# Agentic AI Governance Gap for Law Firms

> Assistive AI drafts; agentic AI acts. When AI files, sends, and executes on its own, the human checkpoint your firm's policy assumes can quietly disappear.

[Home](https://rankshieldlegal.com/) / [Blog](https://rankshieldlegal.com/blog/) / Legal AI Agentic AI
# When Your AI Stops Drafting and Starts Acting: The Agentic AI Governance Gap in Law Firms
Most law firm AI policies were written for AI that drafts: a human reviews the output before anything happens. Agentic AI acts. It files, sends, schedules, and chains steps across a firm's systems, often with less human review at each step. That shift collapses the human checkpoint most policies quietly assume, and it is why governance now has to reach the level of individual actions, not just generated text.

By [Jamie Kloncz](https://rankshieldlegal.com/about/), Founder, RankShield ** 22 min read ** Published July 16, 2026

Most law firm AI policies were written for AI that drafts: a human reviews the output before anything happens. Agentic AI acts. It files a document, sends an email, schedules an event, or runs a multi-step workflow across a firm's systems, often with less human review at each step. That shift collapses the human checkpoint most current policies quietly assume, and it is why governance now has to reach the level of individual actions rather than just the text a model produces.
The distinction sounds academic until you trace what it changes in practice. A policy that says "review AI output before you file" makes sense when a person stands between the model and the docket. It says nothing about a system that moves a document or sends a message on its own. As legal technology vendors add agentic capabilities, and as an emerging integration layer lets AI tools connect to a firm's document management, calendar, and email systems, the gap between what firms have adopted and what they have governed widens.
This article is written from a security and engineering perspective, not a legal one. It is informational and not legal advice. The aim is to describe the governance gap agentic AI opens, explain why the professional duties firms already recognize still apply, and lay out the action-level controls that close it: agent identity, per-action audit trails, approved-tool gating, human-in-the-loop for consequential actions, and independently verifiable attestation of what an agent did.

## The shift from AI that drafts to AI that acts
Assistive AI drafts and waits for a human to act. Agentic AI acts, chaining steps across a firm's systems. The governance question changes from reviewing text to accounting for actions.
The clearest way to understand agentic AI is by contrast with the AI most firms already use. Assistive AI generates text: a research summary, a first-draft clause, a proposed email. A human reads that output and decides whether to use it, edit it, or discard it. Nothing leaves the attorney's control until the attorney chooses to act on it. The model is a drafting tool, and the lawyer is the checkpoint between the draft and any consequence.
Agentic AI is defined by action rather than output. An agent does not only suggest an email; it can send one. It does not only draft a calendar entry; it can create it. It chains steps together, calling tools and moving between systems to complete a task it was given at a higher level. The instruction shifts from "write me a draft" to "handle this," and the system carries out the intermediate steps on its own. That is a different kind of tool, and it asks a different question of a firm's governance.
The reason this matters now is connectivity. Industry coverage describes Model Context Protocol, an open standard originated by Anthropic and backed by a growing set of vendors, as an emerging integration layer that lets AI tools connect to a firm's systems and act, described by one commentator as the equivalent of a common protocol for AI-to-system integration [[1]](#ref-1) [[2]](#ref-2). As legal research, drafting, and document-management platforms add agentic features on top of that kind of connectivity, the practical capability moves from generating text to executing work.

## What agentic AI means inside a law firm's systems
Inside a firm, agentic capability is not abstract. It means an AI system connected to the document management system can move, rename, or share a file. Connected to email, it can draft and send. Connected to a calendar, it can schedule and invite. Connected to a research or drafting platform, it can assemble a work product and route it onward. Each connection is a place where the system can take an action that has a consequence outside the model.
The value proposition is real, which is why adoption is moving. A commentator writing on legal technology noted that the limiting factor for many firms is not model capability but connectivity, because the AI sits in one tool while the documents sit in the document management system, forcing lawyers to move context by hand [[1]](#ref-1). Agentic workflows built on a shared integration layer are meant to remove that friction by letting the system reach across the stack. The same connectivity that removes friction is also what lets an action happen without a person performing it.
That is the engineering reality a governance discussion has to start from. Firms are not choosing between text generation and autonomous action as distant options. They are adopting tools where the two blur together, where a single request can produce both a draft and the act of sending it. The governance task is not to reject that capability but to make each action it produces attributable and reviewable.

