The Decline (and Fall) of Contracts' Empire?

The long-run of agentic contracting is likely to be quite weird

This Substack has described an imperialist vision of contracting’s role in our lives: more written deals, controlling a widening gyre of social relationships, enforced more or less automatically. And, looking backward, it’s the case that contracts has been on a generation-long march:¹

Not content with dominating the worlds of commercial law and finance, written contracts now govern the most common consumer and employment relationships. Everywhere we look, adhesive terms stare back: they control our lives at the market, at school, at work, on vacation, and online; they constrain our public law rights and our private law duties; and they determine procedure we use to vindicate what’s left of both. Forms, assented to on our proliferating portable screens, have never been more dominant, nor perceived to be less morally legitimate.

Today I want to consider a different perspective, sparked by a growing pile of news stories about consumers using AI to fight back against firms. Basically, if they listen hard enough, I think contracts’ emperors might be able to hear the Huns crossing the Rhine.

Consider Svetlana, New York City tenant who reported that she used ChatGPT to draft a retaliation letter after her landlord raised her rent from $1,389 to $1,395 days after she’d complained about two years of broken laundry machines. “It took a couple of attempts to get it perfect,” she said. “I prompted ChatGPT to add more legalese.” The bot drafted something that cited New York rent-stabilization code and argued the increase was retaliatory; she got the machines fixed.

Multiply this kind of self-help by a 100,000, and we’re cooking with gas. The question is what this kind of back-end AI-aided contract skirmishing will do to front-end contract design.²

---

Are contracts going to get more or less complete?

Obviously, the cost to consumers of interacting with their contracts is falling precipitiously. You no longer need a lawyer to understand what the contract says, or engage in self-help with your counterparty. For generations, lawyers’ first tool at hand at making disputes vanish was a well-drafted nasty lawyer’s letter. Now that tool is in the hands of the laity.

What follows? I see two possibilities.

Maybe contracts get more detailed. If consumers can cheaply attack ambiguity, firms will specify everything in writing. They’ll tighten every exclusion, define terms even more precisely, and spell out each procedural step. The classic incomplete-contracts prediction is that when drafting costs fall, parties produce more complete agreements: AI-assisted disputing speeds that up further by making the cost of leaving something ambiguous higher. And (obviously) completeness is easier when firms can pretty easily test their deals against nudgy consumers by having AI red team them. A semi-viral and ratio’d Twitter thread over the weekend expressed the point:

Or…maybe the precise opposite will follow: contracts get vaguer! Explicit rules are what AI attacks best. Feed a language model a clear exclusion and a sympathetic fact pattern and it will produce three reasons the exclusion doesn’t apply. So, while folks like “biglawbro” think that perfect contracting will eliminate litigation, I’m real skeptical that even a hyper-well-specified contract can be future proof. On some margin, reputational markets will control the amount of text you can write.

But vague standards— “reasonable wear and tear,” “sudden and accidental,” “in the landlord’s reasonable judgment”—are harder to attack because the fight becomes about applying the standard to facts the consumer (and thus the AI) doesn’t fully have. So, if firms wanted to walk a self-protective line against AI bots, they might themselves try to avoid bright line rules and move toward standards.

My guess is the equilibrium looks more like bifurcation.

Firms will make the grounds for denial or exclusion in consumer contracts standard-like—discretionary, fact-intensive, textually slippery, hard for an AI to attack on the face of the contract. And they will make the procedural preconditions more rule-like, more unforgiving, and more numerous. That is, we’d see a growth in strict notice windows, documentation requirements, insurance cooperation clauses, appraisal-process requirements, etc. I am skeptical there will be a net effect on litigation filing rates.

You can already see firms anticipating this. Alan Heimlich, an insurance lawyer interviewed by Insurify, predicted “a rise in early motion denials and procedural dismissals instead of substantive decisions.” He points out that AI right now is just inattentive to procedural law:

“They contain little or no reference to procedural rules. That creates significant exposure for self-represented parties.”

Translation: firms will seek to advance the scope of these procedural defaults to avoid litigating the substantive merits. Whether they’d stand up in court is a different problem — it’s my sense that procedural rules are exactly those susceptible to estoppel based arguments, for those rich enough to hire a lawyer to make them.³

Enter AI-CSR

So that’s step one: consumers with AI agents are going to make contracts more rigid procedurally but more vague substantively.

But now consider step two: firms counter consumers’ bots with their own.

Allstate has been quite direct about this. The company’s CIO Zulfi Jeevanjee told the WSJ that its 23,000 claims agents send roughly 50,000 communications a day, and the AI now writes nearly all of them, with humans only reviewing. He said that “The claim agent still looks at them just to make sure they’re accurate, but they’re not writing them anymore.” And, on the margin, the AI is a better rejector: “When these emails used to go out, even though we had standards and so on, they would include a lot of insurance jargon. They weren’t very empathetic… Claims agents would get frustrated, and so it wasn’t necessarily great communication.”⁴

Though AI insurance denials are going to get a ton of blow-back and some lawsuits, they are inevitable. Firms are going to be forced to employ armies of AI bots — despite the bad press — because consumer-side AI disrupts a carefully calibrated system for escalating complaints.

Historically, a well-drafted lawyer letter was a signal. It meant the person probably would escalate and has enough resources to persevere. But once AI makes that letter free, the signal becomes noisy. Firms stop updating on letter quality and start updating on what they can’t fake cheaply—did this person actually file the regulatory complaint, did they actually retain counsel, did they show show up in person to small claims, do they have the documentation. The median consumer’s position thus depends entirely on whether the backstop institutions she’s now credibly threatening are actually accessible to her.

