The question whether English contract law can accommodate agreements negotiated or performed by autonomous artificial intelligence (AI) agents has shifted, within a decade, from a niche concern in the law of electronic commerce to a problem of immediate doctrinal urgency. Generative and reinforcement-learning systems now make pricing, sourcing and contracting decisions with limited human supervision, sometimes producing outputs their principals neither foresaw nor would have authorised. The orthodox response, exemplified by Chwee Kin Keong v Digilandmall.com Pte Ltd [2004] 2 SLR 594 and by the Law Commission’s Smart Legal Contracts advice (2021), is that English contract law is sufficiently flexible to absorb these developments because contracts formed by machines are, in substance, contracts formed by the humans or companies who deploy them.
This essay argues that the orthodox view is correct in its general direction but underestimates two doctrinal pressure points. First, the formation rules of offer, acceptance, certainty and intention to create legal relations can largely cope with deterministic and even mildly stochastic systems through the “mere tool” fiction, but that fiction strains where the agent’s output is genuinely emergent and not within the rational contemplation of the deploying party. Secondly, the doctrines of mistake, misrepresentation and authority were designed around a cognitive subject, and they translate awkwardly onto systems that have no beliefs, no intent and no capacity to know. The proper response is not to confer legal personality on AI agents — a position rightly rejected by mainstream English scholarship — but to refine the tool/agent boundary, to develop a more honest doctrine of attribution, and to recalibrate objective interpretation to take account of the realistic expectations of human counterparties dealing with machine outputs.
The orthodox position: AI agents as sophisticated tools
The starting point in English law is that a contract concluded by an automated system is a contract between the human or corporate principals who deploy that system. This was the unanimous view of the UK Jurisdiction Taskforce’s Legal Statement on Cryptoassets and Smart Contracts (2019), endorsed by the Law Commission in its 2021 advice to government on smart legal contracts (Law Commission, 2021, paras 2.31–2.45). Both bodies concluded that no statutory intervention was needed because English contract law’s flexibility — and in particular the objective principle of interpretation — could accommodate automated and algorithmic dealings.
The doctrinal anchor is the objective theory of agreement articulated in Smith v Hughes (1871) LR 6 QB 597 and reaffirmed in RTS Flexible Systems Ltd v Molkerei Alois Müller GmbH [2010] UKSC 14: contractual obligations arise from how a reasonable person in the position of the offeree would interpret the offeror’s conduct. Where a company instructs a system to display prices, accept orders or sign electronic documents, the company is bound by what the system communicates, irrespective of whether any human within the company contemporaneously knew or approved of the communication. Software Solutions Partners Ltd, R (on the application of) v HM Customs and Excise [2007] EWHC 971 (Admin) and the Singapore Court of Appeal’s reasoning in Chwee Kin Keong v Digilandmall.com Pte Ltd [2005] 1 SLR(R) 502 — frequently cited as persuasive in English commercial practice — both treat automated systems as the agents or instruments of the operating company.
This “mere tool” account is doctrinally economical. It avoids the need to attribute mental states to machines and preserves the symmetry of contractual obligation by routing legal responsibility to a recognised legal subject. Section 7 of the Electronic Communications Act 2000 and the Electronic Commerce (EC Directive) Regulations 2002 reinforce the position by giving legal effect to electronic communications and electronic signatures without requiring contemporaneous human intervention.
For deterministic algorithms — a fixed pricing engine, an EDI ordering system, a rule-based smart contract — this analysis is convincing. The system’s output is, in principle, traceable to a set of human-authored rules, and binding the principal accords with both efficiency and party expectations. The harder question is whether the same account survives the move from deterministic systems to autonomous agents whose outputs are the product of learned, probabilistic behaviour that the principal cannot fully predict.
