For every company size and skill level
How to adopt AI tools securely
Your team is already using AI. This guide shows how to make it safe without slowing anyone down, in eight steps that work whether you run a 20 person company or a 2,000 person one. Each step has the plain English version, the details for whoever runs IT, and a free way to check it.
In short: give people company accounts, lock down sign in, training and sharing, limit what data AI can reach, write a one page policy, find the AI already in use, secure the agents on laptops, train people, and plan for incidents.
By Jon Rose, Founder & Managing Partner, IOmergent.
The eight steps at a glance
- 1Pick the right kind of accountGive people a company workspace, not personal plans. In a company workspace your admins are in control, and on business plans the vendor doesn't train on your data.
- 2Lock down the workspaceTurn on company sign in, make sure the vendor isn't training on your data, and decide how long chats are kept and who can share them.
- 3Decide what company data AI can reachConnectors let AI read your email, files and code. Turn on only the ones you need, and decide who approves new ones.
- 4Write a one page acceptable use policyTell people what's allowed, what isn't, and who to ask. One page is enough to start.
- 5Find the AI already in usePeople are probably already using AI with personal accounts, and some AI apps may already have access to company email or files.
- 6Secure the AI agents on people's computersCoding agents and desktop AI apps can read files, run commands and use connected tools. Their settings matter more than any chat app's.
- 7Roll it out and train peopleStart with one team, show everyone the safe way to use the tools, then expand. Review once a month.
- 8Plan for when something goes wrongKnow how to cut off access fast, and who to call, before you need it.
Step 1. Pick the right kind of account
In plain words: Give people a company workspace, not personal plans. In a company workspace your admins are in control, and on business plans the vendor doesn't train on your data.
Do this
- Choose the business version of each tool your team uses: Claude Team or Enterprise, ChatGPT Business or Enterprise, Gemini in Google Workspace, Microsoft 365 Copilot.
- Verify your company domain with each vendor, so work sign ups land in the company workspace.
- Ask people to move work out of personal accounts. A work email on a personal plan is still a personal account.
- Decide who approves new AI tools, and budget per seat.
Check it: in your AI assistant with the free AI Hardening Check, ask “Run the IOmergent AI Hardening Check and tell me which account levels we have.”
Go deeper: Which AI account is which?
Step 2. Lock down the workspace
In plain words: Turn on company sign in, make sure the vendor isn't training on your data, and decide how long chats are kept and who can share them.
Do this
- Require single sign on (SSO) and multi factor authentication.
- Confirm training on your data is off. It's off on business, enterprise and API plans; personal plans need each person to turn it off.
- Set how long chats and files are kept, and turn off sharing outside the company.
- Keep the list of owners and admins short.
Check it: in your AI assistant with the free AI Hardening Check, ask “Run the IOmergent AI Hardening Check for our ChatGPT workspace.”
Go deeper: ChatGPT security settings · Claude security settings · Gemini security settings · Microsoft 365 Copilot security settings
Step 3. Decide what company data AI can reach
In plain words: Connectors let AI read your email, files and code. Turn on only the ones you need, and decide who approves new ones.
Do this
- List the connectors each tool can use (Google Drive, Microsoft 365, Slack, GitHub) and turn off the rest.
- Prefer read only access, and require admin approval for new connectors and apps.
- In Microsoft 365, fix oversharing first: Copilot answers from everything a person can already open.
- Decide who can build custom GPTs, Gems or agents, and who they can share them with.
Check it: in your AI assistant with the free AI Hardening Check, ask “Can we safely connect Claude to our Google Drive?”
Go deeper: Workforce AI apps architecture · AI security for CTOs: 50 questions
Step 4. Write a one page acceptable use policy
In plain words: Tell people what's allowed, what isn't, and who to ask. One page is enough to start.
Do this
- List the approved tools and accounts.
- Say what must never go into AI unless the tool is approved for it: customer personal data, passwords and keys, anything under NDA.
- Ask people to check AI output before it goes to a customer or into production.
- Name who approves new AI tools and connectors.
Check it: in your AI assistant with the free AI Hardening Check, ask “Run the IOmergent AI Hardening Check and list the gaps our AI policy should cover.”
Go deeper: AI governance and security cheatsheet
Step 5. Find the AI already in use
In plain words: People are probably already using AI with personal accounts, and some AI apps may already have access to company email or files.
Do this
- Ask each team which AI tools they use, and make it safe to answer honestly.
- Check which third party apps hold access to Google Workspace, Microsoft 365, GitHub and Slack, and remove the ones you don't approve.
- Block new personal AI accounts on your domain where the vendor allows it, for example Claude's 'Restrict organization creation'.
- Look at expense reports for AI subscriptions.
Check it: in your AI assistant with the free AI Hardening Check, ask “Run the IOmergent AI Hardening Check, including the personal accounts people use for work.”
Go deeper: AI Security Assessment (finds AI apps with access to your data)
Step 6. Secure the AI agents on people's computers
In plain words: Coding agents and desktop AI apps can read files, run commands and use connected tools. Their settings matter more than any chat app's.
Do this
- Keep approval prompts on. No 'skip permissions' or 'YOLO' modes.
- Allow only trusted MCP servers, plugins and extensions, pinned to a version.
- Keep secrets out of agent config files; use environment variables or a secrets manager.
- On company machines, enforce the settings centrally with managed settings.
Check it: in your AI assistant with the free AI Hardening Check, ask “Run the IOmergent AI Hardening Check quick local scan on this machine.”
Go deeper: Claude security settings · Cursor security settings · Developer AI architecture
Step 7. Roll it out and train people
In plain words: Start with one team, show everyone the safe way to use the tools, then expand. Review once a month.
Do this
- Pilot with a small group and write down what works.
- Run a 30 minute training: the approved tools, what not to paste, how to spot a wrong answer.
- Review usage, spend and new connectors every month.
- Run the check again after any plan change and every 90 days.
Check it: in your AI assistant with the free AI Hardening Check, ask “Run the IOmergent AI Hardening Check again and compare with last time.”
Go deeper: AI security hardening checklist
Step 8. Plan for when something goes wrong
In plain words: Know how to cut off access fast, and who to call, before you need it.
Do this
- Know how to revoke an AI app's access and remove a connector.
- Rotate any API key or token that was pasted into a chat or a config file.
- Turn on and export audit logs where your plan has them, and know who can read them.
- Add AI to your incident response plan and practice it once.
Check it: in your AI assistant with the free AI Hardening Check, ask “How do I revoke an AI app's access to our Google Workspace?”
Go deeper: Incident response tabletop · AI security for CTOs: 50 questions
Common questions
- What is the first step to adopt AI tools securely?
- Give people a company workspace, not personal plans. In a company workspace your admins are in control, and on business plans the vendor doesn't train on your data.
- Do we need an AI acceptable use policy?
- Yes. Tell people what's allowed, what isn't, and who to ask. One page is enough to start.
- How do we find the AI tools people already use?
- People are probably already using AI with personal accounts, and some AI apps may already have access to company email or files. Check which third party apps hold access to Google Workspace, Microsoft 365, GitHub and Slack, and remove the ones you don't approve.
- How do we secure AI coding agents and desktop AI apps?
- Coding agents and desktop AI apps can read files, run commands and use connected tools. Their settings matter more than any chat app's. Keep approval prompts on. No 'skip permissions' or 'YOLO' modes.
Want an expert review?
The AI Security Assessment looks at what a self-check can't: which AI apps hold access to your email, files and code, and whether your settings really match what was reported.
See the AI Security Assessment