# Shadow AI in Law Firms: Finding Unapproved Upload Tools

- Author: Schema Law (https://www.schemalaw.com/)
- Published: 2022-08-28T03:11:43Z
- Updated: 2026-10-11T19:22:20Z
- Category: Confidentiality
- Canonical: https://www.schemalaw.com/blog/shadow-ai-in-law-firms

> Shadow AI in law firms is the chatbots and converters staff use without approval. How to find the tools, run an amnesty and move people to approved ones.

Shadow AI in law firms is the set of chatbots, converters and browser add-ons that lawyers and staff use on firm work without anyone having approved them. You find the tools by asking people without blame, checking expense claims, sign-in records and browser extensions, and then you sort each tool into approve, restrict or stop.

This post is about discovery and clean-up. It assumes the firm either has a written policy or is about to write one, and it leaves the drafting of that policy to another post. Nothing here is legal advice; supervision rules vary by state.

## What is shadow AI in a law firm?

Shadow AI is a label, commonly used by analogy to "shadow IT", for AI tools that people use at work outside the organization's approval. In a law firm the label fits any tool that receives client information and that no responsible lawyer has vetted. It is not a term defined in the ethics rules.

The tool does not need to call itself AI to belong on the list. What matters for a law firm is whether client information leaves the firm's control and lands with a company whose terms nobody at the firm has read. On that test, the category includes:

- a chatbot used on a personal account to summarize or redraft a client document;
- a free website that converts a scanned PDF to Word or pulls a table into a spreadsheet;
- a browser extension that reads the page you are on in order to rewrite, translate or summarize it;
- a meeting recorder or transcription app that joins client calls;
- a phone scanning app that uploads its images to an online account;
- a translation site that receives pasted contract clauses;
- an AI feature added to software the firm already licenses.

The last item deserves a second look. A tool the firm approved two years ago for one purpose can gain a new feature that sends content somewhere new. Approval of the product then is not approval of the feature now.

Two things are outside the label. A tool the firm has vetted and listed is not shadow AI, even if it is a chatbot. And a lawyer asking a chatbot a general question with no client facts in it raises accuracy concerns, not the upload concern this post is about.

## Why do lawyers and staff reach for unapproved converters and chatbots?

No survey stands behind this post, so treat each cause as a hypothesis to test in your own firm. The candidates: the task was real, the approved route was missing, slow or unknown, and nobody had said the tool was off limits. Ask your people which applies, in a way that is safe to answer.

Microsoft's guidance for administrators hunting unapproved cloud apps makes the same point about apps in general. Its [tutorial on discovering and managing shadow IT](https://learn.microsoft.com/en-us/defender-cloud-apps/tutorial-shadow-it) advises looking at discovered apps by category, "because you'll often find that non-sanctioned apps are being used for legitimate work-related purposes that weren't addressed by a sanctioned app." That is one vendor's observation about its customers, not a finding about law firms, but it suggests where to look first: at the task, before the person.

Questions that separate the causes:

| Possible cause | Question to ask the user | What the answer points to |
| --- | --- | --- |
| No approved tool for the task | "What were you trying to get done?" | A gap in the approved list |
| Approved tool exists but is unknown | "Did you know we have a tool for that?" | A communication problem |
| Approved tool is slower or harder | "What made this one easier?" | A workflow problem to fix |
| Approved tool cannot do the job | "Where did the firm's tool fall short?" | A tool to replace or supplement |
| The rule was unclear | "Did you think this was allowed?" | A policy to rewrite in plain words |
| Deadline pressure | "What was due, and when?" | A staffing or planning issue |
| The tool was already on the device | "How did you first come across it?" | Personal accounts and devices to address |

Write the answers down by task, not by name. Ten people who each used a different free site to turn a scan into Word are telling you one thing: the firm needs a sanctioned way to convert scans.

Staff reach for generative artificial intelligence when the scan still has to become text before morning. The duty of competence includes handing them an approved route whose terms of service someone has already read. Client confidentiality is what a personal chatbot puts at risk, and a conflict of interest can sit in a chat history that already holds another client's facts. Name the approved converter where people will see it on the way to the upload box.

## How can a firm find the AI and OCR tools its people already use?

Use four sources together, because each one misses what another catches: a no-blame survey, expense and card records, the list of third-party apps in the firm's identity system, and the extensions installed in work browsers. A fifth source, network traffic reports, suits firms with IT staff.

### 1. Ask

A short survey catches free tools and personal accounts that leave no record on firm systems. Ask about tasks: "In the last three months, have you used any website or app, other than the ones on the firm's list, to convert, summarize, translate, transcribe or redraft something from a client matter?" Offer a list of task types to tick and a box for the tool's name. Promise, and mean, that honest answers will not be punished.