The fiction of the “mere tool” under strain
Autonomous AI agents differ from earlier automated systems in three legally salient respects. First, their outputs are not reducible to a determinate set of rules authored by the deploying party; they are emergent properties of training data, model weights and runtime inputs. Secondly, they may negotiate in natural language with counterparties or other AI systems, making bargaining moves (price concessions, term variations, side agreements) that the principal has neither expressly authorised nor expressly forbidden. Thirdly, they can, in some architectures, modify their own strategies over time. The “mere tool” framing, which works for a vending machine or a static webpage, becomes increasingly artificial as the gap widens between what the principal authorised and what the system did.
Scholars have responded to this tension along three principal axes. Chopra and White’s influential treatment treats sophisticated electronic agents as functionally analogous to human agents and urges the development of an agency-like doctrine of attribution (Chopra and White, 2011). Lim, by contrast, argues that English law can resolve most cases through a generous application of objective interpretation and the existing rules on apparent authority, without needing to confer any form of legal status on the agent itself (Lim, 2019). A third strand, associated with Bryson and others, warns against any move toward legal personhood for AI on both conceptual and political grounds (Bryson, Diamantis and Grant, 2017).
The better view is that the agency analogy is illuminating but ultimately misleading as a doctrinal route. Agency in English law presupposes a principal–agent relationship between two legal persons, with the agent owing fiduciary duties and possessing the cognitive capacity to form intentions on the principal’s behalf (Bowstead and Reynolds, 2021, paras 1-001 to 1-006). An AI system possesses none of these attributes. Treating it as an agent therefore requires either a statutory deeming provision (as exists in some United States jurisdictions under the Uniform Electronic Transactions Act 1999, §14) or a doctrinal fiction at least as ambitious as the personification of the company. Neither is currently available in English law, and neither is plainly desirable.
Yet the “mere tool” account cannot do all the work either. The tool metaphor suggests something inert, controlled and predictable. An autonomous trading agent or large-language-model-driven procurement bot is none of these things in any meaningful sense. The honest doctrinal position is that the system is treated as a tool because policy and convenience require that some legal subject be bound by its outputs, not because the metaphor accurately describes what the system is. Acknowledging this candidly clarifies where the doctrine will and will not stretch.
Formation: offer, acceptance and the objective principle
Formation doctrine is the area where English contract law copes best. The objective principle, properly applied, has no difficulty in concluding that where party A’s AI agent communicates an offer to party B (or B’s agent) and the offer is accepted, a contract is formed between A and B on the terms communicated. The reasoning in Storer v Manchester City Council [1974] 1 WLR 1403 and Centrovincial Estates plc v Merchant Investors Assurance Co Ltd [1983] Com LR 158 makes plain that internal mental reservations do not displace the outward appearance of agreement.
Three complications nonetheless arise. First, the rules on instantaneous communications, derived from Entores Ltd v Miles Far East Corporation [1955] 2 QB 327 and Brinkibon Ltd v Stahag Stahl [1983] 2 AC 34, were designed for telex and telephone exchanges between human operators. Where two AI agents negotiate in machine time, the question of when and where the contract is formed has practical consequences for jurisdiction and applicable law. The Brinkibon framework, with its emphasis on “the intentions of the parties, by sound business practice and in some cases by a judgement where the risks should lie” (per Lord Wilberforce at 42), is flexible enough to absorb machine-to-machine dealings, but its application will require courts to make policy choices about the locus of the contract that have not yet been tested.
Secondly, the requirement of certainty in Scammell and Nephew Ltd v Ouston [1941] AC 251 raises an unsettled question where AI agents conclude contracts in natural language. If an agent agrees in conversational text to “favourable terms to be agreed in due course”, is that an agreement to agree of the kind condemned in Walford v Miles [1992] 2 AC 128, or a binding obligation with implied content? The doctrine is unaltered by the AI context, but the volume and novelty of natural-language machine outputs will multiply borderline cases.
