
01 / ARGENT FOR LEGAL
AI agents now act like staff
An AI agent is software that works for an employee: it reads, decides and acts using an AI model such as GPT, Claude or Gemini.
- The American Bar Association's ethics opinion on AI says a lawyer needs the client's informed consent before putting client information into a self-learning AI tool, and a boilerplate line in the engagement letter is "not sufficient."¹
- IBM puts the average breach in professional services, a group it says includes legal firms, at $5.08 million, up from $4.56 million a year earlier.²
You used to secure people who use software. Now you also secure software that acts like a person.
02 / ARGENT FOR LEGAL
Five ways it goes wrong
Your data-loss, web-proxy, identity and log tools were built for people and files, so they usually never see an agent's request.
- Client data leaves in a prompt. The night before a filing, settlement terms with a client's account number and matter number go to an outside model.
- Unapproved AI spreads. Across industries, IBM found incidents involving unapproved ("shadow") AI more than doubled, to 43% from 20%, averaging $5.39 million each.²
- Agents reach too far. A knowledge assistant answers a new partner from a client file he may not see. 92% of organizations with an AI-related breach lacked proper AI access controls.²
- Agents outlive their owners. The associate who set up a case-file agent leaves the firm; the agent keeps working, and no one answers for it.
- No proof. After an incident, the firm cannot say which agent sent what, or who owns it.
03 / ARGENT FOR LEGAL
The one idea: a checkpoint in front of every agent
Argent Strategic Intelligence sits between your AI agents and the models and tools they use, and decides each request before it goes out.
- It runs inside your own environment, detection included.
- It governs the agents registered with it. It connects to 17 AI providers and to 65 agent tools and frameworks, including Claude Code, Codex, Cursor and GitHub Copilot, plus any tool with a base-URL or proxy setting.
04 / ARGENT FOR LEGAL
How Argent answers each risk
| Risk | What Argent does |
|---|---|
| 1. Data leaves | Reads each request first and replaces Social Security, card and bank account numbers with <REDACTED> before the model sees them. An admin adds a pattern for the firm's own matter numbers, and the client list can be matched by fingerprint (only one-way fingerprints reach Argent, never the names), so those are caught too. The record names the kind and count, never the value. |
| 2. Unapproved AI | Only models the firm approves may answer; a model nobody approved is refused. A live list shows every agent with its owner. |
| 3. Reaching too far | Tools are approved one by one, and a new or changed tool waits for an admin's decision. A knowledge assistant's document lookups are filtered to what the person the agent works for may read (add-on); if the check fails, it gets nothing. The agent never holds the provider's key. |
| 4. Orphaned agents | Every agent has a named human owner. When the owner is deactivated, the agent is refused on its very next request, and the Org chart shows it until an admin reassigns it. |
| 5. No proof | Every decision goes on a signed record with each row linked to the one before. A client or an auditor checks an evidence package on their own machine and sees "TENANT PACKAGE OK"; change one file and the check fails. |
05 / ARGENT FOR LEGAL
What the bar, courts and states already expect
No bar, court or state is waiting for an AI law. These are the duties your program is judged against.
- ABA Formal Opinion 512 (July 2024), ethics guidance rather than a penalty: for a self-learning AI tool, "a client's informed consent is required prior to inputting information relating to the representation," and boilerplate engagement-letter consent "is not sufficient." It warns such a tool may reach people in the firm "prohibited from access to said information because of an ethical wall," and restates the duty of "reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to" client information. State bars apply their own rules.¹
- Federal Rule of Evidence 502(b): in a federal proceeding, an inadvertent disclosure does not waive privilege if the holder "took reasonable steps to prevent disclosure" and "promptly took reasonable steps to rectify the error."³
- Federal Rule of Civil Procedure 37(e): if electronically stored information that should have been preserved for litigation is lost because a party "failed to take reasonable steps to preserve it," and it cannot be restored or replaced, a court that finds prejudice may order measures "no greater than necessary to cure the prejudice"; on finding "intent to deprive," it may presume the information was unfavorable, tell the jury so, or dismiss the action or enter a default judgment.⁴
- New York General Business Law 899-aa, for example: a business that owns or licenses New York residents' private information must tell them "in the most expedient time possible and without unreasonable delay," within thirty days of discovering a breach; a knowing or reckless failure to notify carries a civil penalty of up to $250,000.⁵
06 / ARGENT FOR LEGAL
How a law firm starts
Watch first. Nothing changes for lawyers or staff until the firm decides.
- Watch. New business units start in watch mode: Argent records what it would have done and forwards the request unchanged.
- Enforce, area by area. Data-loss checks, tools, owners and others switch independently.
- Inside your walls. It runs in your Kubernetes cluster or as a single-node bundle, or as an air-gapped bundle verified before import. Redacted prompts go on to the model provider you chose.
Argent Strategic Intelligence