---
title: The Last Click Belongs to a Person
description: Learn how law firms can limit AI access to client information, email, CRM, websites and social platforms while preserving human judgment at consequential decision points.
image: https://lawgixadvisorygroup.com/hubfs/IMG_1049.png
---

[Skip to main content](https://lawgixadvisorygroup.com/insights/the-last-click-belongs-to-a-person#main-content)

[![Lawgix Advisory Group](https://lawgixadvisorygroup.com/hs-fs/hubfs/raw_assets/public/Lawgix-Advisory-Group-Theme-Staging/assets/images/lawgix-logo.png?width=200&height=74&name=lawgix-logo.png)](https://lawgixadvisorygroup.com/)

Open menu

[About Us](https://lawgixadvisorygroup.com/about-us) [How We Help](https://lawgixadvisorygroup.com/services)

Resources ⌄

[Lawgix Insights](https://lawgixadvisorygroup.com/insights)[Lawgix Advisory Podcast](https://lawgixadvisorygroup.com/lawgix-advisory-podcast)[In the News](https://lawgixadvisorygroup.com/in-the-news)

[Events](https://lawgixadvisorygroup.com/lawgix-events)

Open site search

Search website pages and Lawgix Insights

Submit search

[Contact Us](https://lawgixadvisorygroup.com/contactus)

[Home](https://lawgixadvisorygroup.com/)›[Lawgix Insights](https://lawgixadvisorygroup.com/insights)

Lawgix Insights

# The Last Click Belongs to a Person

By Lawgix Advisory Group · October 16, 2026 | [Technology](https://lawgixadvisorygroup.com/insights/tag/technology), [Firm Operations](https://lawgixadvisorygroup.com/insights/tag/firm-operations), [Artificial Intelligence](https://lawgixadvisorygroup.com/insights/tag/artificial-intelligence)

*Guardrails for AI access to client information and public communications*

A recent security incident at a large U.S. law firm began with social engineering directed at an attorney. The firm reported that a threat actor gained access to a limited number of documents. A proposed class action followed, alleging that the firm's training and safeguards were inadequate. Those allegations have not been decided, and the incident was not attributed to artificial intelligence.

That distinction matters. The incident is useful here because it exposes the weak point that AI can magnify: a trusted identity with more access than the immediate task requires. If an attorney's account can already reach email, documents, CRM records and client portals, connecting an AI assistant to that same identity may expand what can be searched, summarized, copied or acted upon in a few seconds.

Law firms should resist the convenient default of giving an AI tool the same reach as the person using it. The safer design is narrower. The system receives the minimum information and authority required for one defined workflow. A qualified person remains responsible for any decision that affects a client, changes an authoritative record or communicates outside the firm.

## Access Is an Operating Decision

An AI license may look like another software purchase, but its permissions determine its risk. A drafting assistant that works only with documents a lawyer deliberately uploads presents one risk profile. An agent that can search an entire document repository, update CRM records, send email and publish website content presents another.

The difference is agency. A retrieval tool can expose information. An action-enabled tool can also change records, contact people and publish statements under the firm's name. When those capabilities are combined, an incorrect instruction, compromised credential or malicious content hidden in a source document can travel farther than the user intended.

The American Bar Association's [Formal Opinion 512](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf) reminds lawyers that using generative AI does not displace their duties concerning competence, confidentiality, communication and supervision. Those duties should be translated into technical permissions and operating procedures. A policy that says “protect client information” will not prevent an integration from reading every matter available to the user who connected it.

## Protect Two Different Risk Zones

The first risk zone contains client and prospect information: matter files, email, intake notes, consultation recordings, billing records, CRM histories and internal knowledge. Access should be limited by matter, role, data type and purpose. A marketing workflow does not need open access to litigation files. An intake summarizer does not need the authority to export every contact in the CRM.

