Legal education sits at a peculiar crossroads. On one hand, it is one of the domains most hungry for modernisation — dense case law, high-stakes bar prep, regulatory content that changes overnight. On the other, it is arguably the least forgiving environment for AI to get things wrong.
When we were approached by an established EdTech provider in the legal domain to help them build an AI-powered learning experience, we were genuinely excited. Weeks into the project, that excitement became tempered by a very healthy respect for the problem. This post is about what made it hard — and what we'd do differently if we started again today.
The Promise That Brought Everyone to the Table
The original pitch was compelling, as these pitches usually are: AI that could personalise content for each learner, generate practice questions on demand, surface relevant case law instantly, and provide tutoring-style explanations to law students at any hour of the day.
The business case was real. Legal education is expensive to deliver at scale, tutor bandwidth is limited, and learner engagement with dense statutory content is a chronic problem. AI seemed tailor-made to solve it.
But as we moved from discovery into build, we ran into challenge after challenge that the original brief had not anticipated. Some were technical. Most were not.
The Five Challenges That Actually Defined the Project
Accuracy Is Not Good Enough — It Has to Be Verifiable
In most EdTech domains, an AI that is right 90% of the time is impressive. In legal education, the 10% where it's wrong can cause a student to carry a misunderstanding into an exam, a client consultation, or a courtroom. The problem isn't just accuracy — it's that learners can't easily tell when the AI is wrong. We had to build citation-first responses, where every AI-generated answer surfaced its source before its conclusion. This slowed output, but it changed the trust dynamic entirely.
The Content Was Not AI-Ready
Our client had years of high-quality course material — lectures, case summaries, mock papers. But it existed in formats that were deeply hostile to AI consumption: scanned PDFs, proprietary LMS structures, and legal prose written for human experts, not language models. The first month was almost entirely spent on content architecture. Not a line of model code was written. Founders underestimate this phase at significant cost.
Jurisdiction Is a Hidden Variable That Breaks Everything
Legal rules are not universal. A correct answer about contract formation under English law is wrong under New York law. Our client served learners across multiple jurisdictions, and the AI had no reliable way to know which rulebook was in play unless we explicitly built that context in. Every query, every piece of generated content, every practice question had to be jurisdiction-aware. This single requirement multiplied the complexity of our retrieval and prompting architecture by a factor we had not budgeted for.
Faculty Trust Was the Real Adoption Barrier
No one warned us about the internal politics. Senior instructors — many of them practising barristers and solicitors — were deeply sceptical of AI-generated legal explanations appearing under their institution's brand. And frankly, their scepticism was well-founded. The path to adoption wasn't a better model; it was a workflow where faculty could review, approve, and annotate AI outputs before they reached learners. We built an editorial layer that wasn't in the original scope. It turned out to be the most important thing we built.
Regulatory Change Makes AI Content Stale Overnight
Legal content has a shelf life measured in parliamentary sessions. A landmark Supreme Court ruling, a statutory amendment, an SRA guidance update — and suddenly a large portion of your AI's training context is not just outdated but actively misleading. We had to move away from any static fine-tuning approach and toward a retrieval-augmented architecture with a living knowledge base that the client's content team could update without engineering involvement. This was a significant architectural pivot made late in the project.
What We'd Tell Any Founder Considering This Space
The AI model is rarely your biggest problem. Your biggest problem is content governance, stakeholder trust, and the infrastructure required to keep legal knowledge current. Budget for those first.
Start with trust architecture, not feature architecture.
Legal professionals do not adopt tools they cannot audit. Before you build a feature roadmap, design the transparency layer — how will faculty see what the AI said? How will learners know what it's drawing on? How will errors be reported and corrected? These are not nice-to-haves; they are the product.
Treat content as infrastructure.
Your source material is the foundation of everything. If it's trapped in legacy formats, inconsistently structured, or not maintained with discipline, your AI will reflect that chaos. The investment in content architecture typically delivers more ROI than the investment in model sophistication.
Jurisdiction-awareness must be a first-class feature.
Do not let it be an afterthought. Design your data model, your retrieval logic, and your prompting strategy with jurisdiction as a primary dimension from day one. Retrofitting it is one of the most expensive mistakes you can make in this domain.
Build for the sceptic, not the evangelist.
In any established educational institution, the person most resistant to AI will be among the most respected in the room. Design your workflows so that their expertise is amplified by AI, not replaced by it. When sceptics become advocates — because the tool genuinely respects their judgement — adoption becomes self-sustaining.
What We Got Right
The RAG (Retrieval-Augmented Generation) architecture we landed on — where the model always grounds its answers in client-owned, jurisdiction-tagged content rather than its own parametric knowledge — proved to be the right call. It gave the client control, gave faculty confidence, and gave learners answers they could actually trace.
The editorial workflow, though it emerged from necessity rather than design, became the product's defining feature. Instructors told us it was the first AI tool they'd used where they felt like collaborators rather than bystanders. That framing — AI as a force multiplier for expert humans, not a replacement — is the only one that works in high-stakes education.
And the metrics, when they came in, were encouraging. Learner engagement with practice content increased significantly. Time-to-answer for common learner queries dropped. The client's content team was updating and expanding the knowledge base with confidence. These are lagging indicators, but they point in the right direction.
The Bigger Picture for Legal EdTech
AI is not going to stay out of legal education. The efficiency gains are too large, the learner demand too clear, and the competitive pressure too significant for institutions to ignore it indefinitely. But the providers who will win in this space are not those who move fastest — they are those who move most carefully.
The legal domain punishes overconfidence. That is, in many ways, what it teaches. The AI solutions built for it need to embody the same epistemological humility: clear about what they know, transparent about their sources, and honest about the limits of their certainty.
If you are a founder, product leader, or decision-maker in this space, we hope this is useful. The challenges are real, but so are the outcomes when you navigate them with rigour.