Already in use
ChatGPT and its peers are already in the building; the question is how to use them on client and claimant data without putting confidentiality at risk.
Working confidently and defensibly with AI
A two-day workshop for law firms: day one builds genuine LLM and prompting expertise, day two turns it into a working prototype for a matter you are running today.
Designed and delivered by a postdoctoral AI researcher at the University of Oxford.
Day 1 · Training
Method for everyone
How LLMs work and how to direct them professionally, transferable to any task.
Day 2 · Consulting
Your own matter
Theory becomes a working prototype for a task that occupies your team today.
Why now
AI has arrived in legal practice – defensible use is what counts now
AI tools are everywhere. The question is no longer whether they are used, but how they can be used responsibly on client and claimant data.
Real risks
Hallucinated authorities and data protection are not footnotes. Both have to be actively controlled, not hoped away.
Concrete potential
Used correctly, AI absorbs the time-consuming preparatory work – the decision stays with the fee-earner.
The concept
Day 1 · Training
Day 2 · Consulting
Two days, two modes
Training that actually transfers – and a consulting day that leaves you with a concrete result.
Method for everyone
Conventional training: how do LLMs work, and how do you direct them professionally? The techniques taught transfer to any task.
Your own matter
Theory becomes a working prototype for a task your firm is dealing with today.
Day 1 · Training
Every block of theory is applied immediately in a hands-on exercise.
Fundamentals and prompting – comprehensible and practical
From how modern language models actually work through to professional prompting frameworks. Every block of theory is applied immediately.
LLM fundamentals – without the mathematics
From the principle of "super autocomplete" through tokenisation and embeddings to transformer architecture and training (pre-training and RLHF), explained in plain terms.
Strengths, limits and risks
Hallucination, knowledge cutoff, "System 1 versus System 2" and the temperature parameter: what AI does reliably and what it does not.
Hands-on with OpenWebUI
First practical steps: model selection, parameters and four guided exercises: test the strengths, compare models, provoke the weaknesses, see what temperature does.
Role prompting
Assigning the model an expert role: the trainee in their first seat, who has to be briefed precisely before the work comes back reliable.
Prompting frameworks: PARE and CO-STAR
Two structured systems for reproducible results: PARE for logic and workflow, CO-STAR for tone and external effect. Including practical exercises.
Day 2 · Consulting
From knowledge to prototype
Together we build a three-stage pipeline: collate, assess, substantiate. The pattern holds wherever the same test has to be applied to a very large number of individual files, and the result for every one of them has to remain evidenced.
Stage 1Collate
Bring the scattered material of a single file together into one structured record: intake questionnaire, documents supplied, email correspondence, statements of case and correspondence with the court.
Stage 2Assess
Test that record against the criteria that decide it: the elements to be pleaded, the limitation position, the cohort a claimant belongs to, the documents still outstanding.
Stage 3Substantiate
Produce a checkable result for each file, with a citation back to the source: what is missing, what has to be requested, what the conclusion rests on.
RAG – the knowledge base
How the model draws its answer from the material you put in front of it: one claimant's file, a contract, the disclosure set (retrieval-augmented generation, meaning document-grounded AI). The single most effective measure against invented content – though it reduces the risk rather than removing it: retrieval can return the wrong passage or miss a relevant one, and a correct source can still be reasoned about wrongly. Which is why every answer carries its citation.
Agentic search – when the model does the searching
The alternative to fixed retrieval: the model decides for itself whether to search the knowledge base, writes its own queries, and reads whole documents rather than pre-selected fragments. Strong where what to look for next only emerges from what the first search returned.
Guided reasoning
Chain-of-thought and few-shot prompting (step-by-step reasoning demonstrated through examples) break the application of a test to the facts into steps that can be checked one at a time. The reasoning the model writes out is a working aid, not evidence of how it arrived at its answer. What makes the result auditable is the chain the pipeline records: source → citation → fact → criterion → result.
German-language material
Where proceedings run against a German counterparty, much of the material arrives in German. We work with German technical and legal documentation daily, and show how to build a pipeline that summarises and cites it in English, so review does not hinge on a separate translation step in front of it.
Building and running the prototype
Assembling the three-stage pipeline together and testing it on real material from your own files.
Error analysis with self-critique
The model reviews its own reasoning for logical faults and catches a proportion of them; what it cannot do is confirm its own result. The human remains the final auditor, and the check is made against the cited source, not against the model's prose.
Limits and failure modes
What actually goes wrong: retrieval returns the wrong passage, a relevant one is missed, the source is right and the inference is not. We provoke each of these on your own prototype, show how they present in the output, and set down which decisions stay with a fee-earner whatever the model says.
Compliance and documentation
How AI-assisted output should be recorded and marked, plus a concrete roadmap for putting it to work in practice.