Thirdly, intention to create legal relations may need rearticulation. In B2B contexts the Esso Petroleum Ltd v Customs and Excise Commissioners [1976] 1 WLR 1 presumption of legal intention will normally apply. But where a consumer interacts with a chatbot that purports to make commitments — for example, the much-publicised airline cases in which generative chatbots offered refunds the airline disputed — the question whether the chatbot’s output reflects an intention to create legal relations is genuinely live. Moffatt v Air Canada (2024 BCCRT 149), although a Canadian small claims decision, illustrates the issue: the tribunal held the airline bound by misinformation given by its chatbot. An English court applying objective principles would, on the better view, reach the same conclusion, because a reasonable consumer would understand the chatbot as speaking for the company. The case is therefore best read as a vindication of orthodox doctrine, not a departure from it.
Mistake: the limits of doctrinal translation
The doctrine of mistake exposes the deepest tension between English contract law and autonomous AI. Unilateral mistake as to terms requires the non-mistaken party to have actual or constructive knowledge of the mistake (Hartog v Colin & Shields [1939] 3 All ER 566; Statoil ASA v Louis Dreyfus Energy Services LP [2008] EWHC 2257 (Comm)). Translating this requirement onto an AI counterparty is conceptually awkward: a system has no beliefs, and the language of “knowledge” applies to it only by metaphor.
Two responses are available. The first, urged by Lim and others, is to translate “knowledge” into a functional test: would a reasonable system, or a reasonable principal deploying such a system, have detected the mistake? This is essentially the route taken in Chwee Kin Keong, where the Singapore Court of Appeal held that mistaken price postings by an automated system could be set aside where the counterparty knew or ought to have known of the mistake. The court attributed knowledge to the human buyer, not to the seller’s system, but the same logic supports attributing constructive knowledge to a deploying principal whose system fails to detect obvious anomalies.
The second response is to expand equitable intervention. The common law unilateral mistake doctrine is narrow, and the demise of Solle v Butcher [1950] 1 KB 671 following Great Peace Shipping Ltd v Tsavliris Salvage (International) Ltd [2002] EWCA Civ 1407 has left limited scope for equitable rescue. There is a case for a more generous approach where one party’s contractual output is the product of a system malfunction or adversarial manipulation, but this would require either statutory intervention or a willingness to revisit Great Peace — neither of which seems imminent.
The harder problem, however, is not mistake by the AI’s principal but mistake induced by the AI. Where a generative model “hallucinates” a contractual term or a factual representation that no human at the deploying company would have made, the question is whether the principal can be heard to say that the output does not represent its true intention. The answer, on orthodox principles, is no: the objective principle binds the principal to the outward appearance of its communications, and the principal has assumed the risk of deploying a system that produces unauthorised outputs. The position is analogous to Hambrook v Stokes Bros [1925] 1 KB 141 in tort — a party who unleashes a foreseeable risk into the world bears its consequences — though the analogy is imperfect because contract liability rests on consent rather than fault.
This conclusion is doctrinally defensible but creates strong incentives for principals to over-constrain their systems, potentially reducing the commercial utility of AI agents. A more refined approach would distinguish between outputs within the agent’s apparent scope of authority (binding) and outputs so divergent from that scope that a reasonable counterparty should have queried them (not binding, or voidable). This mirrors the apparent-authority reasoning in Freeman & Lockyer v Buckhurst Park Properties (Mangal) Ltd [1964] 2 QB 480 and provides a workable doctrinal bridge between agency and the AI context.
Misrepresentation and the problem of attribution
Misrepresentation law is similarly stress-tested. A fraudulent misrepresentation under Derry v Peek (1889) 14 App Cas 337 requires a representation made knowingly, without belief in its truth, or recklessly. None of these mental states can be attributed to an AI system. Section 2(1) of the Misrepresentation Act 1967 imports a reverse burden of proof but still presupposes a “person” making the representation with a belief as to its truth.
