Shadow AI: Why Nobody in Your Company Knows How Many AI Tools Are Running

^{Key takeaways}

^{- Shadow AI is the default state, not the exception. In almost every organisation we work with, AI tools are running that neither IT nor data protection fully know about.}

^{- It doesn't happen through carelessness. It happens through speed: business units solve real problems faster than central approval processes can respond.}

^{- The three most expensive consequences are rarely the ones people think of first: untraceable costs, unprovable data flows, and duplicated work on the same use cases.}

^{- Banning tools doesn't work. What works: make usage visible, assess it, and then offer an official route that is faster than the unofficial one.}

^{- From our experience: the way out of shadow AI doesn't start with a policy. It starts with an inventory and one productive use case that proves the governed path is quicker.}

^‍

A few months ago we ran a workshop with a client in the industrial sector. The agenda was straightforward: which AI use case should they start with? About an hour in, someone from sales spoke up and said, roughly: "We're already doing this. Have been for six months."

It turned out the sales team had subscribed to an AI tool to help draft proposals. Two departments over, someone else was working on almost exactly the same problem with a different vendor. IT knew about neither. Nor did the data protection officer.

This is not an isolated case, and it is not a failure on anyone's part. It is what happens when a technology becomes so accessible that any department can adopt it on its own, while the organisation is still deciding how to deal with it.

‍

What shadow AI actually means

The term borrows from "shadow IT", but describes something considerably more dynamic. Shadow IT used to mean: a department buys software that isn't in the IT catalogue. That was manageable, because software cost money, required installation, and eventually showed up in accounting.

Shadow AI works differently. Entry is often free, usage requires no installation, and the meaningful costs arise not at purchase but during operation. An employee who feeds customer data into a language model to get a summary doesn't trigger a procurement process. They trigger a data transfer that nobody logs.

Concretely, shadow AI covers three things:

Unregistered tools. SaaS subscriptions for AI assistants, transcription services, image generators or coding aids, signed up for at team or individual level.

Uncontrolled data flows. Company data entered into models whose processing location, retention period and training usage have never been reviewed.

Undocumented automations. Workflows, scripts and agents that affect production processes but belong to nobody officially.

The third category is the most dangerous, because it is the only one that keeps running by itself after the person who built it has left the company.

‍

Why this happens even when everyone acts reasonably

In the conversations we have, we rarely meet someone who deliberately circumvents rules. What we see instead is a conflict of objectives that the organisation hasn't resolved.

The business unit has a concrete problem and a concrete deadline. A tool that solves that problem today is more attractive than an approval process that delivers an answer in six weeks. From the perspective of the person who has to deliver, that decision is rational.

IT, in turn, cannot process requests arbitrarily fast, because every new data processing activity has to be assessed. That is equally rational.

The result is an informal parallel operation in which both sides believe they are doing the right thing. And the longer that state persists, the more expensive the eventual clean-up becomes.

We have lived through this ourselves, not only at client sites. During our own AI transition in 2025, we built over 200 automations internally. We switched more than half of them off again, because they turned out to be duplicates, dead ends, or simply not operable. That was part of the learning curve. But it only stayed manageable because we knew from the start which automation was running where.

‍

The three most expensive consequences:

When companies discuss shadow AI, the conversation usually starts with data protection. That concern is legitimate, but it is not what hurts first in practice.

Costs nobody can attribute

Model usage is billed by consumption. A workflow with a flaw in its retry logic can burn through more tokens in a single night than a department's entire planned annual budget. We have seen exactly this case: a workflow without rate limiting that got stuck in a retry loop.

The individual incident isn't the real problem. The real problem is that without a central access point, nobody can say which team is generating which costs, or whether a given spend is matched by any value at all.

Data flows you cannot evidence

During an audit it isn't enough to state that you don't share personal data with third parties. You have to be able to demonstrate it. When AI usage runs through dozens of individually signed-up accounts, that evidence doesn't exist.

For companies in regulated industries this is not a theoretical risk. In telecommunications and financial services, where we frequently work, GDPR-compliant processing, ISO 27001-compatible processes and full auditability are non-negotiable. The EU AI Act adds a further layer that presupposes an overview of the systems actually in use.

The same work, done several times

This is the consequence least often named and the one that costs the most. When three departments independently build a document summarisation tool, you end up with three solutions nobody maintains, three islands of knowledge, and zero economy of scale.

The real loss isn't the wasted effort. It's what doesn't happen: none of the three approaches becomes a building block others can build on.

‍

Why a ban is the wrong first move

The obvious response to shadow AI is a policy prohibiting the use of unapproved tools. In practice this usually achieves two things: usage disappears from official channels, and the willingness to talk about it openly declines.

Someone using an AI tool that works well doesn't stop using it. They stop mentioning it. With that, the organisation loses precisely the information it would need in order to steer.

