Your Estimator Just Pasted a Customer Drawing Into a Chatbot
Verizon’s 2026 Data Breach Investigations Report has a finding worth pausing on. Shadow AI – employees using AI tools without company approval – has become the third most common non-malicious insider action found in breach data, with detections up roughly fourfold in a year. Around 45% of employees are now regular AI users on corporate devices, sanctioned or not, and separate surveys put the share of office workers using AI against company policy considerably higher.
The security industry writes this up as a data protection story, with customer records and personally identifiable information at the center. For a manufacturer, that framing misses what is actually at risk.
What leaves a factory is different
In a bank, shadow AI leaks customer data. In a machine shop, it leaks the business.
Think about who in your company has the strongest incentive to use AI without asking. The estimator turning around an RFQ under time pressure, pasting in a customer’s drawing and specification to speed up the quote. The process engineer troubleshooting a stubborn quality problem, pasting in machine parameters and tolerances. The buyer comparing supplier quotations. The engineer asking for help with a PLC routine and pasting in the code.
Every one of those is a reasonable person trying to do their job faster. Every one of those is also your differentiation – the specifications, the process knowledge, the pricing, the control logic that took a decade to refine – going into a system you do not control, under terms nobody in your company has read.
IBM’s breach research makes the cost point in a way that lands for manufacturers: intellectual property was the most expensive category of exposed data, and breaches involving shadow AI carried a meaningful cost premium. Treat the exact figures with the usual suspicion, given who publishes them. The structural point does not depend on the number.
Why this one cannot be undone
Shadow IT was recoverable. Someone signed up for an unapproved file-sharing tool, IT found it, revoked access, blocked the domain, and the leak stopped.
Shadow AI does not work that way. Data already submitted to an external model cannot be retrieved, deleted, or audited. There is no revocation step. If a customer’s drawing went into a consumer chatbot in March, it went, and your remediation options run from “notify the customer” to “hope it does not matter.” That is the difference between an access control problem and a disclosure problem, and it is why the usual IT playbook does not apply.
There is a contractual edge too, and it is the one most likely to bite first. Most automotive and aerospace customers impose confidentiality terms on drawings and specifications. Very few manufacturers have checked whether pasting a customer drawing into a public AI tool breaches those terms. It very likely does.
The fix is not a ban
The instinct is to prohibit it. Bans have a poor record here, for a simple reason the research keeps confirming: people reach for whatever tool is nearest when the approved option does not exist. A ban does not remove the incentive, it just moves the usage onto personal phones where you cannot see it at all. The most encouraging finding in this whole body of data is that unauthorized use drops substantially when a sanctioned alternative is actually provided.
So the useful sequence is the opposite of the instinct. Find out what people are already doing, without treating it as a disciplinary matter, because the honest answers only come if it is not one. Provide a sanctioned option that is genuinely good enough for the jobs they are using it for – which for a manufacturer usually means an EU-hosted or self-hosted setup where the data does not leave your control, exactly so that pasting a drawing into it is not a problem. Then write the rule you can actually enforce, which is about what data may go where, not about which tool is forbidden.
And ask your engineering and estimating people directly. Not “are you using AI,” which invites the safe answer, but “what have you been pasting in.” The gap between what your policy says and what happened last Tuesday is the thing worth knowing, and it is cheaper to find out from your own people than from a customer.
“Content generated using AI”