Find your three biggest AI risks in 15 minutes

In one 15-minute session you should leave with three concrete exposures, a 30-day plan, and a baseline policy you can copy. This is operational guidance for founders and small teams already using AI tools. It is not legal advice.

15-minute audit

Answer in a shared doc. If you do not know, treat that as an exposure.

1. Data inventory

• List the five kinds of information that went into an AI tool in the last 30 days: ________.
• Has anyone pasted a full client document into a chat? Yes / No / Not sure.
• Is there a written list of data that must not leave the company (secrets, health data, data about children, keys, full identifiers)? Yes / No.
• Do prompts include names, amounts, or contact details the task does not need? Yes / No.
• The data you would least want to see in a vendor’s history: ________.

2. Approved tools and vendors

• List every AI tool in use, including browser extensions and AI features inside CRM, email, or office software: ________.
• Is there a written list of approved tools and who may use them? Yes / No.
• Does anyone use a personal account for work? Yes / No / Not sure.
• For the two most used tools: who pays, whose name is on the account, and is training on your data turned off? ________.
• Have you read the data-processing terms for those two tools? Yes / No.

3. Access and least privilege

• Names (not job titles) with access to each tool: ________.
• Are passwords, API keys, or shared threads sitting in Slack, Notion, or a password manager with no clear owner? Yes / No.
• Can a contractor see the same chat history as the founder? Yes / No / N/A.
• Is access removed the same day someone stops working with you? Yes / No.
• Does any integration (plugin, API, automation) have more permission than the job it does? Yes / No / Not sure.

4. Human review and prohibited uses

• Can AI output reach a client without a person reading it in full? Yes / No.
• Which outputs need a named reviewer (proposals, contracts, production code, public content)? ________.
• Is it forbidden to present an AI answer as legal, medical, tax, or investment advice? Yes / No.
• Is it forbidden to invent testimonials or case studies, or to impersonate a real person? Yes / No.
• A use the team already does that you would not want written down: ________.

5. Offboarding, export, and incidents

• If a client asks you to delete or export what was pasted about them, do you know which tools to search and how? Yes / No.
• Person to notify if data is pasted by mistake: ________.
• Date of the last access review: ________.
• Is there a four-line incident note (what, when, who, next step)? Yes / No.
• Does anyone take chat history with them on a laptop or personal account when they leave? Yes / No / Not sure.

How to rank risk

Do not invent statistics. Score each finding on impact and likelihood.

Impact

• High: harm to a client, a leak that is hard to undo, or sensitive data (health, children, secrets, identifiers).
• Medium: serious rework, a public mistake you can correct, or lost internal trust.
• Low: inconvenience or a bad habit with no sensitive data.

Likelihood

• Frequent: weekly, or part of the daily flow.
• Occasional: monthly, or whenever people are in a rush.
• Rare: possible, but not how the work usually runs.

Rank high impact plus frequent or occasional first. If two items tie, pick the one that has already happened.

Your top three exposures

Fill this from the audit, not from generic risk lists.

If you cannot find three, leave rows blank. If you have more than three high-impact items, keep the ones one owner can close in 30 days.

30-day plan

Now: stop what you cannot reverse. Move the main accounts to a company plan, turn off training on your data, stop pasting full contracts, and name who to call if something is pasted by mistake.

Next 7 days: tool list, forbidden data classes, and who has access. Remove personal work accounts and access for people who no longer collaborate.

Next 30 days: baseline policy posted in the team channel, one prohibited use that actually stops, and proof that you can export or delete a thread.

An action with no owner does not count. Evidence is a dated artifact, not an intention.

AI usage baseline policy

Copy this, fill the brackets, and publish it. Adapt it to your company. Do not treat it as a legal opinion.

1. Scope. This applies to [company] and to anyone who uses AI in our work, including a personal account if the task is company work.

2. Data classes. Do not put the following into AI tools: [health data, data about children, passwords, card numbers, client secrets, full national IDs]. Client data is trimmed to what the task needs. If unsure, do not paste; ask [name].

3. Approved tools. Only: [tool 1, company account, training off; tool 2]. Any other tool needs OK from [name] before first use. Work accounts stay separate from personal accounts.

4. Prohibited uses. Do not send a client AI text that no human has read. Do not invent testimonials, metrics, or case studies. Do not impersonate a person. Do not treat AI as a licensed professional (legal, medical, tax, investment). Do not use a tool to get around a confidentiality duty to a client or an employee.

5. Human review. [Name or role] reviews before send or publish: proposals, contracts, production code, and any content that names a client. The reviewer owns the send, not the model.

6. Transparency. If a client asks whether we use AI on their work, we say: [short agreed sentence]. We do not hide the use when the result affects them.

7. Intellectual property. Do not paste third-party code, copy, or design if the licence forbids it. Material we generate for a client follows the same rule as other deliverables: [assignment or internal use]. Do not assume a vendor grants rights its contract does not grant.

8. Incidents. If data is pasted by mistake or output leaves the company without review: tell [name] the same day, note what / when / who / next step, and do not delete evidence until [name] says so.

9. Access and offboarding. Each tool has an account owner. Access is the minimum needed. On the day someone leaves, revoke accounts, keys, and extensions, and export what the company must keep.

10. Quarterly review. Every [months] the owner [name] reviews tools, access, and the three exposures, and updates this policy if a tool or a data type changes.

Example: a five-person B2B services firm

A small studio that writes proposals and runs sales follow-up. This is not a template for a clinic, a fintech, or a product team.

In 15 minutes they found: full RFPs in a personal chat account; a copy freelancer still had access; some proposals went out without a partner reading them; nobody knew how to delete a thread if a client asked.

Three exposures: client documents in a personal account (high impact, frequent); client send without review (high impact, occasional); access for someone who should be gone (medium impact, already happened).

Plan: today, company account and a ban on full RFPs. In 7 days, revoke the freelancer and publish what must not be pasted. In 30 days, the partner reviews every proposal and one owner can export threads. The policy fits on one page in the team channel.

A solo founder with no client data can stop after the inventory and two rules, without this level of formality.

When to ask for a business audit

If you are few people, you use two or three tools, and you do not handle sensitive data or client-facing workflows, this checklist is enough.

Consider a business audit (process, vendors, and ownership, not a legal stamp) only if at least one of these is true: sensitive data or client data under confidentiality; several AI vendors at once; recurring workflows that already feed deliverables; a team or outside collaborators; or a compliance risk a client, a contract, or a rule already names.

If none of those apply, do not buy a diagnosis. Close the three exposures and review in 90 days.