Strategy · 12 November 2026
Solid ground
A note before you start: writing this, we went down a lot of rabbit holes, as we usually do. The main text says what we believe and where we’re heading. The small arrows open the rabbit holes, for anyone who wants to follow us down.
In early 2026, we came back from a business trip to Spain and all of us, the four founders, were having an existential crisis.
Affine was starting to succeed with clients, but our product was indefensible in the mid to long term. We could have taken comfort in our distribution, but product and deep tech have always been our core focus, and that was exactly where we lacked any differentiation, let alone defensibility.
This metaphor came from the memory of Indiana Jones fighting against quicksand. We went on to learn that quicksand is just ordinary ground saturated with water until it loses its structure. From the outside it looks solid, so you only know you are in it when you are already standing in it.
The instinct is to fight it, move fast and push through, and that is exactly what pulls you under. The way out is to stay calm, drop the weight you don't need, study where the ground is solid, and move there deliberately.
The legal tech market is saturated in the same way, with AI products and model releases, and barriers that took years to build now fall in weeks. The instinct is to pick vendors, ship features and match every release, but that's precisely what pulls you under.
Every new model release showed that the model providers, which used to be the foundation, were moving up into the application layer and the services layer, to reach the end user: the problem-holder. As models got better and cheaper, whatever a model could do alone became a commodity. And we were building a lot of commodities.
Model providers know their own roadmap and build for the models of the future, so any company building on top of them is permanently one step behind. Without deep tech of our own, our value would slowly tilt toward distribution, implementation and support: a services business we could have moved into, but didn't want.
Many questions came and went in our conversations, but one stayed with us and resonated, and it now guides everything we build at Affine:
What doesn’t become a commodity when the next generation of models is released?
This question is what leads us to solid ground. To answer it, we looked from two angles:
- Our marketWhat do our clients sell that won’t become a commodity with the next generation of models, and which legal businesses will thrive?
- Our productWhat can we build that each new generation of models makes more valuable?
What lawyers taught us
We have been selling to lawyers for about three years now. We have had the privilege of meeting them, sharing our concerns and vulnerabilities, and learning about theirs.
We started thinking about this through conversations with our team of lawyers, who explained to us some theories of the philosophy of law: natural law against legal positivism, the sociology behind legal systems. One author was especially interesting: H.L.A. Hart, a 20th-century legal philosopher who argued in The Concept of Law (1961) that law exists because of this chain:
- As humans, we live with scarce resources, natural vulnerability and limited altruism
- which brings conflict and coordination problems
- which brings custom and morality, two simple ways to guide our behaviour
- which work in small groups but break down at scale: they are uncertain (there is no agreed way to tell which rules count), static (custom takes a long time to change) and weakly enforced
- which brings the need for law: rules (primary rules), plus rules to identify, change and apply them (secondary rules)
- which makes peace, coordination, trade and trust between strangers possible
We came to see law as more than the rules and the machinery to identify, change and apply them, but as a way to create and ensure peace, coordination, trade and trust between strangers. Ultimately, law is the infrastructure for trust.
But that infrastructure does not run on its own. Law balances many interests at once, and its rules are general while every case is particular, so someone has to navigate that gap and answer for the result. That is how the trust holds. And that is why and how legal professionals are paid.
The framework is general enough to apply to any knowledge-intensive work, and general enough that it is easy to poke holes in. The work is never this clean. We have talked to many teams that work through these steps in very different ways, with loops, back and forth, and many failure modes: sometimes the framing step is weak, because it lacks critical context or time to think, or it never reaches the teams that structure and produce. Sometimes the context a senior associate finds while structuring never reaches the partner, who could have reframed the work. Information is scattered across many heads, each with less context and judgement than the partner who framed the matter.
- : understanding the real problem and the outcomes it requires, which is often more than law.
