AI acceptable use policy template

1. Purpose

AI assistants, coding assistants and AI features inside everyday software can save the people at Northwind Cloud Inc real time. They also create new ways for information to leave the company: text pasted into a chat window, a file uploaded for a summary, a repository opened by an assistant, or a feature in Northwind that sends customer data to a model provider. Output can also be wrong in ways that look convincing.

This policy sets out which AI tools may be used, what information may go into them, how AI output is checked before it reaches customers or production, and how AI is used inside Northwind. It expands on the AI rule in the Northwind Cloud Inc Acceptable Use Policy and applies alongside the Data Classification and Handling Policy, the Vendor and Third-Party Risk Management Policy and the Incident Response Policy.

2. Scope

This policy applies to every employee, contractor, intern and temporary worker of Northwind Cloud Inc, and to anyone else granted a company account. It covers any use of AI for Northwind Cloud Inc work, on company or personal devices, including:

• AI chat assistants used through a website, desktop app, browser extension or phone app.
• AI coding assistants in editors, terminals, code review and the command line, including assistants that can run commands or change files on their own.
• AI features inside software Northwind Cloud Inc already uses, such as writing aids, meeting transcription and summaries, search over company documents and support ticket drafting. This includes any AI features in Supabase, Stripe, Datadog and Slack.
• AI models and APIs used in Northwind or in internal automation, whether called directly or through another vendor.

All information is covered according to its classification under the Data Classification and Handling Policy. Using a personal AI account for Northwind Cloud Inc work counts as using an unapproved tool.

3. Roles and Responsibilities

• Executive Management (Priya Natarajan, CEO) approves this policy, decides whether to accept the risk of AI features in Northwind, and makes sure the rules apply equally to everyone.
• Security Owner (Dana Whitfield, CTO) owns this policy and the approved AI tools list, reviews AI tools under the vendor risk process, answers questions about whether a use is allowed, and investigates reported exposures.
• Engineering Lead sets up approved coding assistants for the team, keeps secret scanning in GitHub enabled, and owns the testing, logging and switch-off of AI features in Northwind.
• People/HR Lead includes this policy in onboarding and security awareness training and keeps the acknowledgment records.
• All Personnel use only approved tools, follow the data rules in section 4, review AI output before relying on it, and report mistakes promptly.

4. Policy Statements

• 4.1 Northwind Cloud Inc work uses only AI tools on the approved AI tools list, which the Security Owner (Dana Whitfield, CTO) owns and keeps in [Fill in: where the approved AI tools list is kept, for example the security wiki]. For each tool the list records the vendor, the approved uses, the highest data classification allowed, whether a company account and single sign-on are set up, who reviewed it and when, and when it is next reviewed. A tool that is not on the list is not approved, however widely it is used elsewhere.
• 4.2 An AI tool is approved only after it has been reviewed as a vendor under the Vendor and Third-Party Risk Management Policy. The Security Owner reads the vendor's current terms for the plan Northwind Cloud Inc will actually use and records what they say about whether inputs and outputs are used to train models, how long they are kept, who at the vendor can see them, which sub-processors are involved, where data is stored and which admin controls are available. Terms differ between a vendor's plans and change over time, so they are checked for the specific plan and re-checked at every review; nothing is assumed from marketing pages or from another company's choice.
• 4.3 AI tools are used only through company accounts, with single sign-on through Okta where the tool supports it, and multi-factor authentication where it does not. Accounts are requested and removed through the Access Control Policy process and are removed at offboarding like any other account. Personal accounts and free plans are not used for any information above Public.
• 4.4 Information goes into an AI tool only as the table below allows, based on its classification under the Data Classification and Handling Policy. Credentials, passwords, API keys, tokens, private keys and connection strings are never entered into any AI tool, approved or not. Customer data is entered only into a tool the list approves for Restricted data and whose terms the Security Owner has confirmed do not allow the vendor to train models on it. When unsure, treat the information as the higher level and ask the Security Owner first.

• 4.5 Personal data is entered into an approved AI tool only when the task needs it. Personnel remove or replace names, email addresses, account identifiers and other details the task does not need before pasting, and never use AI tools to build profiles of individuals or to look up information about them.
• 4.6 AI output is a draft. A person reviews it before it reaches a customer, is published, is relied on for a security, legal, financial or contractual decision, or is used in production. The reviewer checks facts, figures, names, links and references, and is accountable for the result as if they had written it. Answers to customer security questionnaires drafted with AI are checked against how Northwind Cloud Inc actually operates before they are sent. AI is not the only basis for decisions about individuals, such as hiring or performance.
• 4.7 Coding assistants are used only if they are on the approved list and signed into with a company account. AI-generated code goes through the same pull request review, automated tests and approval as any other change under the Secure Software Development Policy and the Change Management Policy, and the reviewer is told when a change is substantially AI-generated. Secret scanning in GitHub stays enabled and is never bypassed. Where the tool offers settings to exclude files such as environment and secrets files, or to flag suggestions that match public code, they are turned on. Code that looks copied from a recognizable outside project is checked for its license before it is merged.
• 4.8 Assistants that can run commands, change files or act in other systems on their own are given the narrowest access that does the job. They are never given production credentials or administrator tokens, their changes reach production only through the normal review and deployment path, and connecting any AI tool to company email, documents, chat, GitHub or cloud accounts through an integration or OAuth grant needs Security Owner approval as a separate entry on the list.
• 4.9 When Northwind uses an AI model or API to deliver a feature, the model provider is reviewed and recorded as a vendor, and as a sub-processor where customer agreements require one. Customer data is sent to it only under terms the Security Owner has confirmed on training use and retention. Customers are told about the feature and the provider where customer contracts or applicable law require it. Before release the feature is tested for wrong, harmful or unexpected output and for attempts to make it ignore its instructions or reveal other customers' data. Requests and responses are logged in line with the Logging and Monitoring Policy, keeping only what is needed, and the feature can be switched off without a full release.
• 4.10 The following are prohibited: entering information into an AI tool beyond what the table in 4.4 allows; using AI to get around a security control, write malware or probe systems Northwind Cloud Inc does not own; using AI to impersonate a real person or to produce content meant to mislead customers or colleagues; and presenting AI output to a customer as reviewed when it was not.
• 4.11 Entering Confidential or Restricted information into an AI tool not approved for it, an AI tool showing one person's data to another, or AI output that exposes information it should not, is a security event. It is reported to security@northwindcloud.example as soon as it is noticed and handled under the Incident Response Policy. Prompt good-faith reporting, including of one's own mistake, is never penalized.
• 4.12 Everyone completes training on this policy before using AI tools for Northwind Cloud Inc work and again at each annual review, and acknowledges the policy in writing. The training covers the approved list, the data table in 4.4, reviewing output and how to report a mistake.

