Building Legal AI Lawyers Can Actually Trust

As AI adoption accelerates across the legal sector, the conversation is shifting from automation alone to a more fundamental question: how can legal technology become genuinely trustworthy, usable, and embedded into the way lawyers work?

Martina Domenicali, Co-Founder and Chief Revenue Officer at Lexroom, shares her perspective on why the future of legal AI depends on verifiable data, jurisdiction-specific accuracy, and systems designed alongside legal professionals rather than imposed on them.

In this Q&A, Martina explores interoperability, workflow design, legal operations, and why the most effective AI tools are those that strengthen human judgment rather than attempt to replace it.

Enjoy the interview below.

---

1. “Work Smart, Not Hard”: How do you see AI changing the way legal teams collaborate and manage workflows internally?

The biggest shift isn't about individual productivity: it's about how knowledge moves through a team. Right now, in most law firms and in-house departments, knowledge is trapped: in someone's head, in a folder no one else can find, in a research memo that gets redone from scratch six months later because nobody knew it existed. AI breaks that bottleneck, but only on one condition: it has to be AI that lawyers can actually trust. A system that always cites its sources, that lets you verify every output, and that doesn't hallucinate is the only kind lawyers will genuinely integrate into their workflow. 

When that condition is met, the change is structural. Junior associates stop spending three days on research that should take an afternoon, and senior lawyers stop being the human search engine for their team. The time that gets freed up isn't compressed, it is redirected towards what actually matters: analysis, strategy, judgment. That's what "work smart" really means: not doing the same things faster, but redistributing cognitive work along the seniority pyramid of the firm. And it only happens when the technology is designed with lawyers, not sold to them from above.

2. What does a truly interoperable legal tech stack look like in practice?

Honestly, most firms are still far from a truly interoperable stack. What gets called a "tech stack" today is usually a collection of disconnected subscriptions: a document management system, a billing tool, maybe a generic AI assistant layered on top of everything, and multiple email threads holding it all together. It's a market that has historically evolved more through established processes than through technology architecture.

True interoperability means your research tool talks to your drafting environment, your matter management system knows what research has already been done, and nothing requires manual re-entry.

The prerequisite is indeed having a core layer that's actually specialized: not a generalist AI trying to do everything, but a legal intelligence layer built from the data up, with verified, constantly updated sources (laws, case law, filings) and citable outputs that other tools can plug into reliably. Interoperability requires infrastructure designed data-first, not model-first.

3. How can legal technology help bridge operational and regulatory complexity across jurisdictions?

Regulatory fragmentation across Europe is mostly a structural problem. European legal work today runs across 27 different legal systems, each with its own language, its own codes, and its own case law. The same type of contract, the same GDPR obligation, the same employment clause can have materially different implications in Italy, Germany, or Spain.

A tool that gives you a confident-sounding answer without understanding jurisdictional specificity can be truly dangerous. This is exactly why legal AI built for common law markets (designed in English, on top of Anglo-Saxon precedent) doesn't work for continental Europe, which is a civil law world where the starting point is the codified rule, not the case.

What's needed is depth per jurisdiction, not surface-level breadth across all of them. The right approach is to build genuinely deep, locally curated databases, market by market, with sources verified by jurists who know that system, and then make them accessible through a common interface.

It sounds like a paradox, but it's precisely hyperlocalisation that makes the model scalable: you first build a complete and accurate legal database for a given jurisdiction, then you put the AI on top of it. Once validated, that playbook gets replicated. Pan-European ambition has to be grounded in local precision, not the other way around.

4. Innovation often fails not because of technology, but because of culture. What strategies have you seen work best when encouraging adoption within legal teams?

The resistance is first and foremost about trust and professional identity. A lawyer's reputation is built on being right, and anything that introduces a perceived risk of being wrong is a threat. So the adoption strategies that actually work aren't the ones that ask lawyers to replace their judgment, but the ones that let them validate it faster.

When the tool is framed as something that confirms your analysis, catches what you might have missed, or finds the case that strengthens your argument (rather than something that thinks for you), adoption happens naturally.