The second risk zone contains the firm's public voice: outbound email, website publishing, social accounts, client alerts, chatbot responses and automated follow-up. A bad draft can be corrected before anyone sees it. A bad post, message or bulk email can create a client-relations problem, an ethics issue or a misleading public record almost instantly.

These zones often meet inside the same workflow. An AI tool may use CRM data to personalize an email, draw from a matter summary to prepare a client update, or convert an internal analysis into a website article. The handoff from private information to external communication deserves its own control point.

## Build Guardrails Around the Action

Effective guardrails do more than tell users to be careful. They reduce what the system can reach, separate preparation from execution and make a named person approve the consequential step.

### 1. Inventory every connection and name an owner

Create a register of AI tools and integrations showing what each system can read, what it can change, whose identity it uses, where data goes, how long it is retained and who can revoke access. Include browser extensions, meeting assistants and individual user connections. If nobody owns a connection, it should not remain active.

### 2. Begin with read only and least privilege

Give the system access to the smallest useful dataset. Limit it to the relevant practice, matter, pipeline or content library. Use scoped service credentials when available instead of a partner's full account. Add write access only after the workflow has been tested and there is a business reason for it. [CISA's guidance for agentic AI](https://www.cisa.gov/news-events/alerts/2026/05/01/joint-guidance-securing-agentic-artificial-intelligence-systems) specifically warns against broad or unrestricted access to sensitive data and critical systems.

### 3. Separate drafting from sending and publishing

An AI system can prepare a client email without receiving permission to send it. It can draft a website post without access to the publish button. It can propose a CRM update without altering the record. Keep execution in a separate step controlled by the application or workflow, not by a sentence in the prompt.

### 4. Put human judgment at specific gates

Do not rely on a generic instruction that a human remains in the loop. Identify the person, the evidence that person must see and the exact action requiring approval. The reviewer should see the source material, the proposed output, the intended recipient or destination and any data the system plans to change. Approval should apply to one bounded action, not to an open-ended series of future actions.

### 5. Make review consequential

A steady stream of routine approval requests produces rubber stamping. Reserve mandatory approval for decisions that matter, then reduce low-value prompts through narrower permissions and better workflow rules. The reviewer must have the authority, subject knowledge and time to reject or revise the result. Log the approval and the final action.

### 6. Monitor behavior and preserve a fast stop

Record searches, exports, changes, messages and publication events. Alert an accountable owner to unusual volume, access outside the expected scope, repeated failures or new destinations. Maintain a practical way to revoke credentials, disconnect the integration and restore authoritative data. A response plan should assume that the normal email or document system may be unavailable or untrusted.

### 7. Test the workflow before the incident

Run a tabletop exercise around a plausible event: a lawyer approves a malicious connection, an intake transcript contains instructions that influence an AI agent, or a compromised account attempts to export client records and publish a message. Ask who sees the alert, who can stop the action, which clients may be affected and who approves external communications. Repair the gaps while the exercise is still hypothetical.

## Human Review Must Be Designed

Human approval works only when the person is reviewing a decision rather than clicking through a queue. Before deployment, define what the reviewer is responsible for checking. For client work, that may include factual accuracy, legal judgment, confidentiality, privilege, scope and the identity of the recipient. For public communications, it may include support for factual claims, tone, ethics rules, client consent, links, audience and timing.

The system should show its work in a form the reviewer can evaluate. A proposed client alert should display the approved source material and any client facts used to personalize it. A suggested CRM change should show the current record, proposed value and downstream automation that will fire if the change is accepted. A social post should remain a draft until an authorized person sees the exact platform-ready version.

| Workflow | AI may prepare | A person must decide |
| --- | --- | --- |
| **Matter knowledge** | Search authorized sources and summarize responsive material | Whether the sources are appropriate, the summary is accurate and the result may be used |
| **CRM** | Suggest fields, classifications, next steps or duplicate matches | Whether to change the authoritative record, merge data or trigger another workflow |
| **Client email** | Draft a message using approved facts and templates | Whether the content is accurate, appropriate for the recipient and ready to send |
| **Website and social** | Draft copy, metadata and platform variations | Whether the claim is supportable, ethically compliant and ready to publish under the firm's name |
| **Intake** | Summarize a consultation and flag missing information | Whether the summary is reliable, the prospect is qualified and a follow-up action should occur |

Good design leaves the final decision with the person who can understand the consequence. The AI can retrieve, organize, compare and draft. It should not quietly decide which client information leaves the firm or what the firm says in public.

