AI adoption belongs in the non-billable work
Why law firms should start with filing, deadlines and KYC rather than the brief, and how legal-ops-germany puts that into practice.
The obvious question when a law firm considers AI is whether the model can write the brief. It is obvious, and it leads past the place where there is something to gain first.
Where should a firm start with AI?
A substantial part of a firm’s working day consists of tasks nobody bills and somebody still has to do: sorting incoming post and matching it to the right file, calculating and entering deadlines, checking matters for conflicts of interest and money laundering risk, coordinating appointments, reconstructing time entries, chasing outstanding invoices.
This work has three properties that make it a better entry point than the legal core business. It is rule based, its results are checkable, and an error shows up on a second read instead of hiding inside a chain of argument. Legal assessment is the exact opposite: its quality is precisely what cannot be read off the result alone.
Out of that observation came a project I have been working on for a while: legal-ops-germany,
an open library of tools for the non-billable work of German law firms. The focus on the German
market is not a detail but the point: deadline logic, fee law and money laundering due
diligence cannot be translated in from another jurisdiction.
Calculate, don’t guess
Where this work runs is no side question: it presupposes a hardened AWS account (Amazon Web Services, Amazon’s cloud platform) and EU hosting locked down technically.
The most important design principle concerns numbers. Wherever deadlines, amounts or fees are involved, the calculation is done by a deterministic piece of code, not by the language model. The model understands the task, maps the inputs and phrases the result. The calculation itself runs through code that produces identical output for identical input and lays the full working out in the open.
The difference is practical, not ideological. A language model produces the statistically most plausible continuation. For a deadline calculation under §§ 186 ff. BGB (the German Civil Code rules on computing periods) the most plausible continuation is usually right, and “usually” is the wrong category for deadlines. Working that sets out the start of the period, its end and the shift for a weekend or a public holiday in the relevant federal state, each on its own line, can be verified in thirty seconds. A number without the working can only be believed.
The same goes for fee calculations and for the question of which amount came from which document. Where a value is not literally supported in the source document, it should not appear in the result either.
Human in the loop is a presentation problem
Everyone says a human decides at the end. It gets interesting at the question of whether the way the result is presented makes that decision possible at all.
A proposal that turns up as a finished result gets waved through. A proposal that ships its input data, its working and its uncertainties gets checked. So the tools deliver intermediate steps, sources and, explicitly, gaps: which mandatory detail was missing from the document, which hit is only a possible hit, which assumption was made. Sign-off is then a real decision and not a formality.
Responsibility under professional conduct rules stays with the admitted lawyers regardless. A tool cannot take that on, but it can supply the basis on which it can sensibly be exercised.
No network
The tools work without web access. No crawler, no search, no enrichment from external sources at runtime. That follows partly from professional conduct rules, because some firms have good reasons not to allow direct web access in the working environment. It is also a question of traceability: a result that depends on an external page reachable today and changed tomorrow is not reproducible. Sanctions lists and similar data are therefore kept as local files and updated deliberately, not fetched in the background.
Why is the project open source?
The project is open source, for a practical reason. If you entrust a deadline calculation to a tool, you should be able to read how it calculates. With closed software all that remains is trust in an assurance; with open code the working is in the repository (the project’s public code store).
It is explicitly a work in progress, and criticism is the real benefit of publishing. What helps most is the concrete observation from daily practice: which non-legal work actually causes the greatest effort. That question cannot be guessed from the outside.
The content of this blog was created partly with the support of AI.