An Agentic Data-Handling Workflow
Client: Nemadvokat
Every residential property transaction begins with paperwork: deeds, purchase agreements, insurance documents, easements, municipal records and property registration data. Before a lawyer can draft a document or advise a client, all of this material must be read, organised and understood.
It is essential work—and it can consume much of the working day.
We joined Nemadvokat as a strategic and technical partner to explore whether an agentic workflow could take on the repetitive parts of this process, allowing lawyers to focus their time and expertise where human judgement matters most.
The problem
In residential property law, the lawyer’s day often begins with documents rather than clients. A case may contain numerous files in different formats, and relevant information can be scattered across both digital and scanned material.
Before the legal work can begin, someone must extract the text, identify important facts, check whether required information is present and organise the material into a consistent structure. Although every transaction is different, much of this administrative work follows a familiar pattern.
That raised a clear question: if the documents and outputs are broadly predictable, could a system handle those patterns while keeping the lawyer in control of every important decision?
The solution
We developed NemAgent: a working agentic document workflow designed specifically for residential property cases.
Users can upload PDF, DOCX and TXT files. Text is extracted automatically, while scanned or image-based PDFs can be processed using an OCR fallback. GPT-4o then analyses the collected material and converts unstructured document content into structured data points.
Before anything is generated, the lawyer can review and edit the extracted information. Required buyer and seller fields are validated, and missing information is clearly flagged instead of being silently inferred.
Once the data has been approved, NemAgent turns the case into a structured document workflow with more than 20 preconfigured sections, including master data, the purchase agreement, easements and other recurring areas of a property transaction.
Each section operates as a specialised agent. It receives:
The section title and description
Its generation mode
An optional template
The approved structured data
Content from the source documents assigned to that section
Depending on the task, a section can use one of three generation modes. A fixed section returns predefined text, a template section fills an established structure with case-specific information, and a prompt section uses GPT-4o to draft content from the supplied instructions and sources.
Source management gives the lawyer control over which uploaded documents each agent may use. This keeps the working context focused and helps prevent unrelated material from influencing a section.
The human workflow
The process is designed around human review rather than full autonomy.
First, the lawyer uploads the case documents. NemAgent extracts the available text and offers OCR when a file appears to be scanned or image-based.
Next, the system identifies and structures the relevant data points. The lawyer reviews these results, corrects them where necessary and resolves warnings about missing buyer or seller information.
The lawyer then chooses between automatic and manual generation and assigns relevant document sources to individual sections. In automatic mode, the complete document can be produced as a coordinated sequence. In manual mode, sections are started individually.
During generation, every section remains visible and controllable. The lawyer can review completed content, rerun a single section or restart the entire document. When the work is finished, the combined result can be copied or downloaded as a DOCX file. Master data can also be exported separately as TXT.
The agent workflow
Behind the interface, each document section is treated as an individual agent with a clearly bounded task.
In automatic mode, the agents run sequentially from the first section to the last, with a short pause between calls. In manual mode, only the selected agent is started.
Every agent moves through a visible status flow: pending, writing, completed or error. If one section fails, it is marked accordingly without stopping the remaining sequence. This makes failures local, visible and recoverable.
The generated text is written back to its section, where it can be reviewed and rerun without affecting approved content elsewhere in the document.
The section configuration can also be imported or exported as JSON, making the workflow portable and easier to maintain. A Markdown overview of the available data points supports documentation and prompt development. Keyboard shortcuts provide quick access to upload, workflow and configuration views.
The outcome
Work that once required days of reading, sorting and drafting can now be completed in minutes.
NemAgent does not remove the lawyer from the process. It changes where the lawyer’s time is spent. Repetitive document handling, data extraction and first-draft production are carried by the workflow, while validation, approval and legal judgement remain with the professional.
The result is a practical collaboration between people and specialised agents: the system processes recurring patterns, and the lawyer handles the exceptions, implications and decisions that require expertise.
The point
AI is at its strongest when it takes on well-defined work that follows familiar patterns.
NemAgent shows how an agentic workflow can transform document-heavy legal work without turning it into an opaque, fully automated process. Each agent has a limited responsibility, each source can be controlled, every output can be reviewed, and errors remain visible and recoverable.
The time saved does not disappear into the system. It returns to the clients—and to the legal craft they came for.