Where a chatbot makes a false statement of fact that induces a counterparty to contract, the natural legal response is to treat the statement as a representation by the deploying principal. But on what basis? If the principal had no contemporaneous knowledge of the statement, fraud is unavailable. Negligent misrepresentation under s.2(1) might apply, but the statutory language (“the person making the misrepresentation would be liable to damages in respect thereof had the misrepresentation been made fraudulently”) fits poorly with a non-human speaker. The better analysis is that the principal is the maker of the representation through the system, and the relevant belief and reasonable grounds are assessed at the level of the principal’s deployment decisions: did the principal have reasonable grounds to believe that its system would produce accurate statements?
This recharacterises misrepresentation as a form of organisational fault, closer to negligent provision of information under Hedley Byrne & Co Ltd v Heller & Partners Ltd [1964] AC 465 than to traditional misrepresentation. The doctrinal cost is significant: the fault-based core of s.2(1) is preserved only by translating fault from the speaker to the deployer. Yet the alternative — declining to apply misrepresentation to AI-generated statements — would leave a serious gap in consumer and commercial protection. The Law Commission’s silence on misrepresentation in its 2021 advice is, in this respect, a missed opportunity.
Authority, ratification and the limits of binding the principal
If AI agents are not legal persons, they cannot have authority in the agency sense. Yet courts will sometimes need to ask whether an AI’s contractual output is within the scope of what the principal authorised, particularly where the principal asserts that the system “went rogue”. Apparent authority, as articulated in Freeman & Lockyer and refined in The Ocean Frost [1986] AC 717, requires a representation by the principal as to the agent’s authority, on which the third party reasonably relies.
Translating apparent authority to AI agents requires identifying the relevant representation. The strongest candidate is the principal’s act of deploying the agent in a particular commercial context: by placing a procurement bot in a marketplace, the principal represents that the bot has authority to do what such bots ordinarily do. This is a coherent and workable extension, but it sharpens the question of what AI agents “ordinarily” do — a question on which market practice is still forming. Where an agent’s behaviour is significantly outside the ordinary range, the third party’s reliance becomes harder to characterise as reasonable, and the principal may escape liability.
Ratification, under Bolton Partners v Lambert (1889) 41 Ch D 295, offers an additional safety valve: a principal who learns of an unauthorised contractual output may adopt it, retrospectively binding both parties. This doctrine is unproblematic in the AI context, but it places the burden of monitoring on the principal and may not assist counterparties who require certainty at the moment of formation.
Performance: smart contracts and automated execution
Beyond formation, AI and related technologies raise distinct issues at performance. Smart contracts, in the strict sense of code that automatically executes contractual obligations, were the focus of the Law Commission’s 2021 advice. The Commission concluded that the existing law of contractual interpretation — including the principles in Wood v Capita Insurance Services Ltd [2017] UKSC 24 and Arnold v Britton [2015] UKSC 36 — could be applied to natural-language smart contracts, hybrid smart contracts and even purely code-based smart contracts, with the latter raising the question whether the reasonable reader is a reasonable coder (Law Commission, 2021, paras 4.50–4.70).
The harder performance issue arises where an AI agent makes discretionary decisions during the life of the contract — for example, deciding whether to invoke a force majeure clause or to call a guarantee. The Braganza v BP Shipping Ltd [2015] UKSC 17 duty of rationality requires that contractual discretions be exercised honestly and not arbitrarily, capriciously or irrationally. Whether an AI’s decision can be “irrational” in the Braganza sense is a genuinely novel question. The better view is that Braganza applies to the principal who deploys the AI, and that deploying a system known to be unreliable or biased may itself constitute irrationality. This extension is doctrinally modest and reinforces, rather than undermines, the principal-attribution framework.
Should AI agents have legal personality?
It is sometimes argued that the difficulties canvassed above could be solved by conferring some form of legal personality on AI agents, in the manner of corporate personality. The European Parliament’s 2017 resolution on civil law rules on robotics famously floated the idea of “electronic persons” (European Parliament, 2017), although the proposal was widely criticised and has not been pursued in the EU AI Act.
