Draft a Trade-Secret AI-Risk Policy Memo for Employee Device Use

The task

In-house counsel and outside employment/IP attorneys get asked to update device and confidentiality policies when the business starts rolling out AI tools — Copilot, Claude, custom agents, browser extensions. The specific ask here: take the company's existing acceptable-use or BYOD policy, flag where AI agents on personal devices create new trade-secret exposure, and produce a redlined memo partners can send to the client. AI agents create new trade secret risks by allowing employees to access expertise or inadvertently expose proprietary information through personal devices, so the old policy almost certainly doesn't cover it.

Before AI

Pull the current policy. Read it against a mental checklist of AI-era exposures (prompt logging, model training on inputs, agent memory, browser-level DLP gaps on unmanaged devices). Mark up the PDF. Draft a cover memo. Three internal citations later, you're at 90–120 minutes for a first pass — and that's before the associate review. Guidance from firms like Grellas Shah, which warns companies to adopt strict AI policies, restrict data access on unmanaged devices, and update confidentiality agreements, has to be translated into actual policy language each time.

The workflow

Step 1 — Paste the policy and get a structured gap analysis.

The first prompt takes the current policy text (sample input below) and produces a gap table keyed to AI-specific exposures. Keep the taxonomy tight so Step 2 can act on it.

You are a senior employment and trade-secrets attorney reviewing a client's existing device/confidentiality policy for AI-agent exposure. Below is the policy text.

Produce a gap analysis as a markdown table with these columns: Exposure Category | Covered by Current Policy? (Yes / Partial / No) | Specific Gap | Risk Rating (High/Med/Low).

Use exactly these exposure categories, in order:
1. Use of generative AI tools on company-managed devices
2. Use of generative AI tools on personal/unmanaged devices
3. AI agents with persistent memory or browser access
4. Prompt and output logging by third-party AI vendors
5. Model training on employee inputs (opt-out / enterprise tier requirements)
6. Trade-secret marking and labeling of information before AI ingestion
7. Confidentiality agreement language covering AI inputs/outputs
8. Return-of-information obligations on separation (including AI chat histories, agent memory, synced browser data)
9. Monitoring, audit, and incident-response for AI misuse

After the table, write a two-paragraph "Legal Exposure Summary" explaining which gaps most threaten reasonable-measures defense under the DTSA and state UTSA analogs. Do not propose redlines yet.

Policy text follows:
---

NORTHWIND ANALYTICS, INC. — EMPLOYEE DEVICE AND CONFIDENTIAL INFORMATION POLICY (rev. March 2022)

1. PURPOSE. This Policy governs employee use of computing devices and handling of Confidential Information of Northwind Analytics, Inc. ("Company").

2. DEFINITIONS. "Confidential Information" means non-public business, technical, financial, and customer information of the Company, including source code, model weights, pricing data, client lists, and forecasting methodologies. "Company Device" means any laptop, phone, or tablet issued by the Company. "Personal Device" means any device owned by the employee.

3. ACCEPTABLE USE. Employees shall use Company Devices for business purposes. Limited personal use is permitted. Employees shall not install unapproved software. The IT department maintains a list of approved applications.

4. BYOD. Employees may access Company email and calendar on Personal Devices enrolled in the Company's mobile device management ("MDM") profile. Access to the Company's data warehouse, Git repositories, and CRM from Personal Devices is prohibited.

5. CONFIDENTIALITY. Employees shall not disclose Confidential Information to any third party without written authorization from the General Counsel. This obligation survives termination.

6. INTERNET USE. Employees shall exercise good judgment when using internet services. Posting Confidential Information to public websites, forums, or social media is prohibited.

7. RETURN OF INFORMATION. Upon separation, employees shall return all Company property and delete any Confidential Information in their possession.

8. MONITORING. The Company reserves the right to monitor use of Company Devices and Company networks.

9. VIOLATIONS. Violations may result in discipline up to and including termination and may give rise to civil or criminal liability.

Step 2 — Turn the gaps into a redlined memo with draft provisions.

Feed the gap analysis straight into drafting. Ask for strikethroughs and insertions in a format partners actually read.

Using the gap analysis you just produced, draft an attorney work-product memo titled "Redline and Recommended Amendments — Device and Confidential Information Policy (AI-Agent Exposure)."

Structure:

A. COVER MEMO (max 300 words). Addressed to "General Counsel, Northwind Analytics, Inc." From "[Firm]." Mark it "PRIVILEGED AND CONFIDENTIAL — ATTORNEY WORK PRODUCT." State the scope, the three highest-risk gaps, and the reasonable-measures rationale under DTSA § 1839(3)(A) and the UTSA. Do not invent case citations.