- **Document systems:** an agent can move, share, or modify files, not just summarize them.
- **Email:** an agent can compose and send messages, not just propose language.
- **Calendar:** an agent can create events and send invitations on its own.
- **Workflows:** an agent can chain several of these steps to complete a higher-level task with less review per step.

Source: ABA Formal Op. 512 (2024); ABA Model Rules Download SVG

## Why the human checkpoint your policy assumes is disappearing
Current policies assume a person acts on AI output. Agentic systems can act on their own, so the review step the policy relies on may never happen. That silent gap is the governance problem.
Almost every current law firm AI policy contains a version of the same rule: verify AI output before you rely on it, and check citations before you file. That rule is sound, and it addresses a well-documented failure mode. It also carries a hidden assumption, which is that a human being stands between the model and any consequence. The rule governs a review step that the attorney performs before acting. It presumes the attorney is the one who acts.
Agentic AI removes that presumption. When a system can send the email or file the document itself, the review step the policy describes may never occur, because there is no longer a natural pause where a person inspects the output before it takes effect. The checkpoint the policy relies on was implicit in how assistive tools worked, not something the policy explicitly required at the level of each action. Take away the human who acts, and the policy is silent on the moment that now matters most.
This is the governance gap in a sentence. A policy written for "review before you file" does not cover an agent that files. It is not that the existing rules are wrong; it is that they stop short of the new surface. The failure is quiet precisely because the policy still reads as reasonable. It simply no longer maps to a workflow in which consequential actions can happen between, or after, the points where a human is looking.
The point is not that verification rules are wrong, but that they were written for a workflow where a human performed the consequential action.

## How agentic AI differs from assistive AI for governance
Laying the two side by side makes the governance shift concrete. The table below contrasts assistive, drafting AI with agentic, acting AI across the dimensions that matter for oversight. The left column is the world most firm policies were written for; the right column is the world firms are moving into. The change is not the model's intelligence but where the consequence sits and who, or what, produces it.
The comparison is a general illustration, not a compliance standard. It is meant to show why controls designed for reviewing text do not fully transfer to accounting for actions, and where a policy needs new language.
Dimension Assistive AI (drafts) Agentic AI (acts)
Primary output Text a human reads Actions taken across systems
Human checkpoint Before the human acts on output May be absent unless designed in
Unit to govern The generated document Each individual action
Record needed That output was reviewed What the agent did and what data it touched
Access required Read the prompt, return text Reach into email, files, calendar, tools
Failure surface A bad draft a human catches An action taken before anyone looks

Illustrative comparison for governance planning, not a legal or compliance standard. Adapt with your own counsel.

## The professional-responsibility duties that still apply
Nothing about agentic AI creates a gap in professional responsibility. The existing duties still apply, and they extend to how a firm supervises the tools it uses. Competence under Model Rule 1.1, confidentiality under Model Rule 1.6, and the supervisory duties under Model Rules 5.1 and 5.3 covering the conduct of lawyers and nonlawyer assistants reach the AI systems a firm deploys, including ones that act. A lawyer who directs an agent to handle a task remains responsible for the result.
The anchor guidance remains ABA Formal Opinion 512, issued in July 2024, which confirms that generative AI use implicates competence, confidentiality with informed consent, communication, and supervision [[3]](#ref-3). That opinion addresses generative AI generally; as of this writing there is no dedicated ABA opinion specific to agentic AI. The absence of agentic-specific guidance does not mean the duties pause. It means firms apply the duties they already have to a capability the guidance did not separately anticipate.
Regulators are beginning to look at the category directly. The State Bar of California has published proposed amendments to its Rules of Professional Conduct related to artificial intelligence for public comment as part of its 2026 process [[4]](#ref-4). That work is proposed and in progress, not adopted, and it should be read as a signal of direction rather than a settled rule. No bar has adopted a rule specific to agentic AI at this time. The responsible posture is to govern to the duties that already exist and to watch the rulemaking that is underway.
Duties described here are drawn from the ABA Model Rules and ABA Formal Opinion 512 [[3]](#ref-3). This is informational, not legal advice, and rules vary by jurisdiction.

