Cloud and AI Were Supposed to Kill Shadow IT. They Industrialised It

What the data says

Gartner has put shadow IT at 30 to 40 percent of IT spend in large enterprises for years. Everest Group puts it above 50 percent. Those were pre-AI figures.

Here is where 2026 sits:

• Torii's annual benchmark counts 831 applications in the average enterprise, with more than 61 percent of them classified as shadow IT, discovered through usage, browser activity or direct signup rather than through procurement. More than half of the top shadow apps are pure-play AI tools, and close to 700 new AI applications entered enterprise environments in a single year.

• Zylo's 2026 index shows 48 percent of SaaS spend now driven by business units outside IT's control, while spend on AI-native applications grew 108 percent year on year.

• Gartner expects 75 percent of employees to acquire, modify or build technology outside IT oversight by 2027, up from 41 percent in 2022.

• 77 percent of IT leaders report costs that only surfaced after the SaaS contract was signed.

So the trend line is not flat. It is steep.

Why the cloud made it worse

The promise of cloud ERP was consolidation. One platform, one vendor, less technical debt, fewer bespoke integrations to maintain.

What actually happened is that SaaS ERP is deliberately thin at the edges. The core is standardised, which is the point, and every gap that used to be closed by a customisation now gets closed by a departmental point solution bought on a corporate card. You retire on-premise technical debt and re-acquire it as an untracked subscription estate. The debt did not disappear. It changed cost centre and stopped being visible.

AI accelerated the same pattern for a simple reason: the tools are cheap or free, they need no implementation, and the user sees the value in ninety seconds. Nobody is going to write a business case for something that already works.

The part I have lived through

I have been on the inside of this. In a previous role we ended up deploying browser-activity monitoring, not for productivity policing, but because it was the only reliable way to see which free-tier tools people were actually using during the working day. There was no other discovery mechanism. Nothing hit the AP ledger, nothing hit the card statement, nothing hit an intake form. Zero spend, therefore invisible, therefore not a procurement problem. Except it was.

Two categories showed up immediately: translation and writing assistance. DeepL is the honest example, because you do not have to speculate about the risk. DeepL's own free terms of use state that they reserve the right to process uploaded text and documents for a limited period in order to train and improve their neural networks, and the same terms explicitly prohibit using the free service for content containing confidential or personal data of any kind. The paid tiers are a genuinely different data handling model, with no training on customer content and deletion after processing.

Now picture what actually gets pasted into a browser tab at four in the afternoon. A supplier contract that needs reading in Arabic. A term sheet. A disciplinary letter. An RFP response from a bidder. The employee is not being reckless in their own mind. They are trying to finish a task, and the tool is right there, and it is free.

Same story with browser-based writing and spellcheck assistants. Read the terms on the free tier and compare them against the enterprise tier before you assume they are equivalent products. They usually are not. What you are buying with the enterprise licence is very often the data clause, not the features.

That is the ballpark change. It stops being a licensing conversation and becomes a confidentiality one. A free tool has no contract, no DPA, no data residency commitment, no IP position, and no counterparty to go back to when something surfaces.

What this does to procurement, specifically

1. You lose the demand signal. Category strategy, consolidation, volume leverage, benchmark work: all of it depends on knowing what is being consumed. When half the portfolio is invisible until renewal or audit, you are negotiating a fraction of the estate and calling it category management.

2. Consumption pricing breaks the commercial model you were trained on. Around two thirds of IT leaders now report surprise charges from consumption-based and AI pricing. A per-seat contract quietly becomes a per-token bill nobody forecast. Fixed-price benchmarking and unit-rate cards do not protect you against that. Your commercial construct has to change before your negotiation does.

3. The risk lands on your desk regardless. You did not review the tool. You still own the exposure when the data protection question is asked, because the question is asked of the function that is supposed to control third-party engagement.

The uncomfortable diagnosis

Shadow IT is mostly a discovery problem wearing the costume of a discipline problem.

People do not bypass procurement because they enjoy risk. They bypass it because finding and requesting the sanctioned tool is slower and harder than opening a new tab. If your intake process takes six weeks to approve a 200 dollar per month subscription, you are manufacturing the shadow IT you are then asked to police. Every additional gate raises the incentive to route around the gate.

What actually works

• Tier the process to the risk, not the value. A low-spend, high-data-sensitivity tool is more dangerous than a high-spend, low-sensitivity one. Most intake forms have this backwards.

• Build a genuine fast lane. Pre-approved catalogue, standard terms, 48 hour turnaround for anything under a set threshold that touches no sensitive data. Make the sanctioned route the easy route.

• Discover before you govern. Card and expense data, SSO logs, network and browser telemetry, AP line items. Build the picture first. You cannot write policy for a population you have not counted.

• Buy the data clause deliberately. For any tool that touches contracts, employee data, pricing or supplier information, the paid tier is not a nice-to-have. It is the control.

• Publish the alternative. Half the shadow adoption stops if people know there is an approved translation tool and where to find it.

Shadow IT is not going to be eliminated. Anyone selling you that outcome is selling a dashboard. What is achievable is shrinking the gap between what employees need and what procurement makes easy to get, and then making sure the tools that handle your commercially sensitive material are ones you actually have a contract with.

The rest is just spend you have not found yet.