A trade-compliance firm whose best precedents were locked in twenty years of files no one could search
Built a private, locally-run assistant that drafts filings from the firm's own twenty-year archive of precedents and watches every client entity's compliance calendar, alerting on each deadline with the rule behind it and the document to prepare. It runs on the firm's own hardware, so the work continues even when the connection drops and no client file ever leaves the building.
A small trade-compliance firm runs on its own history. Every new declaration, classification, or filing is a variation on something the firm has already done well, somewhere in roughly twenty years of documents, correspondence, and client dossiers spread across a shared drive and an email archive. The partner had tested public AI tools and found them genuinely useful on a single document: catching a bad cross-reference, tidying definitions, drafting a clause from a prior agreement. But on the firm's real archive they fell apart. The context filled up, quality degraded on a large corpus, and the firm had no appetite to hand twenty years of confidential client files to an outside service it could not control or unplug.
There was a second, quieter problem. The firm tracks obligations for hundreds of client entities, each with its own recurring deadlines, filings, renewals, periodic reviews, status declarations. That calendar lived in one person's head and a scatter of reminders. A missed date is not an inconvenience in this work; it is a client's exposure. And the only safe way to stay ahead of it was to never stop checking.
We started from the firm's two hard constraints: the data does not leave the building, and the system keeps working when the internet does not. The assistant runs on the firm's own hardware, with the option to reach an external model only when the firm chooses, never by default. The engine matters less than the architecture, so we put the effort into organizing the corpus: the twenty years of filings, correspondence, and client dossiers became a structured, searchable knowledge base the assistant retrieves from, instead of trying to hold everything in a single oversized context window.
On that base sit two tools. The first drafts: ask for a declaration or a filing and it pulls the firm's closest precedents, proposes a draft that follows the firm's own house format, and leaves every line editable so the professional keeps the judgment. The second watches: a proactive compliance calendar across every client entity that flags each upcoming deadline, cites the rule behind it, and prepares the document to file from the matching prior template, so nothing depends on someone remembering to look.
Drafting a complex filing went from hunting through old files and rebuilding from memory to starting from the firm's own best prior work in minutes, about twelve hours saved on each complex file. The output reads like the firm because it is built from the firm's precedents, not a generic model's house style.
The compliance calendar stopped living in one person's head. Every client entity's deadlines are visible in one view, each alert arrives with its rule and a prepared draft, and the firm moved from checking constantly to being told in advance. The confidential archive that powers all of it never left the firm's own hardware, and the system keeps answering when the connection does not.
“On a single document the public tools were fine. Point them at twenty years of our files and they choke, and I am not handing that archive to a service I cannot unplug. The value was never the model. It was finally being able to search our own work and stay ahead of every deadline.”
Have a mess like this?
Tell us where time and money are leaking. We’ll map the fastest high-ROI win, usually a Pilot that pays for itself.