B. REDLINE. Reproduce each affected section of the original policy. Show deletions as ~~strikethrough~~ and insertions as **bold**. Only redline sections that need changes; leave untouched sections out.

C. NEW PROVISIONS. Draft full text, ready to insert, for:
   - A new Section 10: "Generative AI and AI Agents" covering approved tools, enterprise-tier requirement, prohibition on inputting Confidential Information to consumer AI tools or AI tools on unmanaged Personal Devices, and prohibition on connecting AI agents to Company accounts without written IT approval.
   - A new Section 11: "AI-Related Return of Information" covering export and deletion of chat histories, agent memory, custom GPTs/projects, synced browser profiles, and third-party integrations on separation.
   - A rider to Section 5 (Confidentiality) extending the definition of "disclosure" to include submission to AI systems that log, retain, or train on inputs.

D. OPEN QUESTIONS FOR THE CLIENT. 4–6 bullet points — things you need to confirm (enterprise contracts in place, MDM scope, incident response playbook, whether contractors are covered, choice-of-law for the confidentiality agreement, works-made-for-hire implications of AI-assisted output).

Write in neutral, deal-counsel tone. No marketing language. No emoji.

Step 3 — Pressure-test the draft against a reasonable-measures checklist.

Trade-secret protection turns on whether the company took "reasonable measures." Ask the model to grade its own draft against that standard before you send it.

Review the memo and redline you just produced. Produce a short "Reasonable Measures Self-Audit" (max 400 words) answering:

1. If a court applied the DTSA reasonable-measures standard to the amended policy, which three provisions would carry the most weight, and why?
2. Which two provisions are weakest or most likely to be challenged as vague or unenforceable (e.g., overbreadth, California Business and Professions Code § 16600 concerns, NLRA Section 7 concerns for non-supervisory employees)? Suggest one tightening edit for each.
3. What single document or policy outside this memo should the client update next to make the amendments effective (e.g., employee confidentiality agreement, separation certification, vendor DPA template)?

End with a one-line recommendation on whether this draft is ready to send to the client or needs a second pass.

Step 4 — Hand-check before it leaves the building.

Read the redline against the original. Confirm defined terms are consistent (watch "Confidential Information" vs. "Proprietary Information" drift). Verify no fabricated case citations slipped in — if any appeared in the cover memo, strike them. Run the new Section 10 past IT to confirm the enterprise-tier assumption is real.

Gotchas

• Hallucinated citations. The model will sometimes invent DTSA subsections or state statutes. The prompts tell it not to, but check. If a cite looks oddly specific, verify it.
• One-size-fits-all redlines. If the client is in California, § 16600 and the recent amendments to Labor Code § 925 change what you can enforce. The model's default drafting leans toward broad restrictive language — tighten for CA, MA, and WA.
• Contractor coverage. Original policies often apply only to "employees." AI exposure through contractors and agencies is a bigger blind spot than through W-2s. If the client uses 1099s or staffing firms, you'll need a parallel contractor addendum — the model won't volunteer this unless you ask.
• Privilege labeling. The memo is marked work product, but if it gets forwarded to HR or IT without a privilege waiver conversation, that protection erodes fast. Flag distribution limits when you deliver it.
• "Agent" is a moving target. A Chrome extension that reads tabs, a desktop app with filesystem access, and a cloud agent orchestrating API calls all raise different exposure. The draft Section 10 treats them as one category; refine before finalizing for a client with a mature AI stack. For more context on how this threat model is evolving, see the Crunchbase commentary from Grellas Shah and the Baker Donelson analysis of agent-enabled misappropriation.

Time saved

Roughly 90 minutes per policy review, conservatively. The gap analysis and first-draft redline are the slow parts of a manual pass; the model handles both in one shot, leaving you with partner-level work (judgment calls on enforceability, jurisdiction, and client-specific context) instead of blank-page drafting. The self-audit step in Step 3 catches the most common issues before the draft lands with a partner, saving a review cycle. This is an estimate, not a measurement — if your firm already has a strong policy template, savings are smaller; if you're drafting from the client's text cold, savings are larger.

What we got when we ran it

We executed every prompt above, in order, against the sample input (model: claude-sonnet-4-6). Outputs below are real and unedited apart from length truncation.