### 2. Follow the money

Search expense claims, firm credit card statements and reimbursement requests for software subscriptions. Paid personal plans for chatbots and converters appear here even when nothing else shows them. Free tools do not appear at all.

### 3. Check sign-in records

If staff sign in to outside tools with their work account, the firm's identity system has a list. In Google Workspace, Google's help page on [controlling which third-party and internal apps access Google Workspace data](https://support.google.com/a/answer/7281227) describes an "Accessed apps" list of "Apps that have accessed Google data", with the number of users for each app and the Google services each one requested. The page gives the path in the Admin console as Security, then Access and data control, then API controls.

The same page describes what an administrator can do next. For apps the firm has not configured, the options are to let users "sign in with Google to any third-party app", to limit them to apps "that request only basic profile information", or to block sign-in to unconfigured apps altogether. Choosing among those is a policy decision; the list is the discovery tool.

Know the limit of this source. It shows tools reached with a work account. A tool opened with a personal email address and a password leaves nothing here.

### 4. Look at browser extensions

Open the extensions page in each work browser, or have your IT provider pull the list centrally if the browsers are managed. An extension that can read page content can read a document open in a web-based document system. List every extension that summarizes, rewrites, translates, records or "assists".

### 5. Read network reports, if you have them

Microsoft's tutorial describes "cloud discovery" in Defender for Cloud Apps, which reports "which apps are being used" from device and network traffic data, and says the product "can't discover apps that aren't in the catalog" by default. A firm without that product or an IT team can skip this source and rely on the first four.

| Source | Finds | Misses |
| --- | --- | --- |
| Survey | Free tools, personal accounts, phones | Whatever people forget or withhold |
| Expenses | Paid plans | Free tools |
| Sign-in records | Tools used with a work account | Personal logins |
| Browser extensions | Add-ons in work browsers | Phones, home computers |
| Network reports | Sites visited from managed devices | Unmanaged devices |

## A partner's supervisory duty over tools staff choose for themselves

Under the ABA Model Rules, partners and lawyers with comparable managerial authority must make reasonable efforts to have measures in place so that lawyers and nonlawyers at the firm act consistently with the lawyers' professional obligations. A tool a staff member picked alone is still used on the firm's matters. Your state's version of the rules governs.

[ABA Model Rule 5.1](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_5_1_responsibilities_of_a_partner_or_supervisory_lawyer/), paragraph (a), says a partner "shall make reasonable efforts to ensure that the firm has in effect measures giving reasonable assurance that all lawyers in the firm conform to the Rules of Professional Conduct."

[ABA Model Rule 5.3](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_5_3_responsibilities_regarding_nonlawyer_assistant/) applies the same structure to nonlawyers "employed or retained by or associated with a lawyer". Paragraph (a) requires measures giving reasonable assurance "that the person's conduct is compatible with the professional obligations of the lawyer". Paragraph (b) puts a like duty on a lawyer "having direct supervisory authority over the nonlawyer".

Paragraph (c)(2) of each rule matters most once discovery begins. A partner or supervising lawyer is responsible for another person's conduct that would violate the rules if that lawyer "knows of the conduct at a time when its consequences can be avoided or mitigated but fails to take reasonable remedial action." An inventory creates knowledge. Plan the remedial steps before you run it, so that what you learn is acted on.

[ABA Formal Opinion 512](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf) applies these rules to generative AI: "Managerial lawyers must establish clear policies regarding the law firm's permissible use of GAI, and supervisory lawyers must make reasonable efforts to ensure that the firm's lawyers and nonlawyers comply with their professional obligations when using GAI tools." It adds that supervisory obligations include ensuring that subordinate lawyers and nonlawyers are trained. The opinion is advisory, and it addresses generative AI; whether your state has adopted or departed from it is yours to check.

Three practical consequences:

- Paragraph (a) of each rule asks for "measures" and "reasonable efforts". A dated inventory is a written record of one measure the firm took. Whether it is enough is for your jurisdiction's authorities.
- Paragraph (a) is written about measures put in place in advance. It is not limited to conduct a partner already knows about.
- Tools the firm does approve may still need the client's agreement. The post on [client consent to use AI tools](https://www.schemalaw.com/blog/client-consent-to-use-ai-tools) covers when and how.

## Running a shadow AI amnesty and inventory in a small firm

An amnesty is a fixed window in which anyone can report a tool they have used on firm work without discipline for the past use. Its product is an inventory: one row per tool, with who uses it, for what, on which account, and what the firm decided. Set a short window, such as two weeks.

