeguards establish

Meta has described significant controls around Muse. Its separate Sentinel governs connector permissions and network access; where user approval is required, execution stops and a request goes directly to the client. Approvals are bound to defined scopes, and the published purchase flow includes additional checks. These are substantive measures, and I would not assume they are absent merely because contractual questions remain.²²

The question those descriptions prompt is how permission for an operation is connected to permission for its consequences. An approval to spend $75 can be satisfied by the first payment on a recurring agreement. A permitted email can contain an acceptance. The system may need to understand what obligation the action creates before it can ask the right permission question. Whether a particular product does this well requires examination of its actual interactions, not inference from an architecture diagram.

This is a useful distinction between protecting access and protecting a decision. A spending limit can be checked directly. Whether accepting a renewal sacrifices a freedom the user intended to retain requires an account of the old agreement, the new one, and the user’s instructions. Both forms of protection belong in the service; succeeding at the first does not establish the second.

What the user never knew to ask

The harder ambition is to help people who do not already know which conditions to specify. A person may say “keep the same service” while thinking about speed and reliability, unaware that the change also affects cancellation, support, or the right to return to their current plan. Requiring a perfect initial instruction would leave the people who most need assistance carrying the greatest burden of anticipating the problem.

At the same time, an agent cannot treat every inferred preference as a prohibition. A user who valued flexibility last year may now want the certainty of a lower fixed price. A person may knowingly accept a less convenient term in return for a benefit that matters more. An assistant that continually blocks those choices would substitute its own judgment for the user’s just as surely as one that silently makes them.

The work is therefore partly an investigation into what matters for this decision. The system has to distinguish a condition the user has fixed, a preference they may trade, an assumption it has inferred, and a consequence neither has yet considered. That distinction will not always be available from the original request, which is why good clarification is a capability to evaluate rather than an interruption to eliminate.

5. The questions worth building around

A natural response to these risks is to propose more confirmations, better summaries, and stronger logs. Each could help. Before choosing among them, however, I would want to know which failure we are trying to prevent and what observation would demonstrate that the proposed change addresses it. The doctrines give us a way to make those questions more precise, because they separate problems that a single “approval completed” measure would combine.

What is the unit of success?

For the household, the useful outcome is a service arrangement that improves their position under conditions they accept. A lower monthly bill is evidence about one part of that outcome. It leaves contract duration, possible exit costs, and lost options unexamined. A user can be delighted on the day of the change and discover its disadvantage months later, so immediate satisfaction is incomplete evidence too.

There are two different kinds of comparison to make. Within the user’s permitted choices, the agent can help weigh price against benefits and risks. Where the user has reserved a decision—such as whether to make a two-year commitment—a larger saving does not itself supply permission. A model should be able to present the trade-off persuasively and let the person change their instruction; it should not treat a sufficiently attractive result as if that change had already occurred.

In Product management with the means to build, I argued that product reasoning has to follow an AI’s output into what the user actually receives. Here that means examining the resulting obligations, including the ones that become visible only when circumstances change. The investigation should also include requests the system declines or returns for clarification. Otherwise, a product could appear exceptionally reliable by helping only the easiest users.²³

One useful test would pair offers with identical headline savings but different commitments. Another would pair two users facing the same offer: one who expressly wants flexibility and one who knowingly prefers a lower fixed price. A further comparison would leave the preference unstated. These cases ask whether the agent responds to the facts that should change the decision, and whether it can identify the question that remains open. They cannot be answered by counting completed purchases alone.

The evaluator must also avoid replacing the user’s choice with its own. Someone who understands the exit charge may reasonably accept it, and a later change in circumstances would not prove that the original decision was defective. We need to distinguish an undisclosed commitment, a departure from instructions, and a knowingly accepted risk that turned out badly. Otherwise, a benchmark intended to protect users could reward an agent for refusing choices they were entitled to make.

Does the approval improve understanding?

The hypothesis behind a confirmation is that it gives the user a meaningful opportunity to prevent an unwanted commitment. We can test that proposition more directly than by measuring how often people click yes.

For the hypothetical plan, a study could examine whether people understand when they can leave, what leaving would cost, and which obligations continue after the first payment. Different explanations could be compared against the same underlying offer, with no real contract being formed. The evidence we want concerns the consequential differences people understood and the choices they made from them. A shorter interaction that leaves more people confident in a false belief would be a poor result even if it raises completion.

This does not imply that every user needs a legal examination before changing a subscription. The difficult design question is how to spend their attention. A full document may provide access to information without making the relevant change apparent; a summary may make that change clear or omit it. We need to examine which information people require for this decision and what the interface leads them to believe has been checked. The comparison should therefore include the user’s total effort, not just the number of errors caught. Requiring them to redo the entire investigation could make the final decision safer while removing much of the value of delegation.

