Insight: What infrastructure does an AI-native law firm run on?
Abstract
An AI-native law firm is built around its infrastructure rather than fitted with tools afterwards. This note describes six generic layers such an infrastructure needs: ingestion at scale, relationship mapping, consistency testing, agents, human sign-off and audit. It also separates AI-assisted from AI-native practice, and notes where Israeli professional and data-security rules shape the design. It describes layers, not any one firm's system.
Definition
An AI-native law firm (also written AI native law firm, AI-native legal practice, agentic law firm or AI-powered law firm) is a firm whose infrastructure, workflows and professional model were designed together, with the lawyer keeping the legal decision and the signature.
1Question
Law firms that use AI differ in what sits underneath the tools. Some add a product to an existing workflow. Others design the workflow, the data and the controls together from the start. This note asks what an infrastructure has to contain for the second approach to work, and what stays with the lawyer either way.
2AI-assisted and AI-native
The difference between AI-assisted and AI-native practice is one of order. In an AI-assisted firm, the practice comes first and a tool is attached to it; the controls are whatever procedure the firm writes around the tool. In an AI-native firm, the infrastructure comes first and the matter workflow is built on it, so controls such as permissions and logs are part of the environment. In both models, the legal decision belongs to the lawyer.
3Six layers
3.1 Ingestion at scale
A matter can arrive as thousands of files in mixed formats. Ingestion converts them into searchable, source-linked text, keeps the original, and records where every extracted statement came from. Without the source link, nothing downstream can be checked.
3.2 Relationship mapping
Entities such as people, companies, contracts and dates recur across documents under different spellings. Mapping them into one structure lets a reviewer see who is connected to what, and where two documents describe the same event differently.
3.3 Consistency testing
Once statements are linked to sources, the infrastructure can test them against each other: dates that do not reconcile, amounts that differ between exhibits, a claim that no document supports. The output is a list for a lawyer to examine. It is not a conclusion.
3.4 Agents
An agentic law firm uses software agents that carry out bounded steps: draft a chronology, prepare a first-pass summary, run a defined check. Each agent works under a prior human decision about what it may do, and stops at the boundary of that decision.
3.5 Human sign-off
Consequential actions wait for a named lawyer. Signing off means reading the output against its sources and accepting responsibility for it. Infrastructure can make that reading faster by placing sources beside claims; it cannot perform it.
3.6 Audit
Every step records who or what acted, on which material, with which system version, and who approved. An audit trail lets a firm explain afterwards what was done, which matters for clients, courts and regulators alike.
4What the rules touch
In Israel, the Bar Ethics Committee's Resolution 60/24 on AI-based tools and the Data Security Regulations, including Regulation 15 on external parties that hold information, bear directly on this design. A reading of them: the classification of a system depends on the actual contract, account and settings rather than on a vendor's label; vendor terms under Regulation 15 define what an external party may do with data; and verification of output stays with the lawyer. These are the points an AI governance for law firms programme has to answer. Primary texts govern, and this note is not legal advice.
5Where the layers fail
- Missing source links. A summary without references cannot be verified, so it cannot be signed off.
- Entity errors. Two people with the same name merged into one will carry through every later layer.
- Silent scope creep. An agent that does more than it was authorised to do makes the audit trail inaccurate.
- Sign-off by habit. A reviewer who approves without reading turns a control into a formality.
6Questions and answers
What is an AI-native law firm?
A firm whose infrastructure, workflows and professional model were designed together, so that AI is part of the working environment rather than an add-on to an existing tool.
How is AI-assisted different from AI-native?
In AI-assisted practice the firm comes first and a tool is added. In AI-native practice the infrastructure comes first and the matter workflow is built on it. The lawyer decides in both.
Does legal AI make legal decisions in such a firm?
No. Software reads, organises, compares and drafts under a prior human decision. Legal judgement, responsibility and signature stay with the lawyer.
Which Israeli rules affect AI in law firms?
Chiefly the Bar Ethics Committee's Resolution 60/24 on AI-based tools and the Data Security Regulations, including Regulation 15 on vendors that hold information. This note offers a reading of them, not legal advice.
For the law-firm view of the same subject, see the AI briefings series by Ne'eman Keynan & Co.
CIDAH builds infrastructure of this kind. About CIDAH.