We learned from our first client that a problem can be framed in legal terms, which is what a law student learns, but great lawyers understand when the problem is not law: a negotiation problem, trust issues, power dynamics or ego clashes, where the lawyer’s role is to meet needs that have little to do with the best legal solution. Framing can start with a client who writes “I have this challenge”, or with something that changes the landscape: a new regulation, a new investment, a news article. It is the step where strategy, judgement and taste define the outcomes, and it is usually reserved for partners and senior associates.
- : planning the work: the requirements, what is known and what still needs to be known, the resources and the line of argument.
Once the problem is framed, structuring determines the outputs and the strategy to reach them. This is where research and context are most needed. It can send the work back to framing if the context changes, but usually it is about planning: knowing the requirements, what you know and what you still need to know, what the output should and should not do, and the best resources for the goal.
- : drafting the contract, the pleading, the email or the memo. It used to be the heaviest work, and it is the part AI tools are taking over.
It is where most of the hours were billed at law firms, though not where the outcome was decided, and where most of the resources went: junior lawyers analysing hundreds or thousands of documents, drafting and redrafting with the associates until the work was good enough for the next rung of the ladder.
- : checking the work against the law, the outcomes intended and the firm’s own standard, layer by layer up the ladder.
It happens in layers, and in loops with the other steps. First what can be checked: does every citation exist, is the norm still in force, does the judgment say what the draft claims, has a higher court since ruled the other way. Then substance: does the argument hold, is anything missing. Then the firm’s own standard: is this the best way to present the argument, does it match what good looks like. Each rung of the ladder takes a layer, so when work reaches the partner, the partner assumes every citation has been checked and focuses on outcomes and taste. That assumption rests on the people below having done the checking, and that is exactly what changes when AI does the producing.
- : putting your name to the work and answering for it. It is only part of what clients pay for, alongside judgement, experience and relationships, but it is the part that stays with a named person.
Part of what a client buys from a law firm is a liability sink: a trusted, named professional who takes on the risk and answers for it if the advice is wrong, through the firm’s reputation, professional duties and insurance. That is where the value of the signature lies. AI companies may eventually offer liability sinks of their own, taking on the risk for what their models produce or offering insurance against it, but clients would have to trust them as much as they trust a law firm, and that trust takes years to build. Until then, the more work AI produces, the more the person signing needs to know what they are signing.
What software taught us
The five steps are not that different from any other knowledge-intensive work with a ladder of responsibility and liability. Software, our own field, has roughly the same structure, and as software engineers who build AI and are being transformed by it, we believe we can see early what is coming.
In software development, the bottleneck used to be production: you built a new feature if you had the resources to produce it. Today production is effortless, so the bottleneck has moved, above all to review. Review queues pile up, and teams struggle to review all the AI-generated output. Engineers now specify the outcome and let agents take the actions, so the question has moved from “is this line correct?” to “does this meet the requirements?”. And the quality of what agents produce is bounded by the context they receive.
Whoever solves the review bottleneck changes the economics of legal work. Until now, legal work has been a trade-off between quality and speed: more checking meant more time, more people and more cost. Ease the review bottleneck and that chain starts to run the other way:
- review stops limiting how much a team can deliver and still stand behind
- so the same team can take on more work, without giving up reliability
- so quality legal counsel costs less
- so people and businesses who go without it today, because of the cost, can reach it
- so total demand for legal work grows
That last step is the Jevons paradox: more efficient steam engines increased Britain’s coal consumption, because cheaper energy found more uses. The same can apply to law: we believe many people and businesses that need quality legal counsel don’t get it because of the cost barrier. If the cost falls, people may find new uses for legal services, and demand rises.
How much work each step can get through
Where the ground is solid
Back to the question we started with: what doesn’t become a commodity when the next generation of models is released?
Our market
For our clients, the answer sits at both ends of the work. Framing depends on knowing the client, the counterparty and what is really at stake, much of which is never written down. Deciding stays human because the law requires a name on the work and someone who answers for it. That signature is part of what clients pay for.
Our product
As AI makes production in legal teams easier, the hard problem becomes: can you stand behind what it produces?
