Draft an AI Governance Gap Analysis Memo From a Policy Document

The task

You're the founder. Someone — a prospect's security team, your new Head of Ops, maybe a term sheet checklist — just asked how you govern AI usage internally. You have a policy (probably a Notion page you wrote in 20 minutes six months ago), but you need to know where it's thin before someone else points it out. This workflow turns that existing doc into a structured gap analysis memo you can act on this week.

Before AI

The honest baseline: most founders don't do this at all until a customer forces it. If you do, it means a half-day of reading NIST AI RMF, the EU AI Act summary, and a couple of SOC 2 blog posts, then manually diffing against your own policy in a Google Doc. Call it 3-4 hours and a lot of tab-switching to produce something you're not sure is complete.

Worth doing now because, per Theta Lake's governance research covered by Fintech Global, AI assistants, meeting notetakers and agentic tools are now active contributors to daily business, and the rush to deploy them is outpacing the frameworks needed to keep them in check. The founders who get asked about this first are the ones selling into regulated buyers — and the common failure mode is overcorrecting into restrictive controls that simply block access to AI features or tools like Microsoft Copilot or Zoom AI Companion, which just pushes usage into shadow AI.

The workflow

Step 1 — Extract what your current policy actually covers

Paste your existing AI usage policy into the prompt below. The model's job here is not to critique yet — it's to produce a clean inventory of what the policy claims to cover, so Step 2 has something concrete to compare against a standard framework.

You are an AI governance analyst helping a startup founder review their internal AI usage policy.

Below is the company's current AI policy document. Your task:

1. Produce a structured inventory of what this policy ACTUALLY addresses. Use these categories:
   - Approved tools and vendors
   - Data classification and what may/may not be entered into AI tools
   - Human-in-the-loop requirements (what decisions AI cannot make unsupervised)
   - Agentic / autonomous AI permissions (tools that take actions, not just generate text)
   - Meeting recording, transcription, and notetaker rules
   - Customer-facing AI disclosure
   - IP, confidentiality, and training-data opt-outs
   - Incident reporting (what counts as an AI incident, who to tell)
   - Review cadence and ownership

2. For each category, quote the exact policy language that covers it (or write "NOT ADDRESSED").

3. At the end, list any policy statements that don't fit the categories above under "Other provisions."

Be literal. Do not infer coverage from vague language — if the policy says "use AI responsibly" without specifics, that is NOT ADDRESSED.

Here is the policy:
---

NORTHWIND LABS — AI USAGE POLICY (v0.3, last updated March 2026)

Purpose
Northwind Labs ("the Company") encourages responsible use of AI tools to improve productivity. This policy applies to all employees and contractors.

Approved tools
Employees may use ChatGPT (Team plan), Claude (Team plan), and GitHub Copilot. Use of other AI tools requires written approval from the CTO (Sam Reyes).

Data handling
Do not paste customer PII, source code marked "Confidential," or unreleased financial data into any AI tool. Use good judgment. When in doubt, ask.

Meetings
If you record a meeting with an external party, you must inform them first. Internal meetings may be recorded at the organizer's discretion.

Code
AI-generated code must be reviewed by a human before merging to main. Do not commit AI output verbatim without review.

Violations
Repeated violations may result in disciplinary action. Report concerns to People Ops (Dana Okafor).

This policy will be reviewed annually.

Step 2 — Benchmark against a real framework and surface the gaps

Now the model takes that inventory and compares it to a recognizable governance framework. We're explicitly asking it to be concrete about agentic AI, because that's the fastest-moving gap for most startups right now and the thing enterprise buyers are starting to ask about.

You now have an inventory of what the policy covers. Next task:

Compare that inventory against a composite baseline drawn from the NIST AI Risk Management Framework (Govern, Map, Measure, Manage functions), ISO/IEC 42001 (AI management systems), and common SOC 2 vendor questionnaire items on AI.

