An AI data access audit reviews the information AI could reach through staff accounts, connected apps, folders, inboxes, CRM records, and custom agent permissions. It does not certify security. It finds access risks before AI makes them easier to surface or reuse.
If you are planning Copilot, Google Workspace AI, ChatGPT use, or a custom agent, start by naming one workflow and checking the data it needs. The $1,000 AI Agent Assessment maps the workflow, sources, permissions, approval gates, and safe next step.
An AI data access audit checks what information people and AI-enabled tools can reach, whether that access is appropriate, and what should stay restricted, read-only, draft-only, or human-approved.
In plain English, the audit asks: if staff or AI searches, summarises, drafts from, or connects to this information, what could it see and do next?
It covers files, CRM notes, inboxes, calendars, meeting notes, exports, HR folders, website systems, templates, and third-party tools.
AI usually does not create permission problems from nothing. It makes existing problems visible: overshared folders become searchable, outdated templates get reused, and broad CRM exports feed sensitive drafts.
Run the audit before a broad rollout, not after staff have connected tools.
Most access risk sits in everyday systems, not exotic AI infrastructure. Start with the tools your staff already use.
Check shared folders, Teams channels, meeting recordings, templates, board papers, archived projects, staff files, client folders, and inherited permissions. If staff use Microsoft 365, compare the audit with the Microsoft 365 Copilot Governance Checklist.
Review shared drives, document links, Gmail labels, calendar visibility, Meet notes, Docs, Sheets, and Chat history. If the business uses Gemini or Workspace AI, use the Google Workspace AI Governance Checklist.
CRM data can include lead sources, pricing context, complaints, relationship notes, proposal history, and permission signals. A sales agent may need some context, but not broad access by default.
Website agents may read public pages, forms, analytics summaries, approved offer copy, issue queues, and schema. They should not publish, change pricing, edit forms, alter tracking, touch DNS, or invent proof without review.
Finance, HR, contracts, disputes, credentials, and confidential client material need stronger rules. For many first pilots, exclude these sources or use a human-prepared summary.
Use this checklist before connecting AI to live company systems.
Name approved AI tools, account owners, whether business data can be entered, and whether personal accounts are allowed. If unapproved AI use is common, read the Shadow AI policy guide before expanding access.
Mark customer, staff, finance, legal, commercial, strategy, credential, and confidential project information. Decide whether each source is allowed, restricted, excluded, or human-summarised. If personal information is involved, run an AI Privacy Impact Assessment NZ before connecting tools.
Look for folders, drives, channels, dashboards, or exports visible to wider groups than intended. Broad access is convenient until AI makes search and summarisation easier.
Reading a source is not the same as changing it. Separate read, draft, recommend, update, send, publish, delete, admin, and blocked permissions with the AI Agent Permissions Checklist.
Write simple prompt rules for customer information, staff information, contracts, pricing, credentials, finance data, private disputes, and confidential strategy.
AI-prepared customer emails, proposal wording, public copy, pricing, promises, reports, legal wording, HR material, or privacy statements should stay human-approved. Use AI Approval Gates for Business Automation for the stop points.
Record unusual access, incorrect outputs, staff questions, incidents, and permissions no longer needed. Remove stale accounts, old contractors, unused integrations, and abandoned exports.
The audit changes by tool because each tool reaches business information differently.
Copilot-style workflows make Microsoft 365 access hygiene more important. Check whether the right people can see the right files before AI helps them find those files faster.
Workspace AI may touch Drive, Docs, Sheets, Gmail, Meet, Calendar, and Chat. Define approved folders and draft-only outputs.
ChatGPT use often starts with copy-paste. Define what staff may paste, what must be redacted, which outputs need review, and when an approved business workflow is required.
A custom agent may connect to CRM, inbox, website, finance, project, or reporting systems. Give it the lowest useful permission first.
The audit should produce a practical cleanup list, not a vague warning.
Fix broad access on sensitive folders, archive old material, remove stale sharing links, and separate current source documents from outdated references.
AI should know which offer page, policy, SOP, template, FAQ, price note, or CRM field is current. If sources conflict, the workflow should flag uncertainty. Use the AI Data Readiness Checklist for NZ Businesses when source quality or ownership is unclear.
Review ex-staff, old contractors, unused integrations, shared inboxes, API keys, exports, and personal-account workarounds. Remove or narrow access that no longer supports a current workflow.
Decide who reviews AI output, what they check, where approval is recorded, and which exceptions require escalation.
The AI Agent Assessment turns the access audit into a decision for one real workflow.
The assessment names approved and excluded sources: folders, pages, CRM fields, exports, SOPs, reports, or inbox labels.
The agent might prepare summaries, checklists, draft replies, CRM notes, exception lists, briefs, or website recommendations.
Most first workflows should avoid live updates. If updates are useful, start with low-risk internal notes before changing customer records, public pages, pricing, or permissions.
Customer communication, pricing, legal wording, HR decisions, finance actions, privacy commitments, public publishing, access changes, and sensitive updates should stay human-approved.
It should include approved AI tools, sensitive sources, folder permissions, CRM access, inbox rules, prompt restrictions, read versus edit permissions, approval gates, logs, and access removal.
Copilot-style access depends on the Microsoft 365 environment, user permissions, configuration, and connected data. Review SharePoint, OneDrive, Teams, meetings, and file permissions before broad rollout.
Not by default. Staff need clear rules for customer, staff, finance, legal, credential, and confidential information. Sensitive context needs an approved workflow and review.
Read-only access is usually lower risk than edit, send, publish, delete, or admin access, but sensitive information can still be exposed through summaries or drafts.
The workflow owner should approve it with input from the data-source owner: often the business owner, operations lead, sales lead, finance owner, privacy owner, or IT provider.
Do not connect AI to company files as an experiment. Choose one workflow, name the sources, remove unnecessary access, define what AI may prepare, and keep consequences human-approved.
Book the $1,000 AI Agent Assessment to map one workflow’s data sources, permissions, approval gates, and safe next step before connecting AI to business systems.
