The Legal AI Race Is the Wrong Question

There is a conversation happening in every legal community right now, and it sounds like a product comparison. Harvey or Legora? Claude for Legal or GPT? What about open-source alternatives like Mike? Which tool will dominate legal AI? The debate has the shape of a strategic decision, which makes it feel productive, but the shape is the problem.

The question of which platform wins is the wrong starting point. I said as much in a recent thread on this exact debate, and it is worth repeating here in full, because it is the argument this whole post is built on:

The platform question is secondary to two more fundamental ones: what area of law are you practicing, and how specific is it? A generalist with a broad commercial practice has very different needs from a tax litigator or a labour lawyer with a narrow but deep specialization. And underneath that, there is the training data question: how much domain-specific, jurisdiction-specific material has the model actually seen, and how much of your own context are you bringing to it?

That is the real starting point: what area of law you practise, how specific it is, and how much of your professional judgement you want to retain in any AI-assisted workflow. Everything else follows from that.

Why the platform question keeps coming up

It is worth understanding why this question is so persistent, because it is not irrational. Lawyers are being asked to adopt AI tools by their firms, their clients, and their competitors all at once, and the pressure to choose something, soon, is real. When everything moves this fast, picking a platform feels like a decision you can make now and stop thinking about. It offers the comfort of having a position.

But the tools themselves are moving faster than any comparison can keep up with. Lawyers who tested Harvey and Legora side by side at major firms have reported noticeable improvement in both within two to three months. A comparison you make today will not hold by December. Choosing a tool as if you are choosing a long-term partner is a category error when the tools are updating on product cycles measured in weeks.

The specialization question is more honest

A generalist with a broad commercial practice has very different needs from a tax litigator or a labour lawyer with a narrow, deep specialization. The right tool, or combination of tools, follows from that, not from a general ranking of which platform is best.

Underneath that sits a question most platform comparisons skip entirely: how much domain-specific, jurisdiction-specific material has the model actually seen, and how much of your own context are you responsible for bringing to it? A foundation model trained on broad multilingual data will handle a Dutch employment law question differently from a model fine-tuned on Dutch case law specifically. Neither approach is inherently better. But which one serves you depends on what you practise, where, and how tolerant of gaps your work allows you to be.

This is especially relevant for lawyers working outside the United States. Most foundation models have seen far more US legal material than Portuguese, Spanish, or Dutch jurisprudence. A specialized platform that has invested in training on your jurisdiction may genuinely outperform a more capable general model for the specific task of citing and reasoning about local law. Or it may not, and the only way to know is to test it yourself, with your own matters, not rely on someone else's benchmark.

What a specialized platform is, and what it is not

Tools like Harvey and Legora are not simply interfaces on top of a foundation model. They add proprietary fine-tuning, retrieval pipelines, legal-specific training, source grounding, and workflow features that a raw model does not provide. That is real engineering and real value, especially for firms that want a ready-made system rather than building one internally.

But they are built on the same frontier models that are available directly, which means the underlying reasoning capability is shared. What the platform adds is the layer of specialization around it: how sources are retrieved, how outputs are checked, how the interface guides you toward a particular type of legal task. Understanding what that layer is actually doing is the difference between evaluating a tool and trusting a label.

My recommendation for anyone starting to build an AI-assisted legal practice, and the one I gave in that same thread, is to go to the foundation model layer first: build a workflow in Claude or another model before buying a specialized platform built on top of it. Not because the platform is unnecessary, but because spending time with the base gives you the understanding to evaluate what any specialized layer is genuinely adding versus what it is packaging for convenience. There is a structured way to run that comparison, testing both paths against identical, jurisdiction-specific tasks and scoring them on the same rubric, that produces a number rather than a feeling. That evaluation process matters more than which tool you end up choosing.

A third option: open source

There is a third category worth naming alongside foundation model and closed platform, because it complicates the binary. Mike, released as open-source software by a former Latham and Watkins attorney, replicates much of what Harvey and Legora do: document review, drafting, contract analysis, project vaults that bring a matter's documents into one place. The difference is that the code is public, you self-host it, and you bring your own model and API key rather than paying a platform subscription. That last part is worth understanding clearly: platform subscriptions bundle and often subsidize API costs through bulk pricing negotiations with model providers. When you bring your own key, you pay retail API rates directly, and those costs can add up quickly on document-heavy work. The software is free, the compute is not.

