Industry

AI in legal, and the IT it runs on

Contract review, discovery and drafting, plus helpdesk, security and backup, from one provider that keeps privileged material inside the firm.

Legal work is the closest thing to a native use case for language models: it is text-heavy, precedent-driven and full of high-volume review that reads the same way every time. It is also the field where sending the work to a third party carries the most direct professional risk. Senteras builds legal AI that runs on the firm's own infrastructure, so privileged and confidential material never becomes a disclosure to a model vendor. The same team runs the managed IT underneath it: helpdesk, security, backup and the systems of record, under one agreement with a response time in writing.

Book a free consultation

What makes this hard

Document review still consumes the majority of associate hours on any substantial matter, and it is the work clients most resent paying for
Confidentiality duties make most public AI tools unusable on client matters without consent that firms rarely want to ask for
Model hallucination in a citation is a sanctionable event, not a quality issue. Courts have already sanctioned lawyers for it
Firm knowledge lives in matter files nobody can search, so the same research gets redone every few years

Where AI earns its place

Contract review and abstraction

Extract obligations, dates, termination triggers and non-standard clauses from an agreement set, flagging deviations against the firm's own playbook rather than a generic template.

Hours to minutes per agreement

Discovery triage

First-pass responsiveness and privilege classification across a production set, with every call logged and a confidence score that routes uncertain documents to a human.

Review population cut by 60 to 80%

Firm knowledge retrieval

Search across closed matters, briefs and memos in natural language, with matter-level access controls enforced in the retrieval layer rather than the prompt.

Answers grounded in the firm's own precedent

Drafting from precedent

Generate a first draft from the firm's prior work on comparable matters, with every clause traceable to the document it came from.

First drafts in place of blank pages

Figures are drawn from Senteras engagements and are illustrative of typical results. Outcomes vary by data quality, infrastructure and scope.

The rules that shape the build

These are the constraints that decide the architecture, usually before anyone has picked a model. This is general information about the regulatory landscape, not legal advice on your obligations.

ABA Model Rule 1.6 (confidentiality)

Disclosing client information to a third-party model may be a disclosure requiring informed consent. ABA Formal Opinion 512 treats that as fact-specific rather than automatic, and Rule 1.6(c) separately requires reasonable efforts to prevent unauthorized access. Running the model on your own infrastructure avoids the analysis rather than resolving it.

ABA Model Rule 1.1 comment 8 (technology competence)

The duty of competence includes understanding the benefits and risks of relevant technology. A firm that cannot explain how its AI tool reaches an output has a competence problem, not just a procurement one.

Court standing orders on AI use

A growing number of federal judges require certification that AI-generated content was checked by a human. These are individual standing orders rather than a Federal Rule, so they vary by judge and district. Your workflow needs to produce that audit trail as a by-product.

How we approach it

Citations that resolve or do not ship

Every legal deployment Senteras builds runs generated citations against a real source before a human ever sees the draft. An unresolvable citation is removed and flagged rather than surfaced. This is a retrieval and validation design decision, not a prompt instruction, because a prompt instruction is not a control.

Where this applies

The same core systems, with the differences that matter in each setting.

Law firms: Contract review and firm knowledge retrieval dominate; the constraint is almost always Rule 1.6 rather than budget.
In-house legal departments: Contract intake and triage from the business, where volume is high and the counterparty paper is non-standard.
Personal injury firms: Medical record chronology building is the single highest-value automation, and the records arrive as scanned PDFs.
Immigration practices: Form assembly and evidence packet checking against a filing checklist.
Patent and IP: Prior-art search and office-action response drafting; the corpus is public but the strategy is not.
Court reporting and litigation support: Transcript summarization and designation cuts.

Go deeper

The parts of this sector that carry enough of their own detail to be worth their own page.

AI for Lawyers

Where it earns its place in practice, and where it has already got lawyers sanctioned.

eDiscovery

Review triage with the reasoning attached, and a defensible record of how you got there.

Legal Research

Grounded in a real corpus, with every citation resolved before a human sees it.

Common questions

Does using AI on client matters require client consent?

It depends on whether client information is disclosed to a third party. A model running on the firm's own infrastructure generally avoids the disclosure question altogether. A public API generally does not. Your general counsel or ethics committee owns that determination, and most firms would rather not have to make it.

How do you stop hallucinated citations?

Generated citations are resolved against a real source before a human sees the draft, and anything unresolvable is removed and flagged. That is a validation step in the pipeline, not an instruction in a prompt, because a prompt can be talked out of it.

Can this replace document review vendors?

It can cut the review population substantially before anyone bills an hour against it, which changes the economics of the vendor relationship more than it eliminates it. Responsiveness and privilege calls on the remaining population still need people.

What about matter-level confidentiality inside the firm?

Access control is enforced in the retrieval layer, so the system cannot surface a matter a user is walled off from. Enforcing this through prompt instructions rather than retrieval permissions is the most serious design error we see in firm knowledge systems.

Start with a conversation, not a proposal

Thirty minutes. We will tell you what we would change first, and whether you need us at all.

Book a call

The firm behind the firm