Shadow AI Is Already in Your Portfolio. Governance Has to Start There.

Key Takeaways

• Shadow AI is already running. In Lenovo's 2026 survey of 6,000 employees, between a fifth and a third said they use AI outside IT oversight. Any governance plan that starts from zero is starting from a fiction.
• Bans push usage out of sight. People use unapproved tools because the work gets done faster. Take the tool away without a replacement and the usage just moves to a personal phone.
• Inventory comes before policy. Find what people actually use, with what data, for which tasks. Then write the rules.
• Triage by data sensitivity and business value. Every tool you find gets one of four answers: sanction it, replace it, monitor it, or shut it down.
• Role-based permissions make the policy real. A policy nobody enforces is shelfware. Access tied to roles is what makes the approved path the easy one.

Most AI governance programs start the same way. A committee forms, someone drafts an acceptable use policy, the group picks an approved tool, and the rollout plan assumes the company is starting from a blank page.

It isn't. The page was filled in months ago, by employees, one browser tab at a time.

Lenovo's 2026 survey of 6,000 employees found that more than 70% use AI at work every week, and that between one-fifth and one-third use it outside the oversight of IT. The same survey found 22% say their employer does not give them AI tools at all. So they brought their own.

That second number explains the first. People reach for unsanctioned AI because it helps them get the work done, and nobody gave them a better option.

For a PE firm, multiply that across a portfolio. Ten or fifteen companies, each at a different stage of AI maturity, each with its own mix of customer data, financials, contracts and employee records. Some have an AI policy. Most have a draft. All of them have people pasting work into chat windows right now.

The way I think about it: shadow AI is a demand signal. It tells you exactly where your teams want AI, which tasks they think are worth it, and how far ahead of IT they already are. Governance that ignores that signal fails on day one. Governance that starts from it has a head start.

Your Portfolio Already Has an AI Program. Nobody Approved It.

Shadow AI is any AI tool or feature used for work without the knowledge or approval of the people responsible for security and compliance. Personal ChatGPT accounts. Free browser extensions that summarize email. AI features switched on inside SaaS tools nobody re-reviewed. A note-taker bot that joins every sales call.

None of this is new behavior. It is the same thing that happened with Dropbox and personal Gmail a decade ago. The difference is what goes into the tool. People paste the actual work into AI: the customer list, the draft board deck, the contract redline, the payroll export.

Microsoft and LinkedIn saw this early. In the 2024 Work Trend Index, 75% of knowledge workers said they use AI at work, and 78% of those users said they bring their own tools. That data is two years old now. I would be surprised if the number went down.

Assume usage before you measure it

Start every governance conversation with a simple assumption: people at this company already use AI, with company data, in tools you have not reviewed. You will almost always be right. And it changes the first question from "should we allow AI?" to "what is already happening, and what do we do about it?"

A Ban Pushes Usage Out of Sight

The instinct when leadership discovers shadow AI is to block it. Add the domains to the firewall, send a stern email, done.

It feels decisive. It mostly fails.

Blocking a site on the corporate network does nothing about the phone in someone's pocket. The work still needs to get done, and the employee who saved two hours a day with a chat tool is not going to give those hours back quietly. They will just stop telling you. Now you have the same risk, less visibility, and a team that has learned governance means "no."

The breach data says the cost of that blind spot is real. IBM's 2025 Cost of a Data Breach Report found that one in five organizations reported a breach due to shadow AI, and that organizations with high levels of shadow AI paid an average of $670,000 more per breach than those with low or no shadow AI. Only 37% had policies to manage AI or detect shadow AI.

Make the approved path the easy path

The fix is a better option, offered fast. If the sanctioned tool is slower to get, harder to use, or missing the features people rely on, shadow usage wins. Every time. Give teams an enterprise AI environment that is at least as good as what they found on their own, with single sign-on and their real work files connected, and most of them will switch without being told twice.

Inventory Before Policy

Most AI policies are written in a conference room from first principles. They describe an idealized company. Then they meet the real one and nobody follows them.

Flip the order. Before anyone writes a rule, find out what is in use. Which tools, which teams, which tasks, and which data is going in. That inventory becomes the input to the policy, and the policy ends up describing how people actually work.