## Why current AI policies do not cover autonomous actions
A policy can satisfy the supervision duty for AI that drafts while leaving it unmet for AI that acts, because the actions an agent takes are conduct the policy never described. The gap is structural, not a matter of diligence.
The supervisory duties give the analysis a sharp edge. Model Rules 5.1 and 5.3 concern supervising the conduct of people whose work the lawyer is responsible for. Applied to AI, the sensible reading is that a lawyer is responsible for supervising the tools that produce work in the lawyer's name. Assistive tools make that supervision straightforward, because the lawyer reviews the output before it becomes conduct. Agentic tools make it harder, because the tool itself can produce conduct.
A policy that discharges the supervision duty for assistive AI can leave it unmet for agentic AI without anyone noticing. The policy says review the output; the associate reviews the output; the duty appears satisfied. But if the same tool also sent three emails and moved a document while completing the task, those actions were conduct the policy never described how to supervise. The letter of the policy was followed. The actions it did not contemplate went ungoverned.
This is why the fix is not simply a stricter version of the old rule. Telling attorneys to be more careful does not create oversight of actions the policy does not name. The gap is structural: the unit the policy governs is the document, and the unit that now needs governing is the action. Closing it requires the policy to reach down to the level of individual actions and to require a record of them, which is a different kind of control than reviewing a draft.

## Action-level accountability: the controls that close the gap
If the governed unit shifts from the document to the action, the controls have to shift with it. Action-level accountability means the firm can answer, for any consequential action an agent took, four questions: which agent took it, what it was permitted to do, what it actually did and what data it touched, and whether a human approved it where approval was required. A firm that can answer those questions has governance that reaches the new surface. One that cannot is relying on a checkpoint that agentic tools may have removed.
The controls below are the practical mechanisms for producing those answers. They are drawn from ordinary security engineering rather than anything exotic: identity, access scoping, logging, and human approval gates are how any system that takes actions on your behalf is normally governed. Applied to legal AI, they turn a broad instruction to supervise into specific, checkable practices. None of them is a substitute for lawyer judgment, and none discharges a professional duty on its own.

- **Agent identity:** each agent has a distinct, attributable identity, so an action can be tied to the specific agent that took it rather than to a shared, anonymous tool.
- **Access scoping:** an agent can reach only the systems and data its task requires, not the whole firm, limiting the blast radius of any single action.
- **Approved-tool gating:** an agent can invoke only the tools and actions the firm has cleared, so an unvetted capability is the exception that stands out rather than a silent default.
- **Per-action audit trails:** each consequential action is logged, with what the agent did and what data it touched, so the record exists before anyone needs it.
- **Human-in-the-loop:** consequential actions require human approval before they take effect, restoring the checkpoint the old policy assumed.
- **Attestation:** an independently verifiable record of what an agent did, so the firm can demonstrate the controls operated rather than assert that they did.
4 questions action-level accountability must answer: which agent, what permitted, what done, who approved

## Agent identity, access scoping, and approved-tool gating
The first three controls are about constraining what an agent can do before it does anything. Agent identity is the foundation: if actions cannot be attributed to a specific agent, none of the later controls have anything to attach to. Giving each agent a distinct identity is the difference between a log that says an action occurred and a record that says which agent, operating under which task, took it. Attribution is the precondition for accountability.
Access scoping and approved-tool gating then bound the agent's reach. Scoping limits which systems and which data an agent can touch, so an agent assigned to summarize a matter cannot also reach unrelated client files. Gating limits which actions and tools the agent may invoke, so sending email or moving documents is a capability the firm grants deliberately rather than one that arrives bundled with a platform. Together they shrink the surface where an unreviewed action can happen and make the permitted set explicit.
These controls also map cleanly onto the supervision duty. A firm that has scoped an agent's access and gated its tools has made concrete decisions about what the tool is allowed to do in the lawyer's name, which is closer to what supervising conduct actually requires than a general instruction to use AI carefully. RankShield is building toward this kind of control through an AI-tool attestation gateway that is in development; it is a security vendor, not a law firm, and adopting a tool does not by itself satisfy any professional duty. You can read more about that direction in [AI-tool attestation](https://rankshieldlegal.com/ai-tool-attestation/).