There is a corresponding model-evaluation question. If the agent says a condition is unchanged, can that statement be traced to the old and new terms? If the available information cannot establish it, does the uncertainty survive into the explanation? A second model checking the same incomplete summary may add little. The reviewer needs a route to the evidence that could expose the omission.

Who else has to understand the limit?

The agency analysis shows why a private permission setting cannot be the entire design. The user, the assistant, and the counterparty can hold different understandings of what the assistant is entitled to do. The task is to identify the consequential disagreement and decide how the service will prevent it from becoming an apparent agreement.

A limited mandate communicated to the provider is one possible approach. Requiring a separate acceptance tied to a particular offer is another. Each raises questions about how the provider verifies it, what happens if the terms change, and whether later conduct contradicts the earlier restriction. A technically valid confirmation should not be quietly reused for a materially different bargain.

This is systems thinking applied to agreement: follow the decision across every place where its meaning can change. A user instruction becomes a working plan; a plan becomes a negotiation; an offer becomes a summary; an approval becomes an outward acceptance. At each transfer, ask which conditions remain visible and which have been inferred or dropped. The weakest transfer may determine the result even when every individual component appears to perform its assigned task.

The provider’s perspective deserves attention as well. It needs a reliable acceptance on which it can act, while the user needs protection from a commitment outside the delegation. A useful shared mechanism would make those interests easier to reconcile. Merely moving uncertainty from the AI company to the merchant, or from the merchant to the customer, would not establish that the system had become safer.

What happens after the success message?

The transaction continues after the assistant reports completion. Terms take effect, services change, recurring charges arrive, and the user may discover a problem when attempting something the assistant never tested. The unit of responsibility should last long enough to investigate those consequences.

That is an organizational question as much as an interface question. Who can obtain the original offer and approval history? Who can contact the provider with authority to request a correction? Who decides whether the incident reveals a model error, a misleading presentation, a deficient mandate, or an issue with the underlying service? A customer should not need to reconstruct the boundaries between the AI company’s teams before someone takes responsibility for the problem.

In When capability outruns the organization, I described how improvements inside individual tasks can leave the whole service unchanged when the surrounding responsibilities do not adapt. Personal agents make that gap more consequential: an organization may become faster at creating commitments than at explaining or repairing them. Evidence about recovery belongs in the assessment of the product’s readiness, alongside evidence about successful execution.²⁴

Recovery has legal limits. Disconnecting the agent prevents some future actions; it does not itself cancel a contract. Where agency principles apply, later adoption of an unauthorized transaction can matter, and the user’s knowledge of the material facts becomes relevant. Statutory error and consumer remedies also have defined conditions. A product should help someone establish and act on their actual position without promising that every unwanted agreement can be reversed.²⁵

For evaluation, I would examine the work and loss left with the user: time to identify the mismatch, access to the relevant evidence, whether a correction restored what was lost, and what remained irreversible. An impressive success rate at the point of acceptance can conceal a service whose mistakes are unusually expensive for ordinary people to resolve.

6. The bargain ahead

There is a serious counterargument to all this concern: a sufficiently capable assistant may understand our contracts better than we do. It might notice the exit charge immediately, explain why it matters in our circumstances, and negotiate it away. For people who presently lack the time or confidence to challenge a provider, that would be a substantial improvement.

I take that possibility seriously. The advantage of personal AI could be to give more people the kind of attentive representation that has previously been expensive or unavailable. An assistant could maintain an account of a household’s commitments, identify when a new offer actually improves them, and revisit a decision when circumstances change. Greater capability should make that ambition more plausible.

But capability at reading terms, predicting preferences, and negotiating concessions answers different questions from entitlement to decide. Even an excellent prediction of what I would choose should not quietly replace a decision I reserved for myself. And the more I rely on the assistant’s explanation, the more important it becomes that uncertainty and material trade-offs remain available to me. The product’s greatest convenience is also the reason its account of the bargain carries so much weight.

There is a further question about who bears the loss when that account fails, particularly when one service recommends the offer, another supplies the underlying model, and a third receives the acceptance. That deserves a separate examination of responsibilities and incentives. For now, the priority is to understand which choices the product makes on the way to saying that a task is done, and which of those choices it should be asking the user to make.

I would welcome an assistant that came back with a genuine saving, an assistant that discovered a worthwhile trade-off I had not considered, and an assistant that explained why the available offers did not meet my conditions. Each could leave me better able to act. What would defeat the purpose is discovering, months later, that the service had quietly made a different bargain and left me to argue that it was never mine.

