From Contract Embellisher to Negotiation Machine: The Use of AI in Contracting

In this post, I argue that AI tools can be used as a ‘Contract Embellisher’: the parties to a contract deploy them to improve an agreement they have already concluded. This is less risky and less complex than using AI to negotiate a contract in the first place. At the same time, experience with AI as a ‘Contract Embellisher’ can feed back into using AI as a ‘Negotiation Machine’.

More than four decades ago, Howard Raiffa proposed the idea of ‘Post-Settlement Settlements’. Let’s assume A and B have concluded a contract. Is it value-maximizing, ie, efficient? Likely not, Raiffa hypothesized, especially in ‘complex negotiations where a lot of issues are at stake, where uncertainties are involved, or where settlements could involve transactions and payments over time’.

In comes an ‘analytical intervenor’ from the ‘Contract Embellishment Service’. Think of this intervenor as a mediator or honest broker. The intervenor meets with A and B privately, probing them for confidential information about their interests and values. A and B may have been reluctant to disclose these fully in their bargaining. Then the intervenor solves a ‘mathematical optimization problem’, suggesting to the parties an improved contract which A and B are free to accept or reject. The intervenor’s services would be compensated with a small slice of the surplus created.

For example, consider a negotiated supply contract. The parties have agreed on price, quantities, and delivery dates. The intervenor discovers that the buyer places considerable value on an earlier delivery of one batch, while the supplier can accommodate this at little cost if another batch is delivered later. Both can gain from a revised schedule.

The appeal of Raiffa’s proposal is that parties start from a ‘safe space’. They already have a contract, having settled the distributional problem of finding an agreement within the bargaining range. This contract is their fallback. They appear to have only to gain from the intervenor’s proposal. The contractual fallback may also reduce the parties’ incentives to misrepresent their interests and values. The acceptance process could be structured to preserve confidentiality and encourage acceptance (Raiffa did not discuss this): only if both A and B accept the proposal would the amendment be announced and their decisions disclosed. Otherwise, no amendment would be made, and a rejecting party would not know whether the other party had accepted or rejected the proposal.

Raiffa’s idea has a broad scope. All contracts could, in principle, be improved by an ‘analytical intervenor’. But the intervention could also be used to improve other settlements. Think of an enforceable court judgment A has secured against B. Enforcement may be difficult, costly, and time-consuming. A and B may be able to cut an ‘embellishment’ deal under which B makes an immediate voluntary payment at a small discount.

Despite the intellectual appeal of Raiffa’s proposal, it never gained much (if any) traction in commercial practice. I have not come across a single mediated post-settlement settlement in the last forty years. Parties who successfully negotiate a deal or settlement are usually exhausted and somewhat proud. They may be reluctant, emotionally and cognitively, to hear that they could have done better. More importantly, using a human broker as a ‘Contract Embellisher’ is costly. No human would work for a small percentage of the additional value created in a USD 1,000 transaction, let alone a USD 100 one.

But now we have AI. Companies such as Luminance and Icertis already offer AI tools for analyzing and improving contracts. Why not develop a neutral AI ‘Contract Embellisher’ that identifies Pareto improvements for both parties? Drawing on vast amounts of data about contractual arrangements and commercial practice, such a tool could identify value-creating opportunities that a human mediator might overlook.

The tool could also identify and suggest correcting departures from legal rules, precedents, or commercial practice. Not every such change would constitute a Pareto improvement: complying with a legal standard may impose costs on one or both parties. They should therefore be told whether the recommendation concerns joint value creation, legal compliance, or another normative objective.

Raiffa envisaged compensating the intervenor with a share of the additional value created. But that value may be difficult to quantify. A simpler pricing model would be a fixed fee per contract, possibly scaled to the contract’s value. Such a fee could be much lower than the cost of engaging a human mediator. AI might also overcome the psychological hurdle: parties may be more willing to accept suggestions for improvement from a machine than to be told by another human that they could have negotiated a better deal.