Produce a gap table with these columns:
| # | Control area | Baseline expectation (1 sentence) | Current policy status | Gap severity (High / Medium / Low) | Why this matters for a startup selling to enterprise |

Cover at minimum: approved tool list maintenance, data classification tied to tool approval, prohibited inputs with examples, agentic AI action permissions (what agents can execute vs. only draft), human review thresholds, meeting notetaker consent for external participants, retention and deletion of AI outputs, vendor sub-processor disclosure, customer-facing AI disclosure, employee training requirement, incident definition and reporting workflow, named policy owner, review cadence shorter than annual, and shadow AI detection approach.

Severity rubric:
- HIGH: likely to fail an enterprise security review or create legal exposure (privacy, IP, regulated-industry rules).
- MEDIUM: would require a written explanation to a buyer; not a blocker.
- LOW: nice to have; good hygiene.

Be specific. "Policy is vague" is not useful — say WHAT is missing.

Step 3 — Turn the gaps into a prioritized memo your team will actually read

Last step converts the analysis into something you'd send to your co-founder or paste into a board update. The constraint — "a founder who has 90 minutes this week" — is doing the work; it forces the model to drop boil-the-ocean recommendations.

Final task: write the memo.

Audience: the founder/CEO and one other operator (head of ops or CTO). They have ~90 minutes of focused time THIS WEEK to improve AI governance, and roughly one day per month after that.

Format exactly as follows, in plain markdown:

# AI Governance Gap Analysis — [Company Name]
**Date:** [today]
**Policy version reviewed:** [from the input]

## TL;DR
Three sentences. What's the current posture, what are the top risks, what's the recommended first move.

## Top 3 gaps to close this week
For each: one-sentence gap, one-sentence why it matters, and a concrete remediation written as a policy clause the founder can paste into the existing doc (not "develop a framework" — actual sentences for the policy).

## Backlog (next 30-90 days)
Bulleted list of the Medium-severity items with a one-line action each. No more than 8 bullets.

## Won't do yet (and why)
2-4 items that look like gaps but are reasonable to defer at current company stage. Be explicit about the tradeoff.

## Owner and review cadence
Suggest a named role (not a person) as policy owner and a review cadence. Justify the cadence in one sentence.

Keep the whole memo under 600 words. No hedging language ("it may be prudent to consider"). Direct voice.

Gotchas

• The model will hallucinate framework citations if you let it. The Step 2 prompt deliberately asks for a "composite baseline" rather than exact clause numbers from NIST or ISO. If you need auditable citations, verify them manually against the source documents — don't trust specific control IDs the model produces.
• Agentic AI is the weakest spot in most startup policies and also the hardest for the model to reason about — it will tend to lump "an agent that drafts an email" with "an agent that sends the email and spends money." Re-read the agentic row in the gap table with a skeptical eye.
• If your policy is very short (under ~200 words), Step 1 will mostly return "NOT ADDRESSED" and Step 3's "Won't do yet" section will be thin. That's accurate, not a bug — it means you're starting from near-zero and the memo should say so.
• Don't ship the memo to a customer. This is an internal diagnostic. External-facing trust docs (DPA addenda, security questionnaires) need a human lawyer or compliance review, especially if you touch health, financial, or EU personal data.
• Shadow AI is a people problem the memo won't fix. As Theta Lake's research notes, overly restrictive policies drive employees toward 'shadow AI', where sensitive tasks are performed in unmonitored, unapproved environments. If your Step 3 output reads as pure prohibition, loosen the "approved tools" clause before you circulate it.

Time saved

Estimated 3-4 hours saved per policy review versus doing it manually with framework PDFs open in tabs. The real leverage is cadence: a review that used to happen once a year (or never) becomes something you can rerun each quarter in under 15 minutes as your tool stack changes. For a founder prepping for an enterprise security review or a Series A diligence, this is the difference between "we have a policy" and "we have a policy and know exactly where it's thin."