## Per-action audit trails and human-in-the-loop review
Log each consequential action as it happens, and require human approval before the ones that carry real consequence take effect. Records plus a restored checkpoint are what convert supervision from an instruction into a practice.
Constraining what an agent can do is only half of accountability. The other half is recording what it actually did, and requiring a person to approve the actions that carry real consequence. Per-action audit trails capture, for each consequential action, what the agent did and what data it touched, at the moment it happened. The reason to log at the action level, rather than the session level, is that the action is the unit that has a consequence, and it is the unit a client or a court would ask about later.
Human-in-the-loop review restores the checkpoint that assistive workflows contained by default. Instead of relying on a person happening to look, the control requires approval before a consequential action takes effect: before an email is sent, before a filing is submitted, before a document is shared outside the firm. Not every action needs a gate; the design task is deciding which actions are consequential enough to require one, and building the pause into the workflow rather than hoping it occurs. That decision is itself a supervisory judgment the firm should make deliberately.
A verifiable record then ties the two together. It is not enough to log actions in a system only the vendor controls; the value for a firm is being able to demonstrate, to someone examining the matter closely, that the controls operated as claimed. RankShield's roadmap includes a governance dashboard, in development, intended to give firms a single view of what agents did and which approvals occurred. It is a work in progress, not a shipped product, and it is described here so the direction is clear rather than as a current capability. The broader case for records you can independently check is in [why verifiable matters](https://rankshieldlegal.com/why-verifiable/) and in the emerging discipline behind a [governance dashboard](https://rankshieldlegal.com/ai-governance-dashboard/).

## Attestation, privilege isolation, and honest limits
The strongest form of action-level accountability is attestation: an independently verifiable record that a given action happened, under a given agent, touching given data, so the firm can prove the controls operated instead of asserting it. Attestation is what lets a record stand up when a client, an opposing party, or a court asks whether the firm's AI governance was real. Its value is evidentiary. It does not make the underlying action correct; it makes the action accountable.
Precision about what attestation proves is essential, because overstating it would defeat the purpose. Attesting that privileged material was isolated proves an architectural fact and a consent fact: that the material was kept separated as configured and that consent was captured. It does not decide the legal question of whether privilege attaches or survives, which remains a matter of law and lawyer judgment, and it does not by itself prevent a waiver. The claim the evidence supports is isolation and consent, not preserved privilege. That boundary is what keeps the record trustworthy. The reasoning is developed further in [privilege isolation](https://rankshieldlegal.com/privilege-isolation/).
Two more honest limits belong here. First, no verification or attestation makes a model reliable in every respect; nothing described in this article should be read as a claim that agentic AI is hallucination-free, because it is not. Attestation narrows a specific risk and creates a record that the narrowing occurred. Second, the cryptography behind a verifiable record is designed to be quantum-safe, not quantum-proof; those are different claims, and the honest one is the former. RankShield's attestation gateway and governance dashboard are both roadmap items in active development, not shipped products, and they are named here to describe a direction, not to advertise a capability the firm can rely on today.
Attestation evidences architecture and consent, not the legal status of privilege. It is informational context, not legal advice.

## Where a firm can start before the tools mature
A firm does not have to wait for a finished product to close the most obvious part of the gap. The first move is to read the existing AI policy specifically for the action-level silence: find every place the policy assumes a human acts on output, and ask what the policy says if a tool acts instead. That single pass usually reveals that the policy governs documents and says nothing about sent emails, moved files, or scheduled events. Naming the gap is what makes it addressable.
From there, the sequence below is a general illustration of how to bring agentic use under governance. It is not legal advice, and the specifics should be adapted with counsel to the firm's jurisdiction, tools, and practice. The goal at each step is the same as the rest of this article: not only to do the sensible thing, but to be able to show it was done. A related, broader treatment of policy structure is in the [law firm AI policy](https://rankshieldlegal.com/law-firm-ai-policy/) framework.

- **Inventory where agents can act** List the systems your AI tools connect to and the actions they can take, so the firm knows where autonomous action is possible before governing it.
- **Decide which actions are consequential** Mark the actions that need a human approval gate, such as sending external email, filing, or sharing documents outside the firm, and treat the rest as lower risk.
- **Scope access and gate tools** Limit each agent to the systems, data, and actions its task requires, so unvetted capabilities are deliberate exceptions rather than defaults.
- **Require per-action logging** Insist that consequential actions are recorded with what the agent did and what data it touched, so the record exists before anyone needs it.
- **Rewrite the policy to name actions** Update the AI policy so it governs the action, not only the document, and specify approval and record-keeping for consequential actions.
- **Have counsel review for your jurisdiction** Because duties and rules vary and rulemaking is in progress, have qualified counsel review the approach before the firm relies on it.
This sequence is informational and not legal advice; adapt it with your own counsel.

