Industries Legal & Compliance
Move faster on contracts,prove every conclusion.
AI Engineering and software development for legal companies.
Speed only counts if the answer survives review, which is where fluent output becomes a liability rather than a feature. We build research, contract, and compliance systems where every answer carries the clause it came from, privilege is enforced by the architecture, and a reviewer can check the machine's work in seconds rather than redoing it.
Five DegreesCore Banking Loan Management
We helped build the loan management platform behind a Dutch core banking provider whose software ran at 40+ banks across Europe and North America. Five Degrees is now Akkuro by Topicus.
- $45B
- Annual loan volume
- $13B
- Deposits managed
- 133B+
- Assets under management
Why legal teamstrust us with delivery.
A fabricated citation never surfaces as an error message; it surfaces in front of a client, a regulator, or a judge. This is why we pay special attention to the most sensitive parts of the system.
The citation is the deliverable
Courts have sanctioned lawyers for filing an answer whose source did not exist. Fluent output without provenance is a liability, so retrieval has to trace every proposition to the clause it came from.
Privilege is an access boundary
Your matter walls, conflicts checks, and client confidentiality only hold if the software enforces them. A system that indexes across matters to give better answers is a disclosure incident with a search box on the front.
Contracts are structured, badly
The obligation you miss is in the document nobody could parse: definitions incorporated by reference, amendments that rewrite three clauses in a schedule. Extraction has to survive all of it.
The rules move underneath you
Guidance gets reissued and a threshold shifts in one jurisdiction and not the next, while a system frozen at build time keeps answering anyway. Confident answers that were true last year are worse than none.
Review time is the whole economics
First-pass review is where the cost sits, and clients have stopped paying for it by the hour. A human stays accountable either way; the question is how much of the document set they have to read first.
What legal demandsfrom engineering and AI.
01The firm's memory, searchable
Most of what a firm knows lives in closed matters, precedent folders, and the heads of the partners who ran them. Making that searchable takes retrieval where every result carries its source, date, and jurisdiction, and a system honest enough to return nothing when nothing adequate exists.
02Contracts as a live data source
Renewal dates, notice periods, indexation, liability caps, and change-of-control terms sit locked in prose the business cannot query. Meeting that demand means an obligation register kept in sync with the documents, so finding out what was agreed no longer means a file hunt.
03Regulatory change as a managed diff
Rules are amended faster than any team can re-read them, and a feed of regulatory news is not a control. The demand is a maintained map from obligations to controls and owners, where every change surfaces as a diff: what moved, what it touches, who has to act.
04Review that is defensible at scale
Document volumes grow while budgets for first-pass review shrink, and a court can ask exactly how the review was run. Prioritisation, near-duplicate detection, issue coding, and sampling-based quality control put reviewer hours where they matter, with the audit trail designed first.
Real-life stories of triumph.
Get In Touch
The feedback for Nortik engineers is truly outstanding, and we are very satisfied with the collaboration. I’m eagerly awaiting a moment to bring on additional people and expand this partnership going forward.
Nortik’s Impact
Engineering across Florence's trial operations platform: eBinders, SiteLink, eTMF, eConsent and Site Feasibility, for a network of 65,000+ research sites in 90+ countries.
Discover more of our insights.
Recent Posts
The eval suite is the product: testing LLM features like an engineer
Prompt changes without a test harness are blind bets. How we build eval suites from real traffic, and why they outlive every model you'll ever swap in.
Why AI prototypes fail in production (and how to ship one that doesn't)
Most AI prototypes stall somewhere between the demo and the deploy. The problem is rarely the model. It's the evals, guardrails, and infrastructure nobody scoped.
What an AI agent actually costs to run in production
Token bills are the visible line item. The real costs hide in the architecture: retries, tool calls, human review loops. A breakdown from live systems.
Frequently asked questions
Everything legal and compliance teams usually want to know before trusting a model with a document set. Not here? Ask us directly.
By never letting the model answer from memory. Every response is generated from retrieved passages and carries the citation, so a reviewer verifies in seconds instead of trusting a fluent paragraph.
Where retrieval finds nothing adequate, the correct output is that it found nothing. We tune and evaluate for that behaviour explicitly, because a system that always produces an answer is the failure mode this industry has already been burned by.
It stays in your tenancy, in the region you specify, and it is not used to train anyone's model. That's contractual with the providers we build on, not a policy statement.
Where the material is sensitive enough, we deploy models in your own environment so nothing leaves it at all. That's a cost and capability trade-off we'll walk through openly.
No, and anyone selling that hasn't practised. It removes the first pass of locating, extracting, comparing, and summarising, so the judgement work gets the hours instead of the search.
Accountability stays exactly where it is. What changes is how much unreviewed material sits between a question and a defensible answer.
Yes, but with the caveat that matters: jurisdiction has to be a first-class attribute in the retrieval layer, or you get authoritative-sounding answers from the wrong legal system.
Multilingual document sets are routine. Cross-jurisdiction reasoning is scoped deliberately with your lawyers, not assumed to generalise.
Every retrieval, prompt, model version, and human decision is logged in a form that reconstructs how a given output was produced, months later.
In a regulated or contentious setting the audit trail is the product. We design it at the start rather than adding logging once someone asks.
You do. Code, trained models, pipelines, and documentation are yours, along with the infrastructure access to run them without us.
We hand over a codebase your own team can pick up. Nothing we build is designed to keep you dependent on us.