The steps in between are where each new model release does more: drafting first, and more and more of structuring and review. But what those steps need does not come with any model. No model release brings, on its own, a checked record of which norm was in force in Portugal on a given date, which version applied to the facts, or what a firm’s own standard is. Each release makes that record worth more, because more gets produced and all of it has to be checked. So our answer for Affine is the two layers legal professionals need before they can stand behind what AI produces.
Our first instinct as engineers was to redesign the system: make it easier to analyse, test and iterate, the way we would refactor a codebase. We learned the hard way that complexity cannot be designed away. The legal system was never optimised for efficiency and maintainability, as software was: it was built to serve justice, balance competing rights and interests, and accommodate the complexity of human society. In software, you redesign a system to make it easier to test. In law, simplifying the knowledge base means changing what law fundamentally is. So the knowledge layer has to respect that complexity without flattening it.
How the knowledge layer is built
From the sources to the answer
Datawhich official sources we have, and how many we can reach
If we rely on web search, as general-purpose assistants do, we have two problems: reliability, because results sometimes come from unofficial sites, and access, because an official source may not be searchable through Google at all. Other legal AI platforms use official sources too, so at this level what sets Affine apart is access: how many official sources we have added to Affine and made searchable through our own search system.
Processingrespecting each source’s own structure and links
Many platforms process their data for search. What sets ours apart is how the processing respects the structure of each source, how it connects to other sources, and the specifics of each legal system: from links between documents, such as case law and the legislation it applies, to splitting each document by its own structure, since legislation is divided into articles and judgments almost always follow the same sections.
Searchseveral ways in, built on that processing
That processing lets us build a far more specialised search system, with several ways of searching a document that give better results: navigating a set of documents through their citations, searching only documents with certain content in a certain section, or semantic search. In Affine, or inside the AI tools lawyers already use, this makes legal research more complete, correct and verifiable.
We can have the right information and still give the wrong answer, so we build for the answer:
- Completeevery relevant official source searched
- Correctanalysis that respects the legal system and the question asked
- Verifiableevery statement traceable to its source
One day, at lunch with a partner from one of our earliest clients, we heard how they saw the future of law firms with AI. Among many ideas, three stuck with us:
The quant firmmodels do the work, people stop them when it goes wrong
Like the quantitative investment firms where models screen the market and trade, with a few people watching to stop them when the numbers turn red. Most of the legal work runs through AI, at scale and at low cost. Lawyers set the limits, supervise, and step in when something looks wrong.
The venture firma few people with great judgement bet on what comes next
Like a venture fund: a small group of people with exceptional judgement who bet on what comes next. They take on problems that do not have an answer yet, such as new regulation, new kinds of transactions or disputes no one has seen before. The value is almost entirely in framing and deciding.
The travel agencyeverything prepared for you, competing on care and peace of mind
Like a good travel agency: the client says where they want to go and the firm prepares everything. The legal work is expected to be right; the firm competes on service, care and peace of mind, on what it is like to be its client.
It’s likely that none of these will be what law firms end up looking like, but it was still a useful exercise. In all three, framing and deciding remained with people, and the main value stayed in the liability sink a firm provides to its clients.
A client sends a request with a draft its own AI has written. Before anyone at the firm opens it, the review layer has gone through it: every citation checked against the law in force on that date, each judgment it relies on traced to see whether later case law or the CJEU went the other way, and the passages that fall short of the firm’s standard flagged. The partner starts from the few questions that need judgement. The associate answers with whatever AI tool the firm uses, and before the answer goes out, the review layer checks it again, against the law and against the firm’s own standard, and keeps a record of what was checked and against which version of the law. The partner signs knowing what has been verified, and spends their time on what can only be judged. The firm can match its people to the real effort each matter requires.
Who works where across the five steps
This is our solid ground: the belief that framing and deciding will remain human steps, and that we can build what lets the person who signs stand behind everything produced in between.
Affine exists to build the knowledge layer that makes AI useful and reliable for law, and the review layer that makes it trustworthy.
Trust is what legal professionals sell.
Trust is, ultimately, what law was created for.