Test yourself
## Test yourself: governing AI that acts
Four questions on the governance gap agentic AI opens and how to close it.

- 1 What distinguishes agentic AI from assistive AI? It writes longer drafts It takes actions across systems, not just generating text It is only used by large firms **Answer:** It takes actions across systems, not just generating text Assistive AI drafts text a human reviews; agentic AI acts by sending, filing, or scheduling, which is why governance has to reach individual actions.
- 2 Do professional-responsibility duties apply to agentic AI even without a bar opinion specific to it? No, duties pause until a specific opinion issues Yes, existing duties under Model Rules 1.1, 1.6, 5.1, and 5.3 still apply Only in California **Answer:** Yes, existing duties under Model Rules 1.1, 1.6, 5.1, and 5.3 still apply ABA Formal Opinion 512 confirms the duties apply to generative AI, and no bar has adopted an agentic-specific rule, so firms govern to the duties that already exist.
- 3 What does attestation of privilege isolation actually prove? That a court will find privilege preserved That material was isolated as configured and consent was captured That the AI is hallucination-free **Answer:** That material was isolated as configured and consent was captured Attestation evidences architecture and consent, not the legal status of privilege, which only a court decides. No tool makes AI hallucination-free.
- 4 Which is a control that closes the gap? Telling attorneys to be more careful Per-action audit trails and human-in-the-loop approval Turning off all AI **Answer:** Per-action audit trails and human-in-the-loop approval The fix is structural: log each consequential action and require human approval for consequential ones, rather than a stricter version of the old rule.
Honest self-check. There is no sign-up, and nothing is stored.

Questions answered
## Straight answers to the common questions
The questions readers ask about this topic, answered directly. **No forms, no sales pitch.**

JAMIE KLONCZ · SEO AGENCY NAPLES ************** ONLINE
Pick a question on the left, or search above. You will get the direct answer, the way an answer engine would give it.

← PREV NEXT → [REQUEST ACCESS →](https://rankshieldlegal.com/contact/)

