Legal AI vs The Adversarial Incentive

Most major law firms use legal AI tools, yet there has been little significant change to how those firms operate. Why is that? The answer is: the adversarial incentive. This factor explains how the entire legal sector is impacted by legal AI.

Framing The Adversarial Incentive

How is it possible that generative AI technology that provides so much time saving and that delivers so much efficiency, especially in legal work which is grounded in text and language understanding, is still having so little impact on how many law firms operate?

Artificial Lawyer has watched with great excitement as firm after firm brought in AI tools, platforms, point solutions, general models and more. All capable of allowing lawyers to deliver their work to the clients in accelerated ways that were only an hypothesis some years ago.  

And then this site has also watched as it has become apparent that all this AI investment is not moving the dial – at least for many law firms. Why is this? The answer is as follows.

To understand the problem this site looked at the available ideas that might explain this phenomenon. As seen below, the ideas helped explain part of the challenge, but didn’t get to the root of it….until, after an afternoon of brainstorming this site’s founder came up with an axiom that seems to work and gets to the heart of the matter.

Here’s how the logic flowed via a series of possible solutions:

• Jevons’ Paradox: as technology increases the efficiency of using a resource (e.g. more efficient internet download speeds), total consumption of that resource increases.

Although this helps to explain why legal AI is leading to more legal matters in general, as it’s now easier for everyone to launch a case, or as one might say, ‘with ChatGPT everyone is a lawyer now’, it still doesn’t explain why major law firms are not seeing transformative efficiency gains.

• Amdahl’s law: Improving one component of a process has limited impact if other components remain unchanged.

This is a good one, but it also doesn’t answer the whole challenge. Why? Because today, legal AI can impact in some way most of what a law firm does. It’s not just a case of AI helps with one little bit, and then has no other influence across the rest of the work.

• Parkinson’s law: Work can expand to fill the time available. I.e. if you give someone one week to do something, guess what, the work will take one week.

This is also a neat one and initially was framed in relation to bureaucracy. It also helps somewhat with legal work. E.g. if a client expects a contract review will take three days, then the law firm will spend three days on it. And of course, as legal work has no inherent borders, then lawyers can easily fill those three days checking everything possible in the time allotted, even if they could get it done faster. But, it also doesn’t explain the whole picture.

• Time-Based Legal Labour Pyramid: this is one you’ll have seen AL write about since 2016. But, even that doesn’t fully capture what’s going on here.

The core economics of the ‘Big Law Business Model’, which evolved in the 1970s and 1980s and then continued largely unchanged since then, does indeed contribute to the reduction in legal AI’s impact. I.e. if you sell time and the work is conducted by a pyramid of legal labour, then even if you use AI tools across the firm – and in every aspect of that work – you remain under economic pressure and are incentivised to keep billing time to hit your targets.

But, after a lot of reflection, even this site has to say that the pyramid alone, as well as the other ideas, are not enough to explain the entire phenomenon.

So what is it? And Artificial Lawyer has got to say it really took some theory crunching to get beyond the above ideas, all of which make a lot of sense, yet don’t quite get to the central issue related to how lawyers work and how AI may, or may not, change that work.

The Law of Adversarial Incentives

So, how can we frame what’s happening? Here’s this site’s attempt at putting it into an axiom that should work for any legal activity, litigation or transactional work, and whether it’s work done by a High Street law firm, a new model AI-native firm, or a giant global firm.

This site is going to call it ‘The Legal AI Law of Adversarial Incentives‘, (and if you don’t mind, perhaps even call it Tromans’ Law of Adversarial Incentives)….it goes like this:

The greater the adversarial incentive in a legal interaction, the more legal AI efficiency gains drive additional work.

Conversely, the smaller the adversarial incentive, the greater the reduction in total work, as AI gains do not lead to additional inputs.

This is independent of whether the matter involves litigation or transactional work.

–

Explanation and Examples of Adversarial Incentives

The table below helps to illustrate the axiom:

Here are some examples:

Two very large companies are locked in a major dispute. Neither wishes to back down and both have plenty of resources to fund a legal battle. Any AI tools that either side use simply empower the lawyers to ramp up the battle. Legal research tools that help develop new lines of attack don’t shorten the litigation, they lead to multiple new avenues being explored. The other side responds in kind. Using AI to dig deeper in discovery also doesn’t reduce the length of the conflict – it actually surfaces more and more ideas and details that can go into the fight. The lawyers on the other side then have to counter. And so on.