A worked example
Group litigation: one case, several thousand claimants
A firm runs one and the same case for several thousand clients. In England and Wales those claims are managed together under a Group Litigation Order, on a group register: the generic issues are decided once, but every claimant's individual issues still have to be established. The legal argument is therefore written once; what does not scale is the factual position of each individual claimant, spread across intake questionnaires, documents supplied, email threads and correspondence with the court. The recurring question is not what the law says; that is settled and identical for everyone. It is: where does this particular claimant's matter stand, and does it meet the criteria we have to plead? That is precisely what RAG is built for: every answer carries the citation it was drawn from, so an answer taken from the file next to it shows up rather than passing unnoticed.
Method and confidentiality
Self-contained, secure, intensive
Confidentiality is not an afterthought in this workshop; it is part of the design.
Local LLMs instead of cloud
What decides this is the architecture, not the interface: model server, embeddings, document parsing and OCR, logs and backups all run inside your own environment, with no calls out to an external service. OpenWebUI sits on top as the interface – it is what you see, not what draws the boundary.
UK GDPR and professional duties
The UK GDPR, the Data Protection Act 2018 and your confidentiality duties under the SRA Code of Conduct are addressed directly rather than worked around. Where EU parties or EU-held data are involved, the EU AI Act is covered as well.
Infrastructure you control
Works with your firm's existing local AI platform, or with a test instance we provide for the workshop.
Intensive supervision
A maximum of eight to ten participants and hands-on throughout, for genuine transfer rather than a lecture.
Who is behind this
Scientific depth, transferable method
The workshops teach established techniques – prompting frameworks such as PARE and CO-STAR, document-grounded AI (RAG) and guided reasoning – correctly, and applied to your own matter.
Dr Till Hofmann
Academic lead
Dr Till Hofmann designs the curriculum and is responsible for its content.
- Doctorate in computer science (summa cum laude), RWTH Aachen University 2023, awarded the Friedrich Wilhelm Prize and the Borchers Medal.
- Postdoctoral researcher at the University of Oxford, researching framed autonomy, a subfield of AI safety.
- Teaching: sole lecturer for the master's course "Uncertainty in Robotics" at RWTH Aachen University.
- Industry experience at Intrinsic (Google X) and Mercedes-Benz R&D (Silicon Valley).
Delivery
Dr Till Hofmann currently delivers the workshops personally. The 43 IT GmbH team of instructors is growing; each works to the same curriculum and the same quality standard.
About 43 IT GmbH
43 IT GmbH has supported courts and law firms with technical expert reports since 2020. We already work with UK firms running group litigation against German companies. Defensibility and traceability shape our work, and therefore these AI workshops too.
Format and terms
At a glance
- Duration
- Two days (approximately 09:00–16:30 each)
- Group size
- Eight to ten participants, for intensive supervision
- Prerequisites
- No technical background required; laptop and platform access
- Audience
- Fee-earners, paralegals and knowledge management staff
- Technology
- OpenWebUI / local LLM infrastructure
- Delivery
- On site at your offices or remote – agreed in the scoping call
- Fee
- On request – day two is tailored in the scoping call
Common questions
Worth knowing
Do our people need a technical background?
No. The workshop is practical throughout and assumes no IT or AI knowledge. Every concept is explained in plain terms, without the mathematics.
Is client material safe, and does this sit with our confidentiality duties?
Yes, provided the whole chain runs locally – which is how we set it up: model server, embeddings, document parsing and logs on your firm's own platform, or on the test instance we provide for the workshop. OpenWebUI is only the interface; the assurance rests on the architecture beneath it. Which components are involved in your case is set down in writing in the scoping call. The UK GDPR, the Data Protection Act 2018 and confidentiality under the SRA Code of Conduct are a fixed part of the workshop.
Does the AI replace our paralegals?
No. The AI absorbs the time-consuming preparatory work; the decision and the final check stay with the fee-earner, who remains accountable for the work. That principle runs through the entire workshop.
Can day two be tailored to our own matter?
Yes. That is the core of the offering. In the scoping call we agree together which concrete process is prototyped on the second day.
Can it handle German-language documents?
Yes, and this is a particular strength. We are a German firm that works with German technical and engineering documentation daily. Where disclosure arrives in German, the pipeline can summarise and cite it in English without a separate translation step in front of it.
What if we have no AI platform of our own?
Then we provide a test instance for the workshop; your own infrastructure is not a prerequisite. If you already run a local AI platform, we work directly on that instead. Which route fits is settled in the scoping call.
Contact
Let us talk about your matter
Arrange an initial conversation with no obligation. Tell us briefly what you are dealing with and we will come back with a proposal that fits.
Prefer to go direct? ai@43.gmbh · +49 (0)711 6333686