Another aspect is competition. The market for human ‘Contract Embellishers’ (mediators) is largely local and often opaque. In contrast, AI providers offering contract improvement services compete globally and fiercely, given the potentially substantial economic gains at stake. This should drive down prices and provide strong incentives to develop and market superior, value-creating products. Providers able to demonstrate high-quality recommendations, trustworthy handling of confidential information, and independence from either party may have a competitive advantage.

While Raiffa was concerned mainly about creating additional value in one-off contracts, his idea can be extended to long-term contracts. Let’s assume A and B have agreed on a commercial lease and have included a renegotiation clause. The contract provides for such renegotiations every two years to accommodate changes in market conditions. Maybe it would be more efficient to automate these renegotiations with a ‘Self-Driving Contract’? The AI Contract Embellisher could suggest an arrangement that ties the lease terms, especially the rent, to a set of relevant commercial variables that are continuously updated over time.

AI could also be used to resolve ambiguities in existing contracts. Ambiguities create uncertainty and generate costs when disputes arise over their interpretation. A contract with an efficient mechanism for resolving ambiguities is a better contract. But there is a problem. If the parties are already in dispute about a specific interpretative issue, a particular AI-generated interpretation may benefit A at B’s expense. Why should B accept it? The solution might be to agree on a binding AI interpretation mechanism before a particular dispute arises. This could happen when concluding the contract or subsequently, by amending it. At that stage, the parties are behind a partial ‘veil of ignorance’: they do not yet know what disputes will arise or who will benefit from a particular interpretation.

This example also leads to a natural extension from AI as ‘Contract Embellisher’ to AI as a negotiation assistant. As discussed, parties may initially be more willing to use AI to improve an existing contract because they already have an agreement to fall back on. But as increasingly sophisticated and competitively priced AI Contract Embellishers emerge, why not use these tools to negotiate new contracts in the first place? The idea is similar to deal-making mediation, but potentially much cheaper and more sophisticated. A jointly appointed AI tool could question the parties privately about their interests, preferences, and constraints, identify opportunities to create value, and suggest mutually beneficial contract structures and agreements. The parties would remain free to accept or reject its proposals. Of course, they would still have to solve the distributional problem. One possible approach is Barry Nalebuff’s ‘Split the Pie’ principle, under which the additional value created by cooperation is divided equally between the parties.

But why stop at AI as a negotiation assistant? One can also imagine a fully automated ‘Negotiation Machine’. The parties would authorize the machine to negotiate and conclude a contract on their behalf, based on their stated interests, preferences, and constraints. Unlike with an AI assistant, they would relinquish control over the final decision to accept or reject a particular agreement. Each party would have to decide in advance how hard a bargain it wants the machine to drive on its behalf, creating the risk of bargaining failure. But the potential benefits are substantial: dramatically lower transaction costs and, quite possibly, better agreements. Machines could also overcome some of the psychological obstacles that make human negotiations difficult, such as negative emotions, mistrust, and reactive devaluation.

There is a broader implication. In an article on AI and negotiation published two years ago, ‘Game Over’, I envisaged a future in which large corporations deploy sophisticated, customized AI negotiation tools to outmanoeuvre smaller businesses and consumers, capturing much of the additional value created by AI. The little guy would be left fighting back with ChatGPT. But perhaps the future will look different, at least in certain markets. What if some businesses start offering their customers access to jointly used AI ‘honest brokers’ that create more value and divide it fairly? Other businesses might then come under competitive pressure to follow suit rather than lose market share. Providers of such tools would compete on quality, reliability, fairness, and price. In essence, whether AI Negotiation Machines will serve the common good may depend crucially on competition at two levels: between providers of these machines and in the markets for the goods and services being negotiated.

Horst Eidenmueller is Statutory Professor for Commercial Law at the University of Oxford and Professorial Fellow of St Hugh’s College, Oxford.