In short, AI here is just like a military technological breakthrough that two armies at war both have access to. It actually widens the war, prolongs it even.

Now let’s take an example at the other end of the spectrum. A small startup, with no internal lawyers needs to get some basic commercial contracts reviewed. If it could avoid even working with a lawyer it would. But, the CEO is sensible and realises they need legal input. They turn to an AI-native firm as it promises speed, and lower costs. Here AI seems to work to its full efficiency extent. Why?

The answer is that there is very little adversarial incentive at play – although the total incentive is not zero – as the client can still push back on what the lawyer says in the contract review, maybe they even double check it with ChatGPT? On the other side, let’s say it is a large organisation or regulator that equally has little interest in fighting anything, they want this done quickly as well. For them, those contracts are also a formality, a ‘tick-box’ exercise. In short, there is almost zero adversarial incentive, so legal AI effects are at their maximum potential.

And let’s look at something in the middle. A medium-size regional law firm is doing an M&A deal for an equally medium-scale company. There is a lot at stake – it’s a business with shareholders, first generation owners, there is a buyer, there are loans and more involved. It’s complex and there is plenty worth arguing over, or rather negotiating about.

But, as neither party has endless funds to pay for the legal work here, nor sees this transaction as hostile in any way, there is far less incentive to battle over every point. Where AI can have some immediate effect it’s used, e.g. due diligence. When that work surfaces issues that need to be explored, the clients are happy – up to a point – to allow their lawyers to keep working. But – and this is crucial – neither party wants its lawyers to do any more than what is essential to get that deal over the line and tick those regulatory boxes, and ensure their interests are suitably taken into account.

In short, AI comes into play, its speed and efficiency in surfacing issues to explore does generate some more work, as both parties do want to do a legally sound deal and they are not fools: they naturally want what’s best for them and their shareholders. But, the incentive to keep battling away on every single minor point just isn’t there. AI in this case does result in a slightly faster deal, with perhaps slightly fewer hours billed. But, to an outside observer they still would not notice that much change.

In fact, of the three examples we can see that the firm in the middle is most challenged by AI. The large firm, which represents giant companies with endless legal budgets, only uses AI to help drive even more regular lawyer work. The far smaller new model firm handles work that no-one wants to fight over, and so their AI usage is also a great fit for its business model of fixed fees.

But the medium-scale firm, with clients that don’t have limitless budgets, who do care very much about costs, and do want to see some acceleration of work, are in a tricky position. Clients want them to use AI, so they do, and some of that use does elongate and deepen the work that follows – but in some areas AI really does curtail the need to keep working on that aspect, and importantly the adversarial incentive is not that strong here. The clients ‘want the job done’ – carefully, correctly, which meets their needs, yes – but the desire to go on exhaustively is not there. Thus, AI does indeed reduce the net hours put into the deal.

Thus, law firms that have business models that can’t easily adapt, or are not immune, or cannot fully benefit from AI efficiency, have a challenge – and AL will explore this later.

Conclusion

When adversarial incentives are low, then legal AI’s net economic impact is high.

When adversarial incentives are high, then legal AI’s net economic impact is low.

Across the entire spectrum of legal work, it’s the clients who ultimately decide this. AI is powerful, but is still subject to the demands of the clients for that particular matter.

An NDA can take a few minutes to review, if the parties accept it, e.g. they use the famous OneNDA. Or, they can argue for months about it, if they wish to and are willing to pay the lawyers on either side to keep going. It’s all in the incentive of the client.

The challenge for law firms is how they fit their business model around the incentives created by the clients and the impact of AI.

All law firms will experience both ends of the spectrum. An AI-native firm will occasionally find itself handling a matter that the client wants far more than a fast review for. That won’t fit their business model. Likewise, major firms will encounter clients who even if massive in size occasionally just ‘want it done’ and want maximum AI use and maximum AI benefits from that efficiency gain. Such firms – working on the hour – will also find that tricky unless they can build internal systems for this.

Overall, the adversarial incentive when combined with legal AI is already shaping the global legal market. This effect will only be felt more strongly.

Richard Tromans, Founder, Artificial Lawyer and Chair of the Legal Innovators conferences, October 2026

—