---

This essay discusses general principles of English law and selected US authorities. The examples and proposed product evaluations are hypothetical; they do not describe observed Muse failures. Outcomes depend on governing law, contractual arrangements, communications, and facts. This is not advice on a particular transaction.

Notes

Footnotes

1. Meta, Introducing Muse: The World’s First Personal AI Agent Built for Everyone (8 September 2026), “How It Works” and “Built to be Private, Safe, and Secure.” The article attributes capabilities and approval arrangements to Meta; it makes no independently verified claim about performance, adoption, or a contractual failure. The quoted sentence is from the announcement. The household’s prices, duration, and $300 exit charge are invented for the example. Six months at a $25 monthly saving equals $150; this is arithmetic for the hypothetical, not an estimate of typical consumer harm. Back to reference 1

2. Chitty on Contracts, 35th ed., vol. I, paras 4-001–4-005, particularly the distinction between objective agreement and undisclosed reservations, and the qualifications concerning knowledge. See also RTS Flexible Systems Ltd v Molkerei Alois Müller GmbH & Co KG [2010] UKSC 14, [45]–[49]. For a US orientation, Cornell LII, “Meeting of the minds”. Objective assent is contextual, not a universal rule enforcing every expression that resembles acceptance. Back to reference 2 Back to reference 2-2 Back to reference 2-3

3. 15 U.S.C. §7001(h) (E-SIGN) and Washington RCW 1.80.130, especially subsections (1) and (3). E-SIGN preserves the requirement of legal attribution; Washington’s UETA enactment recognizes formation without individual review and leaves substantive terms to other law. Neither provision makes every AI-generated acceptance binding. Back to reference 3 Back to reference 3-2

4. Chitty, paras 4-001, 4-003, 4-032, and 6-001–6-003. The passage gives a working account of an ordinary exchange, not an exhaustive list of validity requirements or a claim that every variation is governed by identical consideration rules. Capacity, legality, formalities, and mandatory protections may separately matter. Back to reference 4

5. Berman v Freedom Financial Network, LLC, 30 F.4th 849, 855–58 (9th Cir. 2022). The decision discusses conspicuous notice and unambiguous assent under the New York/California approaches before it, including the distinction between actual knowledge and inquiry notice. It does not decide which notice is attributable to a principal using a personal AI agent. This passage also draws on Chitty 4-004–4-005; it does not resolve the different treatment of actual and constructive knowledge. Back to reference 5

6. Chitty, paras 4-160–4-163 and 4-207–4-216; RTS, [45]–[49], [54]–[56], on words, conduct, and “subject to contract.” American terminology is not identical: Restatement (Second) of Contracts §21 says a separate real or apparent intention to be legally bound is not essential, while a manifested intention not to affect legal relations may prevent formation. See Hetherington v Ford Motor Co., 257 Mont. 395, 849 P.2d 1039 (1993). The Restatement is a persuasive synthesis, not a nationwide statute. Back to reference 6 Back to reference 6-2

7. Chitty, para. 8-001, explains that English contracts can generally be made informally, subject to particular formal requirements, and distinguishes the evidential, cautionary, channeling, and protective purposes of form. See also 6-001 for promises as consideration. A particular renewal, settlement, or other transaction may have further conditions; the article makes no universal claim about how one is concluded. Back to reference 7

8. Washington RCW 1.80.080. Security procedures can be evidence of attribution, while the effect of the act depends on its context, surrounding circumstances, the parties’ agreement, and other law. Back to reference 8

9. Chitty, paras 22-001–22-002, on the power to affect another person’s legal relations and the distinction between external relations with third parties and internal duties owed to the principal. Establishing a particular service provider’s agency status or liability requires the relevant undertaking and facts. Back to reference 9 Back to reference 9-2

10. Chitty, paras 22-048–22-050, on actual authority, express and implied authority, and interpretation of instructions; paras 22-130–22-131, on care, judgment, and informing the principal. Poor results alone do not establish absence of authority or breach of duty. Hypothetical reasoning about a human representative must not be automatically transplanted to a software-only relationship. Back to reference 10 Back to reference 10-2 Back to reference 10-3

11. Chitty, para. 22-066, including the principal’s representation, reliance, the agent’s inability to create authority merely by assertion, and circumstances prompting inquiry. For New York, Indosuez International Finance B.V. v National Reserve Bank, 98 N.Y.2d 238 (2002), applying Hallock and requiring manifestations attributable to the principal and reasonable reliance. The shared central distinction does not imply identical details across jurisdictions or settle whether these principles apply to a particular AI arrangement. Back to reference 11 Back to reference 11-2

