What's in this issue and why it matters

Somewhere in your organisation, someone is pasting data into an AI tool you didn't approve. They've been doing it for months. They're not trying to create a security incident. They're trying to get their job done.

This issue is about what shadow AI actually is, why the standard response — ban it, write a policy, move on — makes the problem worse, and what the smarter organisations are doing instead. Five reports, my honest take on each, and one tool you can use to run your own shadow AI audit before someone else does it for you.

Read time: 4 minutes.

The IT security audit started as routine.

Quarterly review, nothing flagged in advance, the kind of meeting that gets scheduled for ninety minutes and wraps in forty. The CISO sent the program director a spreadsheet the morning before. Standard network activity report. New tools detected on the corporate environment, grouped by department, sorted by volume of usage.

She opened it on her phone while walking to the meeting.

She stopped walking.

Seventeen AI tools had been detected running across the program in the previous ninety days. The program had four approved AI tools. She knew because she had spent three months getting them through procurement, security review, and legal sign-off. Three months, four tools.

Seventeen were running.

Some she recognized — ChatGPT, Gemini, tools she had seen people use casually and assumed were personal accounts on personal devices. But several she didn't recognize at all. And one of them, buried in row fourteen, was connected to a workflow that processed client data.

The CISO was already in the room when she arrived. He had the same spreadsheet on his laptop.

"How long has row fourteen been running?" she asked.

He looked at her. "Best guess? About six months."

Nobody in that meeting had authorized it. Nobody in that meeting had known about it. And nobody in that room could tell her, in that moment, exactly what data had gone where, or under what terms.

That is not a technology story. It is a governance story — and it is playing out in programs across every enterprise right now, whether the spreadsheet has been opened yet or not.

Here is the thing most AI governance policies get wrong: they treat shadow AI as a behavior problem.

Write a stronger policy. Send a reminder about approved tools. Remind people what the acceptable use guidelines say. Block the domains if you have to.

The behavior moves to a different tool. The spreadsheet gets longer next quarter.

Shadow AI is not primarily a security problem, though it creates security problems. It is not primarily a compliance problem, though it creates compliance problems. It is a signal. Every unauthorized tool running in your organization is a gap in your official tooling that someone got tired of waiting for. The employee pasting client data into an AI tool at 9pm is not a bad actor. They are a person with a deadline, an approved tool that couldn't do what they needed, and a workaround that worked.

That is information. Most organizations destroy it by treating it as misconduct.

The questions worth asking are not "who did this" and "how do we stop it." They are: what problem were they solving that the approved tools couldn't solve? How long had they been solving it this way? And who else is doing the same thing right now in a row you haven't found yet?

I have watched organizations respond to shadow AI audits in two ways.

The first response: shut it down, block the domains, send a company-wide email reminding employees of the policy, feel satisfied that the risk has been addressed. Three months later, the audit spreadsheet is longer.

The second response: call the person in row fourteen and ask them to walk you through what they built and why. Understand the gap before closing it. Use the finding to improve the official tooling, update the governance process, and get ahead of the next version before it shows up in next quarter's audit.

The second response is harder. It is also the only one that works.

Here are the five reports worth reading before your next governance conversation. I'll tell you what they get right, where they're careful, and what they leave out.

Understand what's happening:

1/ Microsoft, Work Trend Index 2025 🔗 https://www.microsoft.com/en-us/worklab/work-trend-index

What it says: 78% of AI users are bringing their own AI tools to work — tools their employer didn't provide and hasn't approved. The report frames this as enthusiasm for AI. Read the number again.

My take: Microsoft has an incentive to frame shadow AI as energy to be channelled rather than risk to be managed. Both are true. But "enthusiasm" is a comfortable word for what is, in many cases, unsecured data leaving the building in ways nobody has documented.

2/ IBM, Cost of a Data Breach Report 2025 🔗 https://www.ibm.com/reports/data-breach

What it says: The average cost of a data breach reached $4.88M in 2024. Shadow IT — including unauthorized AI tools — is consistently identified as one of the top contributing factors to breach severity and detection delay.

My take: The IBM report is the one to put in front of your CFO, not your IT team. Cost of a breach is a language every finance function understands. The conversation changes when the number is on the table.

Make sense of the debate:

3/ Gartner, Managing the Risks of Shadow AI 🔗 https://www.gartner.com/en/articles/what-is-shadow-ai

What it says: Shadow AI is the fastest-growing category of unsanctioned technology in enterprise environments. Gartner recommends a "govern, don't ban" approach — establishing clear policies and approved pathways rather than attempting to block tools outright.

My take: Gartner is right that banning doesn't work. But "govern, don't ban" still assumes governance frameworks exist and are functioning. In most of the programs I've seen, the governance is six months behind the AI adoption. You cannot govern what you don't yet know is running.

4/ Salesforce, State of AI Report 2025 🔗 https://www.salesforce.com/research/state-of-ai/

What it says: Employees using AI without employer knowledge report higher productivity than those using only approved tools. The gap is significant enough that Salesforce flags it as a competitive risk for organizations that over-restrict AI access.

My take: This is the finding that makes governance conversations uncomfortable. The unauthorized tools are often more capable than the approved ones. Employees know it. That is why they use them. The answer is not to lower the bar on approval. It is to move faster on getting the right tools approved before employees go around the process.

Act on it:

5/ ISACA, Shadow AI: Governance in the Age of Unsanctioned AI 🔗 https://www.isaca.org/resources/news-and-trends/industry-news/2024/shadow-ai-the-risks-of-unsanctioned-ai-use

What it says: Organizations need a shadow AI audit process, an AI tool registry, and a fast-track approval pathway for low-risk tools. Without the fast-track pathway, the audit process just measures the backlog of things people are waiting for permission to use.

My take: The fast-track pathway is the piece most organizations skip. They build the audit process and the registry but leave the approval process at the same speed it has always moved. Employees fill the gap. The spreadsheet gets longer. Read this one last, after you have decided what you are going to do about your own audit findings.

This is a shadow AI audit you can run in the next two weeks without involving IT security. It will not catch everything. It will tell you enough to start the right conversation.

The Shadow AI Conversation Audit:

Step one — ask, don't survey. Pick five people across different teams in your program. Not a survey, not a form. A fifteen-minute conversation. Ask them one question: "What AI tools do you use in a typical week — including anything you've found yourself, outside of what we've officially provided?"

Step two — listen for the gap. For each unauthorized tool they mention, ask: "What were you trying to do that the approved tool couldn't do?" Write down every answer. That list is your governance backlog.

Step three — identify the data risk. For each tool, ask: "What kind of information have you put into it?" You are not looking to discipline anyone. You are looking for row fourteen before it appears in an audit spreadsheet.

Step four — act on the signal. Take the most common gap from step two and ask: "If we solved this through an approved pathway in the next sixty days, would people stop using the workaround?" If yes, that is your next governance priority. If no, you have a deeper adoption problem worth understanding.

For program directors: do this before the quarterly IT audit, not after. Finding it yourself gives you the option to fix it. Finding it in the spreadsheet gives you a different kind of conversation.

The employee in row fourteen wasn't trying to create a risk. They were trying to do their job.

What does it tell you about your AI program that the workaround they found was more useful than anything you officially provided — and that nobody thought to tell you about it?