Your team is already using personal AI agents on company accounts: what to put in your policy now
If someone on your team pays for an AI agent with their own card and connects it to their work inbox, your company data now sits under a contract you never signed, in a tool your admin cannot see or switch off. Do not ban it and hope. Put one clear rule in your shadow AI policy this month: company accounts connect only to AI tools on a plan the company controls. Then make the approved route faster than the workaround.
Prices and plans checked on 16 November 2026.
The short answer
A personal AI agent is not a chatbot someone pastes text into. It holds a standing connection to email, calendar, files and chat, and keeps working after the laptop closes. So four things belong in your AI policy now:
One connection rule. No personal-plan agent on a company account. A fast request route. Approved AI tools named, and requests for new AI tools answered within days. Limits on data and actions. What never goes in, and what an agent may never do without a person. An amnesty and a switch. Thirty days to declare what is connected, then admin settings that stop new connections without approval.
What is shadow AI, and why are personal agents different?
Shadow AI refers to AI tools that people use for work without the company approving or knowing about them. It is the younger sibling of shadow IT, where staff sign up for software on their own. Most companies meet it first as a marketing team drafting copy in a personal chatbot, or an analyst pasting customer data into one of the generative AI tools. Shadow AI usage grows wherever AI adoption runs ahead of the rules.
Personal agents raise the stakes. Unlike traditional shadow IT, an agent reads everything it is connected to, remembers it, and can act:
It sees the whole mailbox, not one pasted paragraph. Client data, financial data, contracts and HR threads are all in reach from the first minute. It reads content written by strangers. Any email can carry hidden instructions aimed at the agent. OpenAI's security page for dots says its models are trained to "resist malicious instructions hidden in emails, files, or web pages". Resist is not the same as immune. It can act in your company's name. Send, reply, book, pay. A wrong email to a client is more than data exposure. It is an incident with your logo on it.
Not all shadow AI is bad news. It shows you where work feels too slow. The problem is that nobody in charge can see it.
Which personal AI agents can your people buy today?
Agent Who can buy it Can your admin see or control it? Claude Duet (Anthropic) Research preview by waitlist in the US, on the personal Pro, Max and Ultra plans. Pro is $20 billed monthly No. The help page says it "isn't available on Team or Enterprise plans yet" ChatGPT dots (OpenAI) Pro plans at $100, $200 or $500 a month, for adults outside the European Economic Area, the UK and Switzerland. Business Premium seats in supported regions. Eligible Enterprise workspaces too Only in a workspace. The Enterprise admin guide covers roles, allowed apps, actions and logs. A personal Pro plan has no company admin Muse (Meta) The US and Canada, and Singapore since 27 October 2026. "Free for most of what people need", with subscription plans. Meta's pages give no prices Meta's small business announcement mentions no admin or team controls OpenClaw and Hermes Agent Anyone. You download and run them yourself, on a laptop or a server No. The only control is the machine it runs on
The hosted agents are sold to individuals before they are sold to companies. Claude Duet connects to Gmail, Google Drive, Slack, QuickBooks and more, and there is no business plan for it yet.
If your team is in Greece or elsewhere in the EU, none of the three hosted agents is offered to them on a personal plan today, going by the pages I read. That buys you time, not safety. Colleagues in the US, Canada or Singapore can sign up, and self-hosted agents work anywhere.
I compare the five products in my guide to always-on AI agents.
What are the AI risks on a personal plan?
Four gaps, and none needs a hacker. This is the shadow AI risk in plain terms: sensitive data leaves, and nobody notices.
The contract is with your employee, not with you. Claude's Team plan lists "No model training on your content by default", and ChatGPT Business says "No training on your business data by default". Personal accounts are different. OpenAI's policy page says of its services for individuals: "we may use your content to train our models", unless the person opts out. On Claude's personal plans, Anthropic's privacy page ties training to a Model Improvement setting in the user's own privacy settings. Either way the switch belongs to the employee. And if the mailbox holds personal data about clients or staff, you have no data processing agreement with that vendor.
Nobody can see it. No audit log your IT teams or security teams can read, no list of what it touched, no record when a regulator or a client asks.
The agent remembers after it is disconnected. Anthropic's help page says disconnecting an app stops access "from that point on", "but memory of past information from earlier conversations is kept". OpenAI's security page says much the same about revoking a dot's access to an app: "The dot may still have information it retrieved earlier." When that employee leaves, the memory leaves with them.
Approvals wear off. The hosted agents have approval steps for sensitive actions. But Anthropic's own safety note admits the weak point: "Approval fatigue is a serious issue, because too many requests teach people to tap through them".
Are self-hosted agents safer, because the data stays with us?
Only if "us" means the company. A self-hosted agent on one person's laptop is the least visible shadow AI: no vendor, no invoice, no admin console.
OpenClaw's trust model supports "one trust boundary per gateway" and warns: "Everyone who can message a tool-enabled agent shares that agent's delegated tool authority." Put that agent in a company chat channel and everyone there can steer it. Hermes Agent has three approval modes, and its security page says of the third: "Setting approvals.mode: off disables all safety prompts."
