Shadow AI: the risks for small businesses and how to manage them

What shadow AI looks like in a small firm

In a large company, shadow AI might mean a department buying its own AI platform. In a 25-person firm it is quieter. A bookkeeper rewrites client emails in a personal ChatGPT account. A project manager installs a free meeting-notes app that joins every Zoom call. A recruiter uses a resume-ranking website. A partner uploads a contract to a free PDF summarizer.

None of these people are trying to cause harm. They found something that saved time, and the firm had not offered an alternative.

The common thread is that each tool solved a real, recurring annoyance faster than the official process did.

The real risks

Shadow AI matters because of what the business loses control of. The main risks:

Of these, the personal account problem is usually the largest in practice. It is invisible to the owner, it grows quietly as more staff find the same shortcut, and it creates a record of client work that the firm does not control. Fixing it is mostly a matter of buying business seats and asking people to move their work across.

• Client data on personal accounts: when an employee leaves, the conversations, uploads and any shared links go with them, and you cannot review or delete them.
• Training defaults: some consumer plans allow content to be used for model improvement unless the user opts out.
• Unknown vendors: free apps and browser extensions may have unclear terms about storage, sharing and resale of data.
• Unreviewed output: AI-drafted text going to clients without anyone checking facts or numbers.
• Regulatory exposure: unprotected health information, taxpayer data or privileged material processed somewhere your obligations do not allow.
• Decisions about people: tools that rank job applicants or tenants can create duties under laws such as Colorado's SB 26-189 from 2027.

A cautionary example

One of the best-known cases involved a large company, but the lesson applies to any size. In 2023 Samsung restricted generative AI tools on company devices after engineers reportedly pasted internal source code and meeting notes into ChatGPT while trying to fix problems. The engineers were doing their jobs. The gap was that the company had not yet set rules or offered a controlled option.

For a small firm, the equivalent might be a client's tax documents or a draft settlement in a personal account. The damage is not always visible, which is why it is worth checking.

The same pattern shows up in small firms in quieter ways: a free transcription service used for a sensitive client call, or a staff member uploading a spreadsheet of client contacts to get help with a mail merge. Each is a reasonable impulse with an avoidable risk.

How to find shadow AI without a witch hunt

Start with amnesty. Tell staff you are choosing approved AI tools and want to know what people already use and like. Make it clear nobody is in trouble. You will get far better information than from monitoring alone. Then check a few technical sources:

Expect some apps that nobody remembers installing. Many meeting tools and writing assistants ask for broad access to email and calendars at sign-up, and people click through. Revoking those connections is quick and rarely disrupts anyone.

• A short survey: which AI tools, for which tasks, on which account, and with what kind of data.
• Sign-in reports in Microsoft 365 or Google Workspace showing third-party apps connected with company accounts.
• Browser extensions on company laptops, especially writing assistants and meeting recorders.
• Calendar invites that include automated note-taker bots.
• Expense reports and card statements for small AI subscriptions.

Replacing shadow AI with approved tools

The goal is to make the approved option easier than the unapproved one. Pick one general assistant on a business plan, such as Microsoft 365 Copilot, ChatGPT Business or Claude Team, and give seats to the people who were already using AI. Choose one meeting-notes approach and switch the others off. Disconnect unknown apps from company accounts.

Write a one-page policy with an approved tool list, three data tiers and a review rule, and explain how to request a new tool. A quick, friendly approval path prevents the next round of shadow AI. If requests take weeks, people will go around it.

Pay attention to why people chose the tools they did. If three people independently found the same transcription app, there is a real need for meeting notes. If a recruiter used a ranking site, the hiring process may be overloaded. Shadow AI is a cheap form of research into where your workflows hurt.

If sensitive data has already gone into an unapproved tool

It is common to discover, during the inventory, that client information already went into a personal account or free app. Respond calmly and consistently. A simple sequence:

Set a target for how quickly tool requests get an answer, such as within one week, and tell staff what it is.

• Find out what was shared, in which tool, on which account, and roughly when.
• Ask the employee to delete the relevant conversations and files, and to turn off model-training settings where the tool offers them. Keep a note that this was done.
• Disconnect any third-party app from company email, calendar or file storage.
• Check whether the data falls under specific obligations, such as health information, taxpayer data, privileged material or a client contract with confidentiality terms. If it does, involve counsel or your compliance lead before deciding on any notification.
• Record the incident in your AI register and adjust the policy or approved tools so the same need is met safely next time.

Keeping it under control

Shadow AI comes back as new tools appear. A light routine keeps it in check: review connected apps and extensions quarterly, ask about AI tools in team meetings, include the AI policy in new-hire onboarding, and have departing employees confirm they have not kept client data in personal AI accounts.

Track a few signals: requests for new tools, the share of staff using approved seats, and any incidents. A rising request count is good news. It means people are asking rather than going around you.

InTheory often starts AI policy work with exactly this kind of no-blame inventory. It usually turns up a few useful tools worth approving, along with a few to switch off.