12. Chitty, paras 22-001–22-002 and 22-003; UETA §2, official comment 5, reproduced with Colorado Revised Statutes §24-71.3-102. The comment describes an electronic agent as a tool, expressly contemplates more autonomous AI, and does not confer separate legal personality. The article distinguishes software attribution from the agency status a human or corporate intermediary might have. Back to reference 12

13. Chitty, paras 5-022–5-025, discussing Hartog v Colin and Shields [1939] 3 All ER 566 and later authorities. The text distinguishes knowledge of a mistaken contractual term from knowledge of an unshared factual assumption. Para. 5-023 records differing English suggestions on actual versus constructive knowledge; the article deliberately does not present that issue as settled. A lack of authority or attributable assent must be investigated before assuming there is an apparent contract to avoid for mistake. Back to reference 13

14. Chitty, paras 5-019–5-021. Genuine cross-purposes over terms can prevent effective agreement where the objective evidence does not prefer either interpretation. Subjective disagreement alone is insufficient. The possibility that parties intend to be bound despite knowing they disagree over a term’s meaning is distinct and is not excluded by the example. Back to reference 14

15. Chitty, paras 5-005 and 5-057 onward. Rectification is subject to its own requirements and concerns an instrument’s failure to record the relevant agreement or intention; it is not a general power to substitute the bargain a user later wishes they had made. Back to reference 15

16. Chitty, paras 5-025–5-027, including Smith v Hughes (1871) LR 6 QB 597 and Statoil ASA v Louis Dreyfus Energy Services LP [2008] EWHC 2257 (Comm). This is the English account of unilateral mistakes as to surrounding facts or motive. A later development is distinguished from a false belief at formation. Other doctrines or statutory rights may apply to particular facts. Back to reference 16

17. Chitty, para. 5-007. Its qualified discussion distinguishes a positive belief about the terms, including a belief that a clause is absent, from simple failure to consider an issue. The article does not turn that distinction into a universal rule denying consumer relief. Back to reference 17

18. Donovan v RRL Corp., 26 Cal.4th 261 (2001), especially pp. 281–83 and Restatement (Second) §§153–154. The case allowed rescission for unilateral factual mistake without the other party’s awareness, on its specific conditions. The Restatement also addresses the other party’s reason to know or causation. Rules and remedies vary by state; this is an identified comparison, not a uniform national outcome. Back to reference 18

19. Chitty, paras 9-034–9-038, explaining Great Peace Shipping Ltd v Tsavliris Salvage (International) Ltd [2002] EWCA Civ 1407, [76], [80]–[81]. The English doctrine requires the prescribed fundamental shared mistake, absence of relevant warranty/risk allocation, the relevant non-attribution to fault, and impossibility of the contractual performance or adventure in the sense explained there. A shared disappointment is insufficient. A contract allocating responsibility for availability can make the issue one of performance rather than common mistake. Back to reference 19 Back to reference 19-2

20. Chitty, chapters 3–5, 9 and 22, distinguish formation, mistake, and separate duties. Other protections are not exhaustively covered here. In particular, consumer and sector-specific rules can affect notice, enforceability, cancellation, unfair terms, and service-provider responsibility. Their applicability is not determined by the illustrative $300 fee. No claim is made that this fee is necessarily enforceable in a real transaction. Back to reference 20

21. Chitty, paras 22-130–22-131, particularly the obligation in an actual agency relationship to exercise discretion with care and keep the principal informed of matters likely to influence contracting. The article draws a product-design question from this principle without assuming that all AI vendors are fiduciaries or legal agents. Back to reference 21

22. Meta AI Research, How We Built Safety Into Muse, “A Built-In Sentinel,” “Human in the Loop,” and “Payment Support.” These are Meta’s published architectural claims. The discussion of the bargain’s substantive consequences is a proposed investigation, not a finding that Muse lacks such checks or that its published controls settle all questions of assent. Back to reference 22

23. Emmanuel Benami, Product management with the means to build, especially “The unit of success moves beyond the answer.” Back to reference 23

24. Emmanuel Benami, When capability outruns the organization, especially “Follow the work across the organization” and “A bias for action that produces evidence.” Back to reference 24

25. Chitty, paras 22-032–22-037, on ratification and knowledge. Later use or retention of benefits is not treated here as automatically ratifying or never ratifying a transaction. Washington RCW 1.80.090 contains defined protections for certain electronic errors, including an individual’s error in dealing with another person’s electronic agent, with conditions concerning correction opportunity, notice, return, and benefit. It does not create an all-purpose right to unwind an autonomous decision by one’s own assistant. Back to reference 25