So self-hosted agents are an IT project, on a machine IT owns, with approvals left on.
What should a shadow AI policy say about agents?
Write the rules as a table, not an essay. If you have an AI acceptable use policy, add a section on agents. If not, this is the first page of your AI governance framework.
Section What to write Scope Any AI tool or AI service that connects to a company account or acts for you: agents, AI writing assistants, meeting note-takers, browser extensions, and embedded AI features in software we already use The connection rule Company email, calendar, files, chat, CRM and finance systems connect only to AI tools on the approved list, on a plan the company pays for and administers Personal accounts Personal AI subscriptions are fine for personal life and public information. Not for company accounts, and not for sensitive company data Sensitive data that never goes in Customer data and client data beyond what the approved tool is cleared for, financial data before it is public, HR and health information, passwords and keys, trade secrets Actions An agent may read, summarise, draft and prepare. A named person approves anything that sends, pays, signs or deletes, and stays accountable for it Self-hosted agents Only with IT, on a company machine, with approvals on New AI platforms and tools Ask the owner. After a short risk assessment you get an answer within five working days: approved, approved with limits, or no with a reason Amnesty and leaving Declare within 30 days what you have already connected. No consequences. We help you disconnect it and reset its memory. Every connection is revoked on your last day Owner and review One named owner. Reviewed every quarter, because the AI landscape changes monthly
The amnesty comes before enforcement. You cannot govern what people hide. Unauthorized AI use that has been declared is a list you can work through. Punish it and it goes back underground.
The request route has a deadline. A blanket ban on unapproved AI tools does not end AI usage. It moves it to personal phones. Five working days is my suggestion.
How do you find shadow AI tools that are already connected?
You do not need a cloud access security broker or a tool that inspects network traffic to start. The office suites already let you monitor AI usage where it matters most: the connection to the mailbox.
Google Workspace. In the Admin console, under Security, then Access and data control, then API controls, your admin can mark each third-party AI app as trusted, limited or blocked, restrict high-risk Gmail and Drive access, and choose what happens with "Unconfigured third-party apps". (Google's steps.) When users ask for a blocked app, it lands in a list called Apps pending review. Unauthorized AI tools that people tried to connect show up there. It is your shadow AI discovery tool, for free. Microsoft 365. Microsoft's documentation is blunt: "By default, all users are allowed to consent to applications for permissions that don't require administrator consent." Microsoft recommends allowing consent only for apps from verified publishers. The setting is in Entra under Enterprise apps, then Consent and permissions. Turn on the admin consent workflow as well, so a blocked request reaches a reviewer. ChatGPT workspaces. If you are on Enterprise, check one date. The admin guide says: "Starting November 18, 2026, members have access by default unless the workspace opts out". Decide before then.
These settings will not catch a self-hosted agent that logs in with a saved password. For that you rely on the policy and the amnesty.
Where do AI governance and the EU AI Act come in?
One duty already applies to companies that use AI systems. Article 4 of the EU AI Act requires providers and deployers to ensure a sufficient level of AI literacy among their staff, and it has applied since 2 February 2025, as Covington's summary of the Commission's guidance sets out. A one-page policy people have read, plus a short session on the risks above, is a sensible part of meeting it. I am not a lawyer, so check with yours. The wider picture is in AI governance and the EU AI Act.
What I would do
I brought AI into the processes and product of a 200-person fintech. People do not route around rules. They route around slowness.
In this order:
This week: send a friendly note. "If you have connected any AI tool to a company account, tell us by the end of the month. Nobody is in trouble." This week: ask your admin to open the app access settings and list the third-party apps connected to company accounts. Within two weeks: publish the one-page policy above and name its owner. Within a month: give people one approved assistant on a business plan. Claude Team is $20 a seat per month billed annually, or $25 monthly. ChatGPT Business is $20 a user per month billed annually, or $25 monthly. Both state that they do not train on your business data by default. Then: switch the admin setting so new connections need approval.
When would I change this?
If you handle health, legal or regulated financial data, switch the admin setting first and explain second. If you are under 15 people with no IT person, start with the policy, the amnesty and one approved plan.
What to check before you decide
Do you know which third-party apps are connected to company mail today? Can users connect a new app without anyone approving it? Is there an approved AI assistant on a business plan, so the rule has an alternative? Does the policy say plainly: no personal-plan agents on company accounts? Is there a named owner and a promised response time for requests? Have you offered an amnesty before any enforcement?
Questions leaders ask me
What is the shadow AI problem?
People use AI tools for work that the company has not approved and cannot see. With chatbots the risk was sensitive data pasted into a prompt. With agents it is a standing connection to the mailbox.
Why is shadow AI not permitted?
Because the company carries the responsibility without having the controls. On a personal plan there is no company contract, no admin, no log, and no way to switch the tool off when someone leaves.
What are the risks of shadow AI?
Data leaks, actions nobody approved, and no record of either. Sensitive data reaches external AI services under terms the company never agreed to, and on some personal plans the vendor may use it to train AI models unless the user opts out.
Want to see where agents fit in your wider AI plan? Take the free AI Growth GPS and see where you stand.