The arguments against personhood are strong and have been comprehensively rehearsed by Bryson, Diamantis and Grant (2017) and by Solaiman (2017). Legal personality without assets is a shell: an “electronic person” cannot meaningfully be sued unless it has a fund or insurance behind it, in which case the fund or insurer is the real defendant. Personality also risks displacing responsibility from the human or corporate actors who design and deploy the system, weakening incentives for safe development. English law has resisted these moves and is right to do so. The doctrinal toolkit of attribution, apparent authority, organisational fault and objective interpretation is sufficient, properly developed, to allocate contractual risk between human and corporate parties without inventing a new legal subject.
Therefore: a qualified affirmative
English contract law can cope with agreements negotiated or performed by autonomous AI agents, but only on three conditions. First, the objective principle of interpretation must be applied robustly to treat the principal as bound by the outward appearance of its system’s communications, even where the principal did not contemporaneously approve them. This is already orthodox doctrine and requires no reform. Secondly, the doctrines of mistake and misrepresentation must be reconceived as forms of organisational fault, with knowledge and belief assessed at the level of the principal’s deployment decisions rather than at the level of the system’s “cognition”. This requires modest but real doctrinal development, most likely through judicial elaboration rather than legislation. Thirdly, the apparent-authority framework should be extended to AI agents on the basis that deployment in a commercial context constitutes a representation as to the scope of the system’s contracting power, with the scope determined by reference to evolving market practice.
What English law should resist is the temptation to confer legal personality on AI agents. The doctrinal economy of treating the agent as the conduit of the principal preserves the integrity of contract as a consensual institution while routing liability to a subject that can meaningfully bear it. The real risk to coherence lies not in technological novelty but in doctrinal evasion: in courts and commentators reaching for the “mere tool” metaphor without acknowledging that, for autonomous systems, the metaphor is doing policy work rather than descriptive work. A more candid doctrine — one that admits the fictional element in attribution and grounds it in the policy of allocating risk to the deploying party — will serve both certainty and fairness better than either an uncritical extension of the tool model or a destabilising move toward electronic personhood.
The ultimate answer to the question posed is therefore a qualified yes. English contract law can cope, because its foundational doctrines — objectivity, attribution, apparent authority, organisational fault — are sufficiently abstract to absorb the AI context. But coping is not the same as governing well, and the quality of the law’s response will depend on whether courts and the Law Commission are willing to articulate, rather than disguise, the policy choices that the AI context forces into the open.
References
- Bowstead, W. and Reynolds, F.M.B. (2021) Bowstead and Reynolds on Agency. 22nd edn. London: Sweet & Maxwell.
- Bryson, J.J., Diamantis, M.E. and Grant, T.D. (2017) ‘Of, for, and by the people: the legal lacuna of synthetic persons’, Artificial Intelligence and Law, 25(3), pp. 273–291.
- Chopra, S. and White, L.F. (2011) A Legal Theory for Autonomous Artificial Agents. Ann Arbor: University of Michigan Press.
- European Parliament (2017) Resolution of 16 February 2017 with recommendations to the Commission on Civil Law Rules on Robotics (2015/2103(INL)). Brussels: European Parliament.
- Law Commission (2021) Smart Legal Contracts: Advice to Government. Law Com No 401. London: HMSO.
- Lim, E. (2019) ‘Attribution and contracts concluded by automated systems’, in Wischmeyer, T. and Rademacher, T. (eds) Regulating Artificial Intelligence. Cham: Springer.
- Solaiman, S.M. (2017) ‘Legal personality of robots, corporations, idols and chimpanzees: a quest for legitimacy’, Artificial Intelligence and Law, 25(2), pp. 155–179.
- UK Jurisdiction Taskforce (2019) Legal Statement on Cryptoassets and Smart Contracts. London: LawtechUK.