- **What is agentic AI, and how is it different from the AI my firm already uses?** Agentic AI is AI that takes actions and chains steps to complete a task, rather than only generating text a person reads. The assistive AI most firms use drafts a summary, a clause, or a proposed email, and an attorney decides whether to act on it. An agentic system can act on its own: send the email, move the document, schedule the event, or run a multi-step workflow across the firm's systems. The practical difference for governance is where the consequence sits. With assistive AI, a human stands between the model and any action. With agentic AI, the tool itself can produce the consequence, which is why oversight has to reach the level of individual actions rather than just reviewing generated text.
- **Why do our current AI policies not cover agentic AI?** Most law firm AI policies contain a rule like verify AI output before you rely on it, or check citations before you file. That rule assumes a human stands between the model and any consequence, performing a review before acting. Agentic AI removes that assumption, because the system can send an email or file a document itself, so the review step the policy describes may never happen. The policy still reads as reasonable, which is why the gap is easy to miss. It governs the document a human reviews and is silent on the actions a tool takes on its own. Closing the gap requires the policy to name individual actions, specify which need human approval, and require a record that the controls operated.
- **Do professional-responsibility rules apply to agentic AI even without a specific bar opinion?** Yes. The existing duties apply regardless of whether a bar has issued agentic-specific guidance. Competence under Model Rule 1.1, confidentiality under Model Rule 1.6, and the supervisory duties under Model Rules 5.1 and 5.3 extend to the AI tools a firm uses, including ones that act. ABA Formal Opinion 512, from July 2024, confirms these duties apply to generative AI, though it addresses generative AI generally rather than agentic systems specifically [3]. The State Bar of California has proposed amendments related to artificial intelligence out for public comment in its 2026 process, which is in progress and not adopted [4]. No bar has adopted a rule specific to agentic AI at this time, so the responsible approach is to govern to the duties that already exist. This is informational, not legal advice.
- **What is Model Context Protocol, and why does it matter for law firm AI?** Model Context Protocol, or MCP, is an open standard originated by Anthropic and backed by a growing set of vendors that lets AI tools connect to a firm's systems in a common way [1] [2]. Industry coverage describes it as an integration layer, likened to a common protocol for AI-to-system connection [1]. It matters because connectivity, not raw model capability, is often the limiting factor for firms: the AI sits in one tool while documents sit in the document management system. As legal platforms add agentic features on top of that kind of connectivity, the same integration that removes friction also lets a system take actions across email, files, and calendars. That capability is the reason governance now needs to account for actions, not only text.
- **What controls actually close the agentic AI governance gap?** Action-level accountability, meaning the firm can answer four questions for any consequential action an agent took: which agent took it, what it was permitted to do, what it did and what data it touched, and whether a human approved it. The mechanisms are ordinary security practices applied to legal AI: agent identity so actions are attributable, access scoping so an agent reaches only what its task requires, approved-tool gating so it can invoke only cleared actions, per-action audit trails so a record exists, human-in-the-loop approval for consequential actions, and attestation so the firm can demonstrate the controls operated rather than assert it. None of these substitutes for lawyer judgment or discharges a professional duty on its own; they make supervision concrete and demonstrable.
- **What can RankShield do about this today, and what is still in development?** RankShield is building toward action-level accountability for legal AI, and it is honest about what is shipped versus planned. The AI-tool attestation gateway and the governance dashboard described in this article are roadmap items in active development, not products a firm can rely on today. RankShield is a security and engineering vendor, not a law firm, and it does not provide legal advice; adopting any tool does not by itself satisfy a professional duty. It is also precise about its claims: attestation evidences that privileged material was isolated as configured and that consent was captured, not that privilege is legally preserved, and its cryptography is designed to be quantum-safe, not quantum-proof. No tool makes AI hallucination-free. What the approach adds is verifiable evidence that governance controls operated.
- **How should a firm start before agentic governance tools are mature?** Begin by reading your existing AI policy for its action-level silence: find every place it assumes a human acts on output, and ask what it says if a tool acts instead. That pass usually shows the policy governs documents and says nothing about sent emails, moved files, or scheduled events. Then inventory where your AI tools can take actions, decide which of those actions are consequential enough to require human approval, scope each agent's access and gate its tools to what a task requires, and require per-action logging. Finally, rewrite the policy so it governs the action rather than only the document, and have counsel review the approach for your jurisdiction, since duties vary and rulemaking is in progress. This is informational, not legal advice.

## References

- Artificial Lawyer (Liam Reid, Legatics). MCP: The Standard that Decides Legal AI's Future. June 2, 2026. https://www. artificiallawyer. com/2026/06/02/mcp-the-standard-that-decides-legal-ais-future/. 2026-06-02.
- Anthropic. Introducing the Model Context Protocol. November 25, 2024. https://www. anthropic. com/news/model-context-protocol. 2024-11-25.
- ABA Standing Committee on Ethics & Prof'l Responsibility. Formal Opinion 512: Generative Artificial Intelligence Tools. July 29, 2024. https://www. americanbar. org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/. 2024-07-29.
- State Bar of California. Proposed Amendments to the Rules of Professional Conduct Related to Artificial Intelligence (2026 public comment; proposed, not adopted). https://www. calbar. ca. gov/. 2026-01-01.

Written by
## [Jamie Kloncz](https://rankshieldlegal.com/about/)
Founder, RankShield
Jamie Kloncz is the founder of RankShield, the verifiable AI and quantum security platform behind RankShield Legal. An engineer by training, he built RankShield after his own devices and business were attacked, including an AI voice-cloning scam that targeted his family, on one conviction: unverifiable security is the real danger, so every consequential action should leave a receipt anyone can independently check.
[More about Jamie →](https://rankshieldlegal.com/about/)

Try it · Free
## Check a citation against live case-law
Paste a citation from an AI-drafted brief and see whether the case actually exists, resolved against live case-law. Free, no sign-up. Then request early access to certify a full filing.
[Try the citation checker](https://rankshieldlegal.com/ai-legal-citation-checker/)

Keep reading
## Related guides
[Legal AI California Is Turning AI Ethics Guidance Into Disciplinary Rules Read guide →](https://rankshieldlegal.com/blog/california-ai-disciplinary-rules/)[Legal AI What Court AI-Certification Orders Require You to Attest Read guide →](https://rankshieldlegal.com/blog/court-ai-certification-orders/)[Legal AI Building a Defensible Law Firm AI Policy Read guide →](https://rankshieldlegal.com/blog/law-firm-ai-policy/)