Step 1 output

# Gap Analysis: Northwind Analytics, Inc. — Device & Confidentiality Policy (rev. March 2022) ## AI-Agent Exposure Review | Exposure Category | Covered by Current Policy? | Specific Gap | Risk Rating | |---|---|---|---| | 1. Use of generative AI tools on company-managed devices | Partial | Section 3 prohibits unapproved software and references an IT-maintained approved-application list, but no provision explicitly addresses generative AI tools (e.g., ChatGPT, Copilot, Claude) as a distinct category. There is no AI-specific acceptable-use rule, no approved/prohibited vendor list for AI services, and no guidance on what classes of Confidential Information may or may not be entered into any AI interface, even on a Company Device. The "good judgment" internet-use standard in Section 6 is legally insufficient to establish the specific, affirmative steps required for a reasonable-measures defense. | High | | 2. Use of generative AI tools on personal/unmanaged devices | No | Section 4 limits BYOD MDM enrollment to email and calendar and prohibits access to the data warehouse, Git, and CRM from Personal Devices, but no provision addresses use of cloud-hosted AI tools on Personal Devices where employees may independently paste or upload Confidential Information outside any MDM-controlled channel. The policy creates no explicit prohibition and no mechanism to detect or deter this vector. | High | | 3. AI agents with persistent memory or browser access | No | The policy contains no reference to AI agents, autonomous workflows, browser-integrated AI assistants (e.g., Copilot in Edge, browser-based agents with memory enabled),

…output truncated at 250 words.

Step 2 output

# PRIVILEGED AND CONFIDENTIAL — ATTORNEY WORK PRODUCT # Protected by Attorney-Client Privilege --- # Redline and Recommended Amendments ## Device and Confidential Information Policy (AI-Agent Exposure) ### Northwind Analytics, Inc. --- ## A. COVER MEMO TO: General Counsel, Northwind Analytics, Inc. FROM: [Firm] DATE: [Date] RE: Redline and Recommended Amendments — Employee Device and Confidential Information Policy (rev. March 2022) — AI-Agent Exposure CLASSIFICATION: PRIVILEGED AND CONFIDENTIAL — ATTORNEY WORK PRODUCT --- Scope. At your request, we reviewed the March 2022 Device and Confidential Information Policy ("Policy") for gaps created by the emergence of generative AI tools and AI agents since the Policy's last revision. Our gap analysis identified nine exposure categories. This memorandum presents redlines to affected existing sections and full-text drafts of two new sections and one rider, prioritized by legal risk. Three Highest-Risk Gaps. First, the Policy imposes no restriction on employee submission of Confidential Information to consumer-tier generative AI products whose terms of service permit training on user inputs. If trade-secret material has been submitted to such products, protectable secrecy may have been compromised through an authorized channel the Company failed to govern, which is among the most difficult postures from which to assert misappropriation. Second, the Policy contains no classification or marking requirement applicable before AI ingestion, leaving the Company without documentary evidence that it identified and segregated its most sensitive assets — a predicate showing in both federal and state trade-secret litigation. Third, the return-of-information obligation on separation does not enumerate AI chat

…output truncated at 250 words.

Step 3 output

# PRIVILEGED AND CONFIDENTIAL — ATTORNEY WORK PRODUCT ## Reasonable Measures Self-Audit ### Northwind Analytics, Inc. — Device and Confidential Information Policy (AI-Agent Exposure Amendments) --- ## 1. Three Provisions Carrying the Most Weight Under the DTSA Reasonable-Measures Standard Section 10.3 (Enterprise Tier Requirement). This provision will carry the greatest evidentiary weight because it directly closes the gap most likely to destroy protectable secrecy before any misappropriation occurs. By prohibiting Consumer AI Tier use for any Confidential Information purpose and requiring a vendor contract that expressly bars training-data use, the Company creates a documented, affirmative, prospective barrier — precisely the kind of specific and consistently enforced measure courts look for under DTSA § 1839(3)(A). A plaintiff relying on this provision can show not only that a rule existed but that it was calibrated to the specific threat mechanism at issue. Section 10.5 (AI Agents and Autonomous Workflows) combined with the AI Systems Register. The written-approval requirement for agent connections and the quarterly-reviewed AI Systems Register operationalize the policy into a running audit trail. Courts assessing reasonable measures give meaningful weight to evidence that a company actively monitored and controlled access channels, not merely prohibited misuse in the abstract. A maintained register with named approvers and access-control records is substantially more defensible than a blanket prohibition standing alone. Section 11.4 (AI-Resident Information Certification). A signed, separation-date certification creates individualized, date-stamped evidence that the departing employee was specifically put on notice of AI-related return obligations. This is particularly valuable in post-separation misappropriation

…output truncated at 250 words.

---

This content is for informational purposes only and is not legal advice. Confirm confidentiality, privilege, and jurisdictional rules before using any AI tool with client matters.