In December 2005, four hospitals on the East Coast of the United States recorded more than 300,000 medication administrations in their logs, and a few years later the researcher Ross Koppel and his co-authors counted how often a nurse confirmed an administration despite a warning from the computer. The answer was 4.2 percent of administrations to patients and 10.3 percent of all medications given. The hospitals had earlier introduced a rule that before each administration the nurse scans the barcode on the patient's wristband and the barcode on the medication package, and the computer checks that the person, the drug and the dose all match.
Koppel's team carried out 62 observations, shadowing 31 nurses on intensive care and general medical units in two hospitals, and collected interviews at five. They saw copies of patient barcodes stuck to desks, medication carts, door frames and clipboards, and also to sleeves and belt loops, so that a nurse could scan the sleeve instead of the wristband on the patient's arm. Sometimes a single package was scanned twice for two doses, because the nurse judged the tablets to be identical.
The twist is that the people described meant well. The authors wrote that none of the workarounds came from malice, and that the nurses regarded them as necessary for the patient's good, for efficiency or for safety. They counted 31 causes: crumpled barcodes on packages, failures of scanners and batteries, staff shortages and slow procedures. People who cared about their patients had therefore built around the safeguard an unofficial circuit that nobody supervised, and it opened the way to exactly the mistakes the safeguard was meant to prevent.
If the safeguard is bypassed by those who care about their patients, what does that say about the safeguard?
The nurses in Koppel's study behaved the way employees in many companies behave today toward artificial intelligence tools: they reach for their own solution because the official one does not fit into their working day. The thesis of this post is that a ban will not stop the phenomenon, and that the only answer that works is an approved route that is more convenient than the workaround.
How many people already do this
In its annual Data Breach Investigations Report for 2026, Verizon states that 45 percent of employees are regular users of artificial intelligence tools on company devices, whether or not the company authorised it. Of those users, 67 percent use non-corporate accounts, and unapproved tools ranked third among the most common non-malicious insider actions detected, a fourfold increase within a year. The material most often sent to external models was program source code, followed by images and structured data.
Your equivalent of source code is customer contracts, price lists, sales results, payroll spreadsheets and candidates' CVs. Verizon's figures come from detecting uploads on company devices, so a private phone lying next to the keyboard does not enter the calculation, and the real scale is probably larger. I rely on this report because it measures behaviour on the network, whereas survey answers depend on how the question is asked.
Why people do it despite the ban
People assess risk by what they see today. Twenty minutes saved on summarising a report is visible at once, while a data leak, if it happens at all, will surface in half a year somewhere else entirely, so the brain gives it little weight. Add the cost of effort: a private account opens in a second, while an approved tool requires an application, a manager's consent and a week of waiting.
Group norms also play a part. If the colleague at the next desk has done it for months and nothing has happened, the employee treats the practice as accepted and the ban as a formality printed for the legal department. A ban raises the cost but does not change the balance of gains, so the phenomenon goes below the surface. Someone who knows they are doing something forbidden will not ask for advice, report an incident or share a good method, and the company loses the one thing it can use in that situation: knowledge of what is going on.
What you really risk
Personal data of customers and employees pasted into a private account goes to a provider with whom the company has no data processing agreement, which infringes the General Data Protection Regulation. The terms depend on the plan: in consumer accounts providers may use conversations to train models if the setting is switched on, and Anthropic, for example, has since 2025 extended retention in that case to five years. Business plans and access through a programming interface fall under separate contracts whose content you must read at each provider instead of assuming it is the same everywhere.
There is also the European Union regulation on artificial intelligence. Article 4 obliges companies that use such tools to look after their staff's competence in the area and has applied since 2 February 2025. According to legal commentaries, the changes adopted in 2026 replaced the duty to "ensure a sufficient level" of competence with a duty to take measures that develop it, and the amount of penalties is set by the member states. An approved tool together with training creates documented evidence of such measures, which cannot be said of a ban that everyone circumvents.
The HR department as a special case
The situation looks worst where the pasted text contains data about people. A recruiter who, to save an hour, asks a private account to rank twenty CVs breaks two rules at once: she sends candidates' personal data to a provider without a processing agreement, and she uses a system that, according to the European Commission's information service on the artificial intelligence regulation, belongs to the high-risk category (Annex III, point 4a), because it evaluates and ranks candidates for jobs. The same person can, without much risk, ask the tool for a draft job advert, a summary of the leave policy or a tidy-up of meeting notes. The boundary is worth drawing for the HR department separately and earlier than for others, because the price of a mistake is highest there.
What even a well-chosen tool will not do
A tool with a business contract settles where the data goes, but it does not guarantee that the answers are correct. The model still errs with full confidence, so a contract summary that goes to a customer needs to be read by a person. Blocking website addresses on the company network works on laptops in the office and misses the phone in a pocket, so technology supports a policy but does not replace it. The licence cost is worth setting against the hours your people already save quietly today, because that figure will justify the budget before the board.
Recommendation and a first step for Monday morning
My recommendation is to make one approved tool available faster than you can write its rules, and to refine the rules as people use it. An employee who gets a working account in the first week stops looking for their own, whereas one who waits a quarter for a committee's consent has a quarter in which to form habits outside the company, and changing a habit costs more later than the tool itself.
Start with a question instead of a policy document, with a clear promise that the answers will not be used to draw consequences. Koppel's observations show that workarounds arise from difficulties at work, so employees will answer honestly only if they do not fear punishment.
- Send a short anonymous survey to 10-15 people from different departments: which tools they use, for which tasks, and what kinds of documents they put into them.
- Choose the three most common tasks, for example summarising documents, translation and drafts of messages to customers.
- Choose one tool on a business plan and check in the contract whether the provider trains models on your data, where it processes the data and whether it will sign a processing agreement.
- Write one-page rules that divide data into three