The more effective approach is less comfortable, because it means work on the central side: the official route has to be faster than the unofficial one. Only then does governance become something people adopt voluntarily.

"Governance isn't overhead. It's the only way to move fast and keep control at the same time." Daniel Pötzinger, CTO, AOE

‍

How we approach it: three steps instead of a policy

Across our projects, an approach has emerged that starts from reality rather than from the ideal state.

Step 1: Take inventory without assigning blame

First, what is actually running has to come out into the open. That only works if it is clearly communicated that this is not about consequences, but about establishing a baseline.

We work with a structured survey: which tools are in use, who uses them, what data goes in, what value comes out. Network and billing data can additionally surface hints of services that were never reported. The result is rarely flattering, but it is the first reliable basis for every decision that follows.

Step 2: Assess rather than blanket-ban

Not every tool needs replacing. Some are uncritical because no sensitive data goes into them. Others solve a problem so well that the right answer is to adopt them officially and secure them properly.

Our assessment runs along three questions: which data is involved? How business-critical is the process? And does this already exist somewhere else in the organisation? That assessment produces the prioritisation: what gets replaced, what gets adopted, what gets consolidated.

Step 3: Build a faster official route

This is the decisive step. As long as the governed path is slower than the ungoverned one, nothing changes.

For our clients this usually means: a central platform on which automations and AI access run, with pre-configured access to approved models, clear project structures, and a deployment path that takes days rather than weeks.

Technically we typically build this on n8n, operated on sovereign infrastructure in Germany. A central LLM gateway consolidates all model requests in one place, so that routing, costs and audit trail converge at a single point. Anyone who wants to build an automation doesn't need to sign up for their own subscription to do it.

One thing matters here: we don't recommend building a complete framework first and the first use case second. The reverse works better. The first productive application becomes the lighthouse project through which the key governance requirements are met by example. That builds trust across the organisation faster than any document. We've described what this looks like in detail in n8n Governance: What Lies Between the First Pilot and a Productive Enterprise Platform

‍

What actually changes

We accompanied one of our clients, along exactly this path. The starting point was comparable: scattered automation efforts, no shared foundation.

The point here isn't the technology. The point is the sequence: first a route that works, then the expectation that it will be used.

How to tell whether this applies to you

Five questions that should be answerable without preparation. If you hesitate on two or more, shadow AI exists in your organisation.

1. Can you say how many AI tools are currently in use?

2. Do you know which company data those tools process?

3. Can you attribute last month's model costs to a specific department?

4. Is there a named person responsible for any given automation currently running?

5. What happens to an automation when the person who built it leaves the company?

In our experience, the fifth question is the uncomfortable one.

‍

The sensible next step

Shadow AI isn't a sign that something has gone wrong. It's a sign that people in your organisation want to solve problems before someone solves them for them. That is, first of all, good news.

The task isn't to slow that energy down. It's to give it a track that holds.

If you want to know where your organisation stands, a structured inventory is the most pragmatic entry point. Get in touch and we'll look at it together.

‍

Frequently asked questions about shadow AI

‍What is shadow AI?

Shadow AI refers to the use of AI tools, models and automations within a company without them being centrally recorded, approved or monitored. This includes unregistered SaaS subscriptions, unreviewed data flows to external model providers, and automations with no documented ownership.

How is shadow AI different from shadow IT?

Shadow IT traditionally involved software whose purchase cost money and therefore became visible sooner or later. Shadow AI is harder to detect, because entry is often free, no installation is required, and costs only arise during ongoing operation. It also frequently involves company data flowing to third parties.

What are the concrete risks?

The three most common are: uncontrolled and unattributable model costs; the inability to evidence how personal data is processed during audits and compliance reviews; and multiple teams working in parallel on the same use cases without a shared foundation.

Should companies ban AI tools?

A ban usually just pushes usage further into the informal space and worsens the information available to decision-makers. It is more effective to offer an official route that is faster and simpler than the unofficial one, and to start by making existing usage visible.

How do you start when you don't know what's running?

With an inventory that assigns no blame. Only once it is transparent which tools are used, which data is involved and what value is being created can you meaningfully prioritise what to replace, adopt or consolidate.

How long does it take to bring shadow AI under control?

The inventory itself typically takes a few weeks. A first productive use case on a governed platform is generally achievable within a few weeks. Migrating all existing applications takes longer and should be prioritised by criticality rather than completeness.

What role does the EU AI Act play here?

The EU AI Act presupposes that companies know which AI systems they operate and how those systems are classified. Without an overview of the systems actually in use, that requirement cannot be met. Shadow AI is therefore not only a security topic but a compliance topic.

‍

Further reading: n8n & AI Automation Gvernance and AI Workflow Automation

‍