Shadow AI: 45% of your AI adoption is invisible to IT

Roughly 45% of enterprise AI adoption happens outside formal IT procurement. The Larridin Q1 2026 report put the number on the page; every CIO we've spoken to in the last six months has nodded and said it sounds about right or possibly low. Half of your business is already using AI for real work. Most of it isn't in any inventory. Most of it isn't covered by any policy. And most of the IT and security teams attempting to control it are going to lose, because the underlying demand is genuine.

The instinct to ban shadow AI doesn't work. Banning it is the response that produces the worst outcome: staff continue using the tools, but they hide it, which means you can no longer see what's being used, what data is going where, or which decisions are being affected. The right response is to channel demand, not block it. Build a sanctioned set of tools that's faster and easier than the unsanctioned ones, then make ongoing visibility cheap enough to maintain.

Why staff use shadow AI

Three reasons, in roughly equal measure, repeated across every business we've seen.

The official tool isn't good enough or doesn't exist. A marketer who needs to summarise a transcript and the only sanctioned tool is the corporate OneDrive that doesn't have AI summarisation will paste the transcript into ChatGPT. Not because they're rebellious. Because the work needs to ship by 4pm.

The sanctioned tool is too slow to access. A sales rep who needs to draft a proposal and the official AI assistant requires a procurement-led licence that takes six weeks will use a personal account in the meantime.

Nobody told them what was sanctioned. Most "shadow AI" is shadow only because there's no list of sanctioned tools, or the list is buried three levels into the intranet behind a search bar that doesn't return it.

Notice that none of these is a discipline problem. They're operational problems. The fix is operational, not punitive.

The risks worth taking seriously

Not every shadow AI use is dangerous. Most isn't. The cases that warrant active management are the ones with one or more of these signals:

• Customer or client data going into a public LLM with terms permitting training. Legal exposure, contractual exposure, and depending on the data, regulatory exposure under the Privacy Act amendments.
• Trade secrets, source code, or proprietary methodology being pasted into chat interfaces with no enterprise data isolation.
• Decisions being made or filtered by AI without human review, in domains where regulators expect human decision-making (lending, insurance, healthcare, employment).
• AI-generated content being shipped to customers without disclosure, in jurisdictions where disclosure may be required.

The remaining 80% of shadow AI use is staff using a chat interface to summarise a meeting, draft an email, brainstorm options, or check their reasoning. Low-risk. Often genuinely productive. Trying to manage it is a tax with no return.

The four-step operational fix

Step 1: build a sanctioned alternative that's actually good. This is the step most businesses skip and then are surprised when shadow AI continues. If your sanctioned AI tool requires three logins and produces worse output than what the staff member can get from a free tier of ChatGPT, you have not built a sanctioned alternative. You've built a checkbox. Real fix: enterprise ChatGPT or Claude or Gemini, paid licences, single sign-on, default availability for everyone who would benefit. Pay for the seats. Cheaper than the alternative.

Step 2: publish the approved-tool list, prominently. One page on the intranet. Five tools. What each is approved for. What each is not approved for. Updated quarterly. Linked from every onboarding deck and every IT communication. Most shadow AI is shadow because nobody knows what isn't.

Step 3: deploy a data-loss prevention layer at the LLM gateway. For the use cases you really do want to prevent (customer PII into a public LLM, source code into an unsanctioned chat), use a DLP product that intercepts at the network or browser layer. This is the small fraction of shadow AI worth actively blocking, and it should block at the technology layer rather than at the policy layer. Policy alone never works.

Step 4: run quarterly amnesty surveys. Anonymous, non-punitive: which AI tools are you using that aren't on the approved list, and what for? The first round will surprise you. The second round, six months later, will tell you what's becoming popular before it becomes a problem. Make this culturally safe by being explicit that nobody gets fired for using an unsanctioned tool, only for hiding incidents.

Shadow AI exists because the official tools didn't ship. Punishing the symptom doesn't make the official tools ship faster, it just makes the shadow harder to see. Channel demand. Build the better path. Block the genuinely dangerous fraction at the technology layer, not the policy layer.

What this catches that policy-alone misses

Three specific incident types we've seen mid-market businesses experience in the last twelve months that lightweight DLP plus an approved-tool list would have prevented:

• A finance analyst pasted a working capital model into a public LLM to ask for help formatting it. The model included client names, contract values, and supplier terms. The LLM provider's terms permitted training on the input.
• A sales engineer used an unsanctioned coding assistant to debug a customer's deployment. The repository was set to public by default in the assistant's web interface. Customer source code was indexable by search engines for 36 hours before the team noticed.
• An HR coordinator used a CV screening tool that automatically scored candidates against a job description. The tool had no audit trail and no demographic-bias monitoring. A rejected candidate later asked, under privacy law, what factors had been used in the decision. There was no answer to give.

None of these were intentional misuse. All three would have been prevented by an approved-tool list plus a basic DLP gateway.

The single most useful first move

Conduct an inventory. Do it once, properly, before you build any policies or buy any tools. Walk every department. Ask "what AI tools are you using, including ones on personal accounts, including ones you signed up for via free trial." You will find double what you expected. The map of what you find is the basis for everything else: what to sanction, what to block, what to procure properly, and what to leave alone.

The businesses that get shadow AI right start with curiosity rather than control. The ones that get it wrong start with prohibition. The first approach produces a manageable footprint inside two quarters. The second produces an underground footprint that's unmanageable forever.