A new format stopped being a project.
Their product already extracted corporate documentation, but every new format meant retraining classifiers or writing rules. We rebuilt the pipelines as agents on language models: an unseen document takes the same path as any other, every field arrives with its source page, and anything inconclusive goes to a person.
- Context
- Legal SaaS - live product
- Window
- Phased delivery, format by format
- Delivery
- Extraction pipelines rebuilt as agents on language models, integrated into the existing product, with a trace for every extraction.
Where each figure comes from is detailed below.
Where it came from
The product was already live and solved the hard part: sourcing and structuring corporate and representation documentation. The toll was on growth: bringing in external documents ran on a pipeline of BERT-style classifiers and template-and-rule extraction, which worked as long as the document looked like the template. Every new format took model or rule work before it was covered — and in a domain where the document calls the shots, the template is always late.
The approach
We rebuilt those pipelines as agents on language models, and the change is one of kind: instead of classifying against a closed catalog, the agent reads the document, works out what it is looking at and extracts the fields by reasoning over the text. An unseen format stops being the exception that breaks the pipeline: it takes the same path as everything else. And because the domain is legal, confidence is not assumed: every field comes back with the page and passage it was taken from, and when the extraction is inconclusive a person decides before anything reaches the case file.
What reached production
The AI features live inside the product their customers were already using, with no separate tool: a document goes into the case and every extracted field shows where it came from. And covering a new format is no longer a modelling project: the schema and the evaluation cases are extended.
Delivery components
- Extraction agents on language models
- Source citations and human review
- Evaluations over real documents on every change
Anonymized case. We work under NDA: we do not publish the client name, their logo or any detail that would identify them. We discuss the technical detail on a call.
Shall we talk about your project?
In 30 minutes we review what you want to build or what you already have running, the data, the integrations and what it still needs to remain reliable once users arrive. Whether the project is yours or a client’s, the call is the same.
Free 30-minute review, no commitment. We sign an NDA beforehand if you need it. We reply the next business day.