Shadow AI: the unmanaged AI tools quietly connected to your business
Jordan Gall, Head of Cybersecurity, Buchanan Technology
What it is
Shadow AI is what happens when your team starts using AI tools without you knowing. A marketing assistant signs up for a generative AI writing tool and pastes in three months of client copy. A bookkeeper uses a free AI spreadsheet helper to "speed up" reconciliation and uploads six months of transaction data. A consultant drops a client proposal into a public chatbot to get feedback on the tone.
None of them did it with bad intent. Each one just wanted to get the job done faster. And none of them told you, because in their minds it felt like using a new spell-check or a slightly smarter search bar.
But the tool is now processing your data, your clients' data, and your intellectual property. Through a third party's server, under a third party's terms, with your business having zero visibility, zero control, and zero contractual protection.
This is not a hypothetical. A 2025 survey by Deloitte found that over 70 percent of enterprise employees had used an unapproved AI tool, and the same pattern is now showing up in small businesses where there is no IT team to catch it. The difference is that at the enterprise, a security team will eventually notice anomalous data flows. At a small business, nobody is looking.
Why you need to care
Your client data is going somewhere you cannot see. When staff paste client names, financials, legal briefs, or personal health information into a public AI tool, that data may be stored, logged, or used for model training depending on the tool's terms of use. Your Privacy Act obligations and any contractual confidentiality duties to your clients do not care where the data ended up. They care that it left your control.
Your intellectual property is no longer private. Trade secrets, proprietary processes, pricing models, and unpublished work product all lose their protection the moment they are fed to a third-party system. If your competitive edge is in how you structure a service or price a project, and that gets processed by a model whose training data policies you never agreed to, you have a problem that is very hard to walk back.
Compliance exposure multiplies. If you operate in a regulated space, be it financial advice, legal, health, or education, sending controlled or sensitive data through an unauthorised channel can breach your professional obligations, your cyber insurance policy, and in some cases your licence conditions. Insurers are already writing AI usage exclusions into new policies, and the fine print increasingly requires you to demonstrate you have governance over your AI tools.
One click can go wrong in public. The riskiest shadow AI tools are not the passive chatbots. They are the active ones: AI agents that can draft emails, book meetings, update CRM records, or trigger workflows. When an agent acts on your behalf using client data it was never authorised to see, the resulting error is yours, not the vendor's.
What you can do about it
You do not need to ban AI. That is the same as telling people not to use the internet in 2003. But you do need to know what is in the building and what it can see.
1. Find out what is actually in use. Have a straight conversation with your team. "What AI tools have you signed up for, and what have you put in them?" No blame, no punishment. You are building the inventory, not running a witch hunt. Common offenders include: chatbots (ChatGPT, Claude, Gemini), AI writing assistants, AI meeting note-takers, AI coding helpers, and AI image or design tools.
2. Draw a line between approved and unapproved. Decide, as a business, which tools are allowed to process business or client data and which are not. The rule does not need to be complicated. A simple one-page policy that says "we use X for Y, and nothing else touches client data" is enough to start. The key is that the rule exists and everyone knows it.
3. Lock down what you can technically control. If you run Microsoft 365, check which AI features are turned on in your tenant. Microsoft Copilot, the AI summary features in Outlook, and the generative features in Teams all process data within the Microsoft cloud, but they still process it. If you have not configured the data boundaries, retention, and access controls for those features, you have a shadow AI problem inside a licensed product. This is a configuration and governance task, not a purchase decision.
4. Put the guardrails in the platform, not just the policy. A policy that says "do not paste client data into AI" is a hope, not a control. A control is a DLP rule that blocks or alerts when sensitive data is shared with an external app. A control is a Conditional Access policy that limits which tools a user can connect to from a company device. A control is a tenant-level switch that turns off the features your team does not need. These are all doable within your existing stack.
5. Make the approved path the easy path. If your team needs AI help with writing, summarising, or analysis, give them a sanctioned tool that they are allowed to use with business data. The moment there is an approved alternative that is as fast as the unapproved one, the shadow tools lose their appeal. Most small businesses do not need a complex AI governance platform. They need one or two tools, configured correctly, with a clear "this is what you can and cannot put in it" instruction.
6. Review your AI exposure quarterly. Tools change, features switch on by default, and staff roles evolve. A 15-minute check each quarter, "what is new, what is unused, what is no longer needed", keeps your footprint honest without a dedicated governance team.
The bigger point
Shadow AI is not a technology problem. It is a visibility problem. Your team is solving real work challenges with real tools, and those tools are processing your business data in ways you have not authorised, reviewed, or insured.
The fix is not to scare people into stopping. It is to make the right tools visible, the wrong ones invisible, and the rules simple enough that a busy operator can follow them without thinking.
If you want to know what AI is quietly running in your Microsoft 365 tenant, or you want a one-page AI usage policy your team will actually follow, get in touch. A short conversation today is a lot cheaper than a data breach investigation next month.