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

Northwind Labs AI Policy — Structured Inventory --- ### 1. Approved Tools and Vendors **ADDRESSED** > "Employees may use ChatGPT (Team plan), Claude (Team plan), and GitHub Copilot. Use of other AI tools requires written approval from the CTO (Sam Reyes)." --- ### 2. Data Classification and What May/May Not Be Entered into AI Tools **PARTIALLY ADDRESSED** *(prohibition exists; classification framework does not)* > "Do not paste customer PII, source code marked 'Confidential,' or unreleased financial data into any AI tool. Use good judgment. When in doubt, ask." **Gap note (literal):** The policy names three prohibited data types but provides no classification tiers, no definition of PII, no list of what *is* permitted, and no escalation path for "when in doubt" beyond the vague "ask" (ask whom is unspecified). --- ### 3. Human-in-the-Loop Requirements **PARTIALLY ADDRESSED** *(code review only; no general human-in-the-loop rule)* > "AI-generated code must be reviewed by a human before merging to main. Do not commit AI output verbatim without review." **Gap note (literal):** This requirement applies exclusively to code. No other decision types (hiring, legal, financial, customer-facing content, etc.) are covered. --- ### 4. Agentic / Autonomous AI Permissions **NOT ADDRESSED** The policy contains no language governing AI tools that take actions (e.g., browsing, sending emails, executing code, interacting with APIs) rather than generating text. --- ### 5. Meeting Recording, Transcription, and Notetaker Rules **PARTIALLY ADDRESSED** *(consent rule exists; AI-specific transcription/notetaker bots not mentioned)* > "If you record a meeting with an external party, you

…output truncated at 250 words.

Step 2 output

Northwind Labs AI Policy — Gap Analysis Against NIST AI RMF, ISO 42001, and SOC 2 Enterprise Baseline --- | # | Control Area | Baseline Expectation | Current Policy Status | Gap Severity | Why This Matters for a Startup Selling to Enterprise | |---|---|---|---|---|---| | 1 | **Approved tool list maintenance** | A versioned, actively maintained register of approved AI tools exists, with each tool's vendor, plan tier, data residency, and approval date recorded. | Partially addressed. Three tools named (ChatGPT Team, Claude Team, GitHub Copilot). No register, no version dates per tool, no data residency information, no process for removing tools from the approved list. | **MEDIUM** | Enterprise buyers running vendor questionnaires (SOC 2 Type II, SIG Lite) will ask for an asset inventory of AI tools in use. A three-item list in a prose policy is not a maintainable register. If a tool is deprecated or has a data breach, there is no mechanism to revoke approval and notify employees. | | 2 | **Data classification tied to tool approval** | Each approved AI tool is mapped to the data classification levels it may process (e.g., Public, Internal, Confidential, Regulated), so employees know which tool is appropriate for which data type. | Not addressed. The policy prohibits three input types (customer PII, Confidential source code, unreleased financials) but does not define classification tiers and does not map any classification level to any specific tool. | **HIGH** | Enterprise customers in regulated industries (healthcare, financial

…output truncated at 250 words.

Step 3 output

# AI Governance Gap Analysis — Northwind Labs Date: 2026-07-14 Policy version reviewed: v0.3 (March 2026) --- ## TL;DR The current policy controls three things adequately: approved tools, code review, and external meeting recording consent. The top risks are legal exposure from undisclosed AI sub-processors, missing incident response procedure, and zero governance over agentic AI tools — any of which can kill an enterprise deal or trigger a breach notification obligation. Start this week by patching the three clauses below directly into v0.4. --- ## Top 3 Gaps to Close This Week 1. Sub-processor disclosure and vendor DPAs The policy doesn't treat OpenAI, Anthropic, or GitHub as data processors, so there's nothing to show enterprise buyers or regulators when they ask. GDPR Article 28 and standard enterprise DPAs require it. > Paste this into the policy: "OpenAI (ChatGPT), Anthropic (Claude), and GitHub (Copilot) are approved AI sub-processors. Employees must use only the designated Team/Enterprise plan accounts — never free-tier or personal accounts — because those plans include data processing agreements with training-data opt-outs. The CTO maintains a sub-processor register. Customers are notified before any new AI sub-processor is added." --- 2. Incident definition and reporting workflow There is no definition of an AI incident and no path to escalate one. If an employee pastes customer data into the wrong tool, no one knows who to call or when. Enterprise contracts typically require breach notification within 48–72 hours. > Paste this into the policy: "An AI incident is any of the

…output truncated at 250 words.