There are a few places to look. Expense reports and corporate card statements show paid AI subscriptions. Browser and network logs show which AI domains get traffic. SaaS admin consoles show AI features that vendors switched on by default. And the most useful source is the simplest: ask people.

Run an amnesty-style usage survey

Tell every team, plainly, that nobody gets in trouble for what they report. Ask what AI tools they use, what for, and what kind of data goes in. You will get more honest answers from a two-week amnesty than from a year of log analysis. You will also learn which workflows people value most, which is the best possible input for deciding where to invest next.

Sort What You Find by Data and Value

An inventory with 40 tools on it is not a plan. You need a fast way to decide what happens to each one.

Two questions do most of the work. How sensitive is the data going into this tool? And how much real value is the team getting from it? That gives you four answers.

The "replace it" bucket is where the risk and the opportunity both live. These are the workflows where people found something that genuinely helps, and they are doing it with data you cannot afford to leak. That is your rollout priority list, written for you by your own employees.

Decide in weeks

Speed matters here more than precision. Every week a high-sensitivity workflow stays in a personal account is another week of exposure. Get a rough sort done, act on the "replace" and "shut down" buckets first, and refine the rest later. A good-enough decision this month beats a perfect one next year.

Role-Based Permissions Turn a Policy Into Practice

Here is where most governance programs quietly die. The policy gets written, approved, and posted to the intranet. Then nothing in the actual tools changes. Everyone still has the same access to everything. The policy says "don't put customer PII into AI," and the platform happily accepts customer PII from anyone who logs in.

A policy without enforcement is shelfware. I'll say that flat.

What makes governance real is access control inside the AI environment itself. Who can use which models. Which connectors each role can reach, and whether those connectors can only read or can also write. Which agents can take actions on their own and which need a human to approve. When finance can query the ERP connector and marketing cannot, you don't need a paragraph in a PDF to explain the rule. The platform is the rule.

This is also what lets you say yes more often. Once access is scoped by role, you can open AI up to more teams with less worry, because each one only reaches what its job requires.

Map permissions to the roles you already have

Don't invent a new permission model for AI. Start from the identity system and org chart the company already runs on, connect the AI platform to single sign-on, and let access follow the roles that exist. When someone changes jobs or leaves, their AI access should change with them automatically. If it depends on someone remembering to update a spreadsheet, it will drift.

Portfolio Governance Needs One Floor and Local Judgment

For a PE firm, the hard part is scale. You can't run a bespoke governance program at every company, and you shouldn't impose one rigid policy on a SaaS business and an industrial distributor alike.

What works better is a common floor set at the fund level, with each company building on top of it. The floor is short: an inventory every company completes, a standard triage, single sign-on on the approved AI tools, role-based access to connectors, and audit logs that someone actually reviews. Above that floor, each management team decides what fits its data, its customers and its regulators.

I'm honest that this part is still unsettled. How much should a fund standardize the AI stack itself across portfolio companies? How do you compare governance maturity across businesses that look nothing alike? I don't think anyone has a clean answer yet. But a shared floor gives the operating team one language to use across the portfolio, and that alone is a big step up from fifteen separate drafts.

Set the floor at the fund, let each company build up

Pick five or six non-negotiable controls and write them down. Make them checkable, so an operating partner can ask "is SSO enforced on the AI tools?" and get a yes or no. Leave everything else to the company. A short floor that every company meets beats a long standard that none of them do.

Start Where Your People Already Are

The usage is already there. You can treat it as a threat to stamp out, or as the clearest map you will ever get of where AI helps your teams. I'd take the map.

Start with the inventory. Sort what you find. Replace the risky, valuable workflows with a governed tool that is actually good. Tie access to roles so the policy enforces itself. Then review it on a schedule, because new AI features and connectors will change the picture every quarter.

I don't know exactly what AI governance looks like in two years, when more of the work is done by agents acting on their own. The permission models for that are still being figured out across the industry. What I'm confident about is that the companies that know what's in use today will adapt faster than the ones still writing policy for a blank page.

This is the work behind PressW's AI Security, Compliance & Governance service: assess current usage, write policies people can follow, and configure access, retention and audit controls directly in the platform. If you want a second set of eyes on what's already running in your portfolio, Book a Discovery Call.

FAQ