## A Practical Starting Point

Choose one workflow that touches client information or an external channel. Map the data it uses, the systems it reaches and the action it proposes to take. Remove any access the task does not require. Place a human decision before every external communication, authoritative data change, bulk export or expansion of permissions. Then test the control with realistic mistakes and a compromised-user scenario.

Lawgix Advisory Group helps law firms map workflows, define AI access, establish decision rights and build human review into practical adoption plans. The objective is controlled usefulness: enough access to improve the work, with firm leadership and qualified professionals retaining authority over client information and the firm's public voice.

[**Explore Lawgix Advisory Group consultation options**](https://lawgixadvisorygroup.com/consultations)

## Source Notes

1. [Reuters, October 8, 2026](https://www.reuters.com/legal/government/sheppard-mullin-faces-lawsuit-over-data-breach-2026-10-08/). Reporting on the recent law-firm incident and proposed class action. The article does not attribute the incident to AI.
2. [American Bar Association Formal Opinion 512](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf), July 29, 2024. Addresses competence, confidentiality, communication, supervision, candor and reasonable fees when lawyers use generative AI.
3. [CISA and partner guidance on adopting agentic AI services](https://www.cisa.gov/news-events/alerts/2026/05/01/joint-guidance-securing-agentic-artificial-intelligence-systems), May 1, 2026. Recommends limiting agent autonomy, applying strong identity management and maintaining robust oversight.
4. [NIST AI Risk Management Framework Core](https://airc.nist.gov/airmf-resources/playbook/). Calls for defined human oversight, documented roles, controlled third-party components and ongoing monitoring.

Related Service

## Law Firm Advisory Services

Explore the strategy, operations, growth, and leadership support Lawgix provides to law firms and professional service firms.

[Explore This Service](https://lawgixadvisorygroup.com/services)

More on Technology

## Related Insights

### [AI Adoption Is Easy to Count. Can Your Law Firm Show Clients What Improved?](https://lawgixadvisorygroup.com/insights/ai-adoption-is-easy-to-count.-can-your-law-firm-show-clients-what-improved)

October 7, 2026

### [The AI-Native Law Firm Is an Operating Model, Not a Technology Stack](https://lawgixadvisorygroup.com/insights/the-ai-native-law-firm-is-an-operating-model-not-a-technology-stack)

September 27, 2026

[View This Topic →](https://lawgixadvisorygroup.com/insights/tag/technology)

[View All Lawgix Insights →](https://lawgixadvisorygroup.com/insights)

Ideas Worth Putting Into Practice

## Receive Lawgix Insights and Firm Updates

[![Lawgix Advisory Group](https://lawgixadvisorygroup.com/hs-fs/hubfs/raw_assets/public/Lawgix-Advisory-Group-Theme-Staging/assets/images/lawgix-logo.png?width=200&height=74&name=lawgix-logo.png)](https://lawgixadvisorygroup.com/)

© 2026 Lawgix Advisory Group LLC.  
All Rights Reserved.

Quick Links

[About Us](https://lawgixadvisorygroup.com/about-us)[How We Help](https://lawgixadvisorygroup.com/services)[Contact Us](https://lawgixadvisorygroup.com/contactus)

[Lawgix Insights](https://lawgixadvisorygroup.com/insights)[Lawgix Advisory Podcast](https://lawgixadvisorygroup.com/lawgix-advisory-podcast)[In the News](https://lawgixadvisorygroup.com/in-the-news)

Stay Connected

## Ideas Worth Putting Into Practice

Receive Lawgix Insights and firm updates.