But the real shift happens when you stop selling technology to firms and start co-building it with them. Involving top-tier law firms in the design of the vertical modules for each practice area changes the dynamic entirely: they're no longer clients receiving a tool, they're partners shaping it. To innovate a structurally conservative sector, you first have to speak its language.

The other thing that matters enormously is output quality. If a tool produces something citable, sourced, and accurate, people use it. If it produces something that sounds plausible but still has to be checked by hand, they stop.

5. How do you balance automation with the need for legal professionals to maintain strategic oversight and human judgment?

“Balance” is not quite the right perspective. It implies a trade-off, as if more automation necessarily meant less judgment. But the goal is exactly the opposite: automation should absorb the tasks that don't require judgment, so that it amplifies the professional's critical judgment. Not replacing it, but freeing time and cognitive capacity for the things that genuinely require it.

Strategic risk assessment, the client relationship, the ethical call, the interpretation of an ambiguous rule in a novel situation: none of that should be automated.

What can and should be automated is information retrieval, cross-referencing of sources, the first draft of standard clauses, the flagging of inconsistencies. When those tasks are handled, lawyers can actually exercise judgment rather than spend it on mechanical work.

The right model is the human-in-the-loop approach: the human stays at the center, validates and interprets what the machine produces. Professional responsibility is therefore reinforced, because every output is verifiable back to its source. Technology amplifies human work; it doesn't take responsibility away from it.

6. What are the biggest inefficiencies still limiting legal teams today?

Three, in order of impact.

First: unstructured knowledge. Valuable work product that disappears into files and folders and gets recreated from scratch instead of being reused. Every firm holds within itself a body of analyses, opinions, and research already produced but that stays inaccessible simply because no one knows where to find it.

Second: research overhead. The ratio between time spent finding information and time spent actually applying it is still absurd in most firms. A lawyer should spend the majority of their time on analysis and strategy, not on the path that gets them to the right information to analyze.

Third: the gap between what AI tools promise and what they actually deliver in a legal context. Generalist AI has raised expectations dramatically, but in legal work "almost right" is genuinely dangerous. The inefficiency created by having to manually verify everything an AI produces can easily be worse than not using it at all. This is precisely the difference between a model-first approach (a generalist LLM given a “legal coat of paint”) and a data-first approach, in which every output is already anchored to a verifiable source.

There is also a fourth inefficiency, more structural in nature: the pricing model itself. When technology drastically compresses the time needed for a given activity, the billable-hours model becomes a brake: it rewards those who are slower and penalizes those who deliver better and faster. It's not an inefficiency that gets solved by a tool: it requires firms to rethink the value of what they sell. But it is probably the deepest change AI is triggering in the profession.

7. How do you see the relationship evolving between lawyers, legal operations, and technology providers?

What's emerging is a tripartite working relationship: lawyers setting the quality and accuracy standards, legal ops translating those standards into workflow requirements, and technology providers who understand enough about how law actually works to build on those requirements.

The providers who will win in the long term are the ones who treat lawyers as genuine partners in product development: building with them, not for them. This doesn't just mean "collecting feedback": it means designing the vertical modules together, practice area by practice area, with top-tier firms inside the process from day one. It's the difference between a product lawyers tolerate and a product they recognize as their own.

This requires having real legal expertise in-house, not just software engineers. And it requires a precise architectural choice: building data-first, not model-first. Because the most reliable signal to tell whether a legal tech company has actually understood the craft is one and only one: can its outputs be cited in a court filing (after the rightful checks)? If the answer is yes, they've earned a seat at the table. If the answer is no, then it’s not the right choice.

---

Martina will be joining Future Lawyer Europe Italy 2.0 Day 1 for the standalone session “Building the European Standard for Legal AI” and the panel discussion “Defining Your AI Strategy: From Experiments to Firm Wide Transformation”, exploring how firms can move from isolated AI initiatives to scalable, reliable, and practice ready legal innovation.

Register now to join the discussion at Future Lawyer Europe - Italy 2.0.