Where open source does change the equation is on two points that matter enormously to lawyers. First, data sovereignty: with a self-hosted platform, client documents never leave your infrastructure. For firms with strict confidentiality obligations, regulated industries, or clients who contractually prohibit third-party data processing, that is not a convenience difference, it is a compliance requirement. Second, auditability: if something goes wrong with AI-assisted legal work, being able to show exactly what code ran, what model was called, what retrieval logic was applied, and what verification checks were performed is a materially stronger position than explaining that you used a closed platform and trusted its output. Open source gives you an audit trail down to the code level.

But auditability is not the same as certification, and it is not the same as liability transfer. A self-hosted open-source tool gives you the capability to produce records. It does not give you a formal conformity stamp from a regulator, and it does not give you a vendor who contractually shares the risk when the tool is wrong. Those are three different things, and conflating them is a mistake lawyers make often enough that it is worth naming.

The honest caveat is sustainability. Mike's creator, Will Chen, has said explicitly that he has no commercialisation plans and that the project was built to start a conversation about how legal software should be distributed and priced. That is a familiar starting point. OpenAI launched in 2015 as a nonprofit with open-source intentions and no business model. By 2019 it had a capped-profit subsidiary and a billion-dollar Microsoft investment. By 2026 it is valued at over three hundred billion dollars and the 'open' is essentially gone. The pattern in AI is consistent: start open and ideological, hit the compute cost wall, take investment, introduce a commercial layer. Mike's trajectory could follow the same path, and that is not necessarily a problem. But lawyers evaluating it today should understand whether they are betting on its current state, free and community-maintained, or its likely future state, some form of paid offering with the code still readable under the AGPL licence.

Either way, the control advantage persists. Even if a commercial layer appears on top, the ability to read and audit the code does not disappear. That is the structural difference between open source and a closed platform, and for legal work, where the ability to explain how a conclusion was reached is not optional, it is a difference worth taking seriously.

The control question underneath everything

Custom workflows let you decide which sources are retrieved, how they are verified, and which checks run before you trust the output. You still do not know exactly what happens inside the model, no one does, but you have substantially more insight into and control over the surrounding system, and more importantly, you know where to look when something goes wrong.

With a platform, some of those decisions have been made for you. That can be exactly what you want. Or it can be a source of hidden risk that only surfaces when a high-stakes matter is on the line and you need to explain, in detail, how you reached a conclusion. A partner asks whether the legal reasoning accounts for a recent regulatory change. Opposing counsel challenges a citation. A client wants to understand how their exposure was assessed. What you can answer in those moments depends directly on how much of your workflow you actually understand.

The lawyers I work with who are adapting most successfully to AI are not the ones who found the right platform. They are the ones who got clear on what they were asking the AI to do at each step, and built their verification habits around that clarity. After that, the tool question almost resolves itself.

When delegation becomes risk

There is a line in every matter where delegation to an AI tool stops being a time-saving and starts being a risk. That line is not in the same place for every lawyer, every practice area, or every client. A volume task like standardising formatting across fifty contracts sits comfortably on one side. A novel regulatory question in a jurisdiction with sparse case law sits on the other. Most work is somewhere in between, and figuring out where is a judgment call that no platform can make for you.

This is not an argument against delegation. AI tools are genuinely useful for legal work, and the lawyers who refuse to adopt them are already at a disadvantage. But the value of the delegation depends entirely on knowing what was delegated and being able to verify whether the output meets the standard you would hold yourself to. A fluent, confident draft is not a reliable one by default. Hallucinated citations look exactly like real ones until someone checks. Legal reasoning that sounds correct can be structurally wrong in ways that only become visible when tested against the specific facts and jurisdiction of the matter at hand.

The practitioners who run multiple tools in parallel, and there are many who do, report the same finding: most of the quality comes from how you structure the input, not from which model you pick. That is uncomfortable for anyone looking for a straightforward product recommendation. It means the investment is not primarily in the subscription. It is in the thinking that precedes the prompt.

The question that actually matters

The legal AI market will not resolve into a single winner. It will resolve into a mix of foundation models, specialized platforms, open-source tools, and custom workflows, with each lawyer or firm assembling the combination that fits their practice, their jurisdiction, and their tolerance for opacity. The tool question is not the strategic question.

The strategic question is the one I keep coming back to whenever this debate resurfaces:

It is how much of your professional judgement you want to retain, and how much do we want to outsource our job, and when does it make sense and when is it too risky.

That line is different for every lawyer and every matter. No platform decides it for you. And the variables that shape it, whether you practise solo or with a team, how you bill, what confidentiality tiers your matters carry, what your professional indemnity cover says about AI-assisted work, are specific enough that a general recommendation from a blog post is not going to get you there.

We have built a structured framework for working through this decision. If you are a lawyer or a firm evaluating legal AI tools right now and want to make that call with clarity rather than guesswork, that conversation is one we are happy to have.