5. Procedures

5.1 Requesting a new AI tool or AI feature

1. The requester sends the Security Owner the tool name, the vendor, the plan, the business need, the people who will use it and the highest classification of information it will touch.
2. The Security Owner reviews the vendor under the Vendor and Third-Party Risk Management Policy and records the terms listed in 4.2, using the vendor's current published terms, its security documentation or report, and the admin settings of the plan.
3. The Security Owner decides within five business days: approved, approved with limits (for example Internal information only), or declined, and records the decision and the reason.
4. If approved, the tool is added to the approved AI tools list with its allowed classification, a company account is set up with single sign-on through Okta where supported, and admin settings such as conversation retention or training opt-outs are set as the review requires.
5. The Security Owner tells the requester and the people who will use it what the tool is approved for.

5.2 Turning on an AI feature in a tool already in use

An AI feature added to an existing tool is treated as a new request under 5.1, because it may send data to a new sub-processor or under different terms. Where the tool lets an administrator keep new AI features off until they are reviewed, they stay off until the review is done.

5.3 Quarterly check of the approved list

1. The Security Owner re-reads the terms of each approved tool and records any change that affects its approval.
2. The Security Owner looks for AI tools in use that are not on the list, for example in the Okta connected apps and OAuth grants, browser extension reports where available, and expense reports.
3. Unapproved tools found are either reviewed under 5.1 or removed, and the people using them are reminded of this policy.
4. Tools no one uses any more are removed from the list and their accounts are closed.
5. The check, its date and its findings are recorded with the list.

5.4 Setting up coding assistants

1. The Engineering Lead sets up each approved coding assistant under a company account or organization, not individual personal accounts.
2. The Engineering Lead turns on the file exclusions and public-code filters the tool offers, and records the settings used.
3. The Engineering Lead confirms that secret scanning and required pull request review are enabled in GitHub for every repository the assistant can reach.
4. Assistants that can act on their own are limited to development environments and given scoped, short-lived credentials.

5.5 Releasing an AI feature in Northwind

1. The Engineering Lead records the feature, the model provider, the customer data it sends and the reason it needs that data.
2. The Security Owner confirms the provider is an approved vendor with confirmed terms on training use and retention, and checks whether customer contracts require notice or an update to a sub-processor list.
3. The feature is tested before release for wrong or harmful output, for attempts to override its instructions and for any path that could expose one customer's data to another. Test results are kept with the change record.
4. Logging, data minimization and the switch-off method are in place and documented.
5. Executive Management (Priya Natarajan, CEO) approves the release and any customer communication the contracts require.

5.6 Handling an AI-related exposure

1. The person who notices reports it to security@northwindcloud.example immediately, with what was entered, into which tool, under which account and when.
2. The Security Owner records it as a security event under the Incident Response Policy and decides its severity.
3. The content is deleted from the tool where the tool allows, and deletion is requested from the vendor where its terms provide for it. The request and the answer are kept.
4. Any credential, key or token that was exposed is rotated straight away, whether or not the tool was approved.
5. If customer or personal data was involved, the notification assessment in the Incident Response Policy is carried out.
6. The cause is recorded and the approved list, training or tool settings are updated to stop it happening again.

5.7 Training and acknowledgment

The People/HR Lead includes this policy in onboarding, before AI tool access is granted, and in the annual security awareness training. Each person acknowledges the policy in writing, and the acknowledgment and training completion are kept with the other policy acknowledgments.

6. Exceptions

Exceptions are approved in writing by the Security Owner, recorded in the exception register with the business reason, the compensating controls and an expiry date no later than twelve months out, and reviewed at each annual review. No exception permits entering credentials, secrets or keys into an AI tool, or sending customer data to a tool whose terms the Security Owner has not reviewed.

7. Enforcement

The People/HR Lead and the Security Owner handle violations together. Consequences are proportionate to intent and impact and range from a documented conversation and retraining, through removal of access to AI tools, to termination of employment or contract. Deliberately sending customer data or secrets to an unapproved tool, or hiding an exposure, is escalated to Executive Management. Mistakes reported promptly in good faith are treated as a chance to improve the rules and the training.

8. Review Cadence

The Security Owner reviews this policy at the annual policy review, when Northwind Cloud Inc adopts a new kind of AI tool or adds AI to Northwind, when a vendor on the approved list materially changes its terms, and after any incident in which AI use was a contributing factor. Revisions are approved by Priya Natarajan, CEO, recorded in section 9 and re-acknowledged by all personnel.

9. Revision History