Fractional AI departments for law firms, privileged by default.

Every off-the-shelf AI tool on the market today is a cloud endpoint. Your associate types a question, your matter content leaves your network, it lands on a server you do not own, the answer comes back. The vendor signed a terms of service that says they will not train on it. The architecture says it crossed the perimeter anyway. For a marketing team that is fine. For a law firm that just routed a privileged communication, a deposition transcript, or a work-product memo through a third party, the partners would not sign off if they knew.

Attorney-client privilege does not survive a thoughtless API call. Work-product doctrine does not protect documents that a vendor logged for thirty days under a standard retention policy. The conflict-check rules at most state bars do not let you let a generic LLM see who you represent against whom. The unauthorized practice of law guidance in most jurisdictions does not let a chatbot give legal advice on your website. None of these rules are about whether AI is useful. They are about whether the firm can defend the workflow if anyone audits it.

So most firms do nothing. The youngest associate has a personal ChatGPT account and uses it for first drafts they would never admit to. The website still has a static intake form that converts under one percent. The marketing budget goes to a junior person who posts on LinkedIn between billable hours. The precedent library is locked in Word documents across server shares that nobody can search efficiently. The firm is leaking time on every front and the only sanctioned answer is to keep doing it the same way until somebody figures out a defensible version. That is what we build.

The split is not philosophical. It is operational. Matter content (privileged communications, work-product memos, deposition prep, contract review, client correspondence) sits behind a wall. It runs on hardware the firm controls, inside the firm network, with no public endpoint in the loop. Sanitized work (BD outreach to general counsel, jurisdiction-aware SEO articles, intake chat that never gives advice, billing reconciliation against time entries) runs on cloud agents because none of that data is privileged in the first place.

The architecture decides what goes where. A partner reviewing a draft motion against ten years of internal precedent never leaves the firm network. A BD agent emailing a GC at a target corporate client never sees a matter file. The intake chatbot on the public website cannot retrieve a single line of work product because it does not have permission to. Each agent runs in the context the firm scoped for it, with the data access the firm authorized, under retention and disposition policies the managing partner signed off on.

We built this configuration first for fintech and healthcare clients where the cloud is genuinely off the table. The legal version is more nuanced because not every workflow at a firm needs the on-device treatment. BD does not. Content marketing does not. Intake routing does not. The matter side does, and the hybrid model lets us put each workload on the right rail without forcing the whole firm to pay for an air-gapped install. The Local Agent Setup page covers the technical mechanics of the on-device half. This page covers what changes when the whole firm runs on a fractional model.

Every firm we work with has the same locked asset. Ten or twenty years of briefs, motions, memos, deposition outlines, and client letters, written by partners who have left or are no longer reachable on a Tuesday afternoon. The work product is brilliant. It is also unsearchable. An associate drafting a new motion knows there is a similar one from three years ago, but they would need to remember which partner wrote it and what client it was for to even begin looking. Most of the time they start from scratch and the firm pays for the same work twice.

A legal-tuned ops agent reads the entire archive once. Indexes by issue type, jurisdiction, opposing counsel, judge, outcome, partner of record. Respects matter-level permissions so an associate on the litigation side cannot accidentally retrieve a transactional matter they were never staffed on. When the same associate now asks "what is our best argument for tolling under California Code 340.6 on a malpractice claim," the agent returns the three best precedents from the firm archive, the partner who wrote each one, and the section of the brief that contains the relevant analysis.

The associate still does the legal work. The agent does not write the brief. It surfaces what the firm already produced, in the format the associate needs, with citations to the source documents. The partner reviews and the work product is grounded in actual firm precedent, not a generic LLM hallucination of what the law might say. This is the workflow that pays for the whole engagement at most firms because the time return is measurable from week one and the partners see it on their own matters before the end of the first month.

The same index quietly powers the BD side. When the conflict-check workflow needs to verify whether a prospect overlaps with an existing matter, it queries the same indexed archive instead of a separate database the firm has to maintain in parallel. When a partner is preparing for a pitch to a target GC, the agent pulls every motion, brief, and client alert the firm has written on the regulatory regime that GC operates under. The same precedent corpus serves four different workflows because the index respects the same matter-level permissions across all of them.

For firms wanting to understand whether the on-device half makes sense for their matter content before committing to the full engagement, the AI Consultancy audit is the front door. It is a written recommendation on what should run on-device, what can run on cloud, and what the compliance defensibility looks like for either path.