### The procedure

1. **Name an owner.** One partner is responsible. In a solo practice that is you, and the exercise is still worth doing for your own devices and any contract staff.
2. **Set the terms in writing.** State the window, state that past use reported in the window will not be disciplined, and state the exception plainly: the firm must still deal with any client document that went somewhere it should not have.
3. **Announce it in person.** Say why: the firm cannot protect clients from tools it does not know about. Say that the aim is to get people working tools, not to take tools away.
4. **Collect.** Send the survey. Pull the expense, sign-in and extension lists during the same window.
5. **Build the inventory.** Use the template in this section.
6. **Triage each tool.** Read its terms. Sort it under the firm's policy tiers; the post on [sorting tools into approved, restricted and banned](https://www.schemalaw.com/blog/law-firm-ai-policy-document-upload#sorting-tools-into-approved-restricted-and-banned) owns those definitions.
7. **Deal with past uploads.** For each tool that received client documents, follow the incident routine in the post on [what to do after uploading a confidential document to the wrong tool](https://www.schemalaw.com/blog/accidentally-uploaded-a-confidential-document-to-chatgpt): record, delete, and decide on client notice.
8. **Report back.** Tell everyone what was found in general terms, what is now approved, and what replaces each stopped tool.
9. **Repeat.** Put the next inventory in the calendar. New tools and new features arrive between inventories.

### Inventory template

| Tool and web address | Task | Users | Account type | Client documents uploaded? | Terms read on | Decision | Replacement |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Free PDF-to-Word site | Converting scanned leases | 3 | None, anonymous | Yes | (date) | Stop | Approved converter |
| Chatbot, personal plan | Redrafting letters | 2 | Personal email | Yes, two matters | (date) | Stop; incident routine | Firm workspace |
| Summarizer extension | Summarizing web pages | 1 | Work Google sign-in | Unknown | (date) | Restrict pending review | None yet |

### A worked example

Castellan and Brooks is a six-person family law firm. Its managing partner, Ms. Castellan, announces a ten-day amnesty at the Monday meeting. The survey returns seven tools. The firm's card statement adds one paid chatbot subscription nobody mentioned. The Google Admin list adds a transcription app that an associate had signed in to with her work account.

Nine tools go into the inventory. Four received client documents: two free converters used for bank statements, the chatbot, and the transcription app. Ms. Castellan reads the terms of each. She approves none of the four, starts the incident routine for each, and notes that six of the nine tools were doing two jobs: converting scans and tidying draft letters. The firm picks one vetted tool for each job. The whole exercise fits in the ten-day window, and the inventory, dated and signed, goes into the firm's risk file.

## Moving staff from an unapproved tool to an approved one

Move staff by making the approved tool do the same task with no more effort than the one they found themselves. Start from the task list in your inventory, give each task a named approved tool, make that tool reachable in one step, and set a date after which the old one is not to be used.

A ban with no replacement leaves the task undone, and the task can then go to a different unapproved site. Pair every "stop" with a "use this instead".

### A migration checklist

- Each task in the inventory has a named approved tool, or a written decision that the task is done by hand.
- Each user has an account on the approved tool, on the firm's plan, under a work email.
- The approved tool is bookmarked or pinned where the old one was.
- A one-page guide shows the task done start to finish in the approved tool.
- Each user has run one real task in it with someone available to help.
- Old accounts are closed: documents deleted first, deletion recorded, then the account.
- Unapproved extensions are removed from work browsers.
- A cut-off date is announced and in the calendar.
- Someone is named to receive requests for new tools, with a promised response time.

### Test the replacement before you mandate it

A replacement that cannot do the task sends people back to the tool they had. Before you announce a replacement for scan conversion, run the firm's own difficult pages through it: the [faxed order](https://www.schemalaw.com/blog/how-to-ocr-a-faxed-document), the bank statement with small print, the handwritten intake form. Compare the output with the paper. Take Schema Law as an example of what to check: its homepage says recognition is strongest on clean print and weaker on faint faxes, stamps over text, and handwriting, and that nothing marks a passage as uncertain, so names, dates and amounts need checking against the original. Whichever tool you test, read its [security commitments](https://www.schemalaw.com/#security) and plan limits with the same care you gave the tools you are retiring. The Free plan on the [pricing page](https://www.schemalaw.com/pricing), at one document a day and up to three pages per document, is enough for a trial page and not for a production.

### Keep the door open

The last checklist item is what prevents the next round of shadow AI. If a paralegal finds a tool that does something the approved list cannot, there should be a short, known way to ask for it and a quick answer. A request that gets a prompt answer brings a new tool to the firm's attention before client documents reach it.

## Try it on a real document

Upload a scanned PDF and export it as Word, a searchable PDF or plain text on the free plan. Your documents are never used to train an AI model. [Start free](https://www.schemalaw.com/api/auth/google)