[Privacy Policy](https://lawgixadvisorygroup.com/privacy)•[Disclaimer & Terms of Use](https://lawgixadvisorygroup.com/terms)

```json
{
  "@context" : "https://schema.org",
  "@id" : "https://lawgixadvisorygroup.com/#organization",
  "@type" : "ProfessionalService",
  "areaServed" : "United States",
  "description" : "Strategic advisory and management consulting for law firms and professional service firms.",
  "founder" : {
    "@id" : "https://lawgixadvisorygroup.com/about-us#jim-jarrell",
    "@type" : "Person",
    "jobTitle" : "Founder and CEO",
    "name" : "Jim Jarrell"
  },
  "logo" : "//47111792.fs1.hubspotusercontent-na1.net/hubfs/47111792/raw_assets/public/Lawgix-Advisory-Group-Theme-Staging/assets/images/lawgix-logo.png",
  "name" : "Lawgix Advisory Group",
  "url" : "https://lawgixadvisorygroup.com/"
}
```

```json
{
  "@context" : "https://schema.org",
  "@id" : "https://lawgixadvisorygroup.com/insights/the-last-click-belongs-to-a-person#article",
  "@type" : "BlogPosting",
  "author" : {
    "@id" : "https://lawgixadvisorygroup.com/#organization",
    "@type" : "Organization",
    "name" : "Lawgix Advisory Group",
    "url" : "https://lawgixadvisorygroup.com/"
  },
  "dateModified" : "2026-10-16T16:17:14Z",
  "datePublished" : "2026-10-16T15:56:55Z",
  "description" : "Learn how law firms can limit AI access to client information, email, CRM, websites and social platforms while preserving human judgment at consequential decision points.",
  "headline" : "The Last Click Belongs to a Person",
  "image" : "https://47111792.fs1.hubspotusercontent-na1.net/hubfs/47111792/IMG_1049.png",
  "mainEntityOfPage" : {
    "@id" : "https://lawgixadvisorygroup.com/insights/the-last-click-belongs-to-a-person",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@id" : "https://lawgixadvisorygroup.com/#organization"
  },
  "url" : "https://lawgixadvisorygroup.com/insights/the-last-click-belongs-to-a-person"
}
```

```json
{
  "@context" : "https://schema.org",
  "@type" : "BreadcrumbList",
  "itemListElement" : [ {
    "@type" : "ListItem",
    "item" : "https://lawgixadvisorygroup.com/",
    "name" : "Home",
    "position" : 1
  }, {
    "@type" : "ListItem",
    "item" : "https://lawgixadvisorygroup.com/insights",
    "name" : "Lawgix Insights",
    "position" : 2
  }, {
    "@type" : "ListItem",
    "item" : "https://lawgixadvisorygroup.com/insights/the-last-click-belongs-to-a-person",
    "name" : "The Last Click Belongs to a Person",
    "position" : 3
  } ]
}
```

```json
{
  "@context" : "https://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Lawgix Advisory Group",
    "url" : "https://lawgixadvisorygroup.com/insights/author/lawgix-advisory-group"
  },
  "dateModified" : "2026-10-16T16:17:14.370Z",
  "datePublished" : "2026-10-16T15:56:55.000Z",
  "headline" : "The Last Click Belongs to a Person",
  "image" : [ "https://lawgixadvisorygroup.com/hubfs/IMG_1049.png" ],
  "mainEntityOfPage" : {
    "@id" : "https://lawgixadvisorygroup.com/insights/the-last-click-belongs-to-a-person",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "url" : "https://lawgixadvisorygroup.com/hubfs/Lawgix%20Logo%20-%20300dpi.png"
    }
  }
}
```