The cheap-talk equilibrium is reflected in a growing market for these products. There are now thriving cottage industries selling AI-dispute templates to consumers (Security Deposit Refund Letter, $15; Insurance Claim Appeal Template, $3) and—on the very same platform—selling AI prompt vaults to landlords (AI Prompts Vault, $49.99, marketed at "self-managing landlords" with bundles like "49 AI Prompts for Novice Landlords: Navigating Tenant Eviction").

It’s a real sell-the-panning-table-don’t-mine-the-gold situation.⁵

Learning From Medicine?

Health insurance suggests a vision for consumer contracts’ future.

Insurers were using algorithmic systems to deny claims long before consumers had access to LLMs. As Professor Jennifer Oliva explained on PBS NewsHour, in 2023, about 73 million Americans on ACA plans had in-network claims denied, and less than 1% appealed. Of the small share who did appeal, a majority were successful.

Now the doctors are responding by using ChatGPT to draft appeal letters and, importantly, to make them long: “If you’re going to put all kinds of barriers up for my patients, then when I fire back, I’m going to make it very time consuming.”

And meanwhile, third-party companies are now selling AI-generated appeal letters directly to patients for $40 or $50 a pop. Oliva’s concern, which seems probably right to me, is that the equilibrium will likely benefit repeat players:

“AI makes it really easy for them to detect people who won’t appeal based on past, a long-standing past history in claims data, and on people who won’t live through an appeal based on the time that the appeals take.”

That is, if consumer contracts go down this path, it’s not the deal itself that will end up mattering, but who has access to the deeper set of data about the strategic considerations in the post-deal negotiation process.

The Long-Term Equilibrium (i.e., the next few years).

Put it all together one future path is that AI-assisted consumer law won’t make mass contracts more or less complete in any clean way. It makes contracts less relevant to real world outcomes, and shifts the weight of welfare outcomes onto three non-contractual buckets of concern.

The first is information the firm has and the consumer doesn’t. That will include an estimate of the consumer’s and tenant’s willingness to fight based on observable and implied pieces of data (gender, income, race, age and past history with other firms).

The second is procedural compliance. Whether you gave notice correctly, filed on time, cooperated with the investigation, used the right form. AI helps consumers here, and this is probably its largest concrete welfare contribution. But it’s also the layer firms are tightening fastest, in direct response to consumer AI.

The third is the public infrastructure of dispute resolution. State insurance commissioners and attorneys general, local landlord tenant regulatory boards, small claims court, tenant unions, and the media: institutions that give credibility to a consumer’s threat to escalate. Svetlana’s letter worked partly because the threat of an HCR complaint was real and the landlord knew it. If the New York Division of Housing and Community Renewal cannot absorb the volume of well-drafted complaints AI now enables, then the signal well-calibrated AI consumer bots produce becomes kind of worthless.

The Fall of Contracts’ Empire

This ends up being a real it’s good it’s bad situation.

1. It’s good! Consumers abandon fewer valid claims because procedural friction no longer eats them alive.

2. It’s bad! But contract design probably shifts against consumers in the bifurcated way I described, and firms pass AI-response costs through to prices.

3. Mabe it’s fine? Consumers pay for their own empowerment, and the distributional balance depends on whether AI-use correlates with who’s paying higher premiums or rents.

4. No, it’s probably real bad. Ultimately, Firms will transfer money (through settlements) to some AI-assisted claimants while continuing to stonewall others, producing a perverse cross-subsidy.

Net-net? Society will burn AI resources on an arms race that just redistributes surplus. Standards-setters and reasonable-consumer doctrines in contract law will begin to assume AI assistance as a baseline, penalizing those who don’t use these tools—who, despite cheap access, will continue to correlate with existing disadvantage, because knowing which AI to trust is itself unevenly distributed.

And if you follow this down the hole, it does suggest that proposals to cabin consumer contract law’s domain—however interesting they may be—are increasingly going to be academic curiosities.

That is, if the binding constraint is public dispute-resolution infrastructure, we ought to be funding state insurance regulators, simplifying small claims access, and standardizing regulatory complaint formats so AI-drafted complaints help agencies identify systemic bad actors rather than clogging queues. And we should consider requiring large landlords to produce claim-file and deduction data in discovery or by regulation, since leveling the informational playing field is going to be increasingly crucial.

AI thus may precipitate the decline and fall of contracts’ empire, at least in its consumer-facing regions. Once use of AI to work with contracts becomes normalized, the form will become a shared fiction that both sides know how to attack and defend with equal fluency. What will matter is everything the contract doesn’t say.

1. Particularly compared to obviously-moribund first-year subjects like torts and property.

2. Yonathan Arbel and Shmuel Becher saw a version of this coming in their 2022 Contracts in the Age of Smart Readers and they deserve credit for getting the technology’s consumer-empowerment potential before most of us had touched GPT-3. Their analysis asked how AI would affect readership and formation pre-signing. But for now, the real action turns out to be post-signing, in the dispute. A related work—though less optimstic!—is Noam Kolt’s Predicting Consumer Contracts also from 2022.

3. The classic estoppel cases are all procedural defaults (typically, renewal timing on a lease or mortgage) with weak merits.

4. Wildly, Allstate’s media team apparently badgered trade press to delete these quotes, perhaps in response to suits like this one.

5. The FTC’s enforcement action against DoNotPay illustrates same dynamic. When the floor of legal self-help rises so fast that quality dispersion blows up, regulators show up to police the bottom of the distribution.