AI for law firms, by email
David YuDeepsolve Legal is an AI assistant that works your firm's task inbox. You email it a task the way you'd hand one to a paralegal; it returns finished work in the same thread: drafted from your firm's own documents, disclosed as AI-prepared, and always for solicitor review.
We built it in collaboration with a Sydney law firm. This post covers how it works, how it's built, and what the results were.
Where it started
The firm is a boutique general practice, working across dispute resolution, property, corporate and family matters among others. They came to this with a specific load rather than a general interest in AI: reviewing the correspondence coming in, and drafting the letters and emails going back out. That work is high in volume, largely patterned, and still has to be right every time. It's the kind of thing that quietly eats an afternoon for everyone who touches it.
The assistant answers from the firm's own workflows, templates and letterheads, and it will intelligently pull in context from different data sources like email, practice management and document systems. The answers that come back read like an email from someone at the firm rather than a generic one. Looking deeper, this is how it's built.
The pipeline
Every task runs through the same steps. The assistant polls the firm's Outlook inbox through the Microsoft API; anything not addressed to it is left untouched.
- 1
A task lands in the inbox
The Outlook inbox your firm already uses. No new platform, no second login.
Not addressed to the assistant? Left untouched.
- 2
It acknowledges
You get a reply confirming the task is underway, rather than an inbox that goes quiet.
- 3
It reads the context it already has
The thread, the attachments, the matter files sitting in OneDrive.
Something missing? It asks you and waits, rather than inventing it.
- 4
It pulls the context it doesn't
Your playbooks and precedents, the web where that helps.
- 5
It drafts
Written against your actual documents, not a general impression of what such a document says.
- 6
It reviews its own draft
A first pass is always too long and too hedged. This step cuts it back and checks it against your standards.
- 7
It replies in-thread
Finished work returns to the same conversation, disclosed as AI-prepared and marked for your review.
Four steps carry most of the weight:
- Acknowledge first. You know the task landed and is underway before any work starts.
- Ask rather than guess. If the file is missing a fact (a commencement date, a counterparty ABN), the assistant asks and waits. A gap in the documents never becomes an invented answer.
- Context before drafting. The thread, attachments and OneDrive matter files are read first; your playbooks and precedents are pulled where the task needs them. The draft is written against your actual documents, not the model's general impression of what such a document says.
- Review before sending. A first draft is always too long and too hedged. A second pass cuts it back and checks it against your firm's standards.
The workflow is the product as much as the model is. A model on its own will happily produce text that looks coherent and, on closer inspection, is wrong. The pipeline is what makes the output grounded, concise and safe to put in front of a lawyer.
The architecture
The assistant runs on a machine your firm controls: documents are opened and processed there, not uploaded to a remote server you don't own. Reasoning runs on Anthropic's Claude under your firm's own agreement, so retention and training terms are the ones you hold, not ones we resell.
Access follows the principle of least privilege:
- Read-only wherever read-only is enough. Write permission is requested only where a task genuinely needs it.
- A sandboxed workspace. The assistant works in an environment no employee shares. If something goes wrong it is contained to that one workspace, and the audit log shows exactly what was read and what was sent.
- One narrow way out. The only write path is a reply to the person who asked, an approved sender at your firm. Nothing goes to a client, and nothing is written back into your systems, on the assistant's own initiative. The AI-disclosure footer is built into the send path and can't be omitted.
For more on how we think about data security in legal work, see Understanding data security for law firms.
Connecting it to your firm
AI tools don’t come with any integration because it requires IT approval. That connection work is part of what we do: we integrate the assistant with your data sources ourselves, at the lowest level of access that works.
What came of it
This comes out of months of close collaboration with the firm. Their solicitors and staff put live matters through the assistant, told us where the drafts missed, and fed in the things a firm's writing actually depends on: house tone and wording, and the formatting conventions down to numbering, indenting and capitalisation.
It has since taken hundreds of hours of manual review and drafting off the firm, work that used to sit in an inbox waiting for someone to have a clear afternoon. AI has changed the way we work, and we hope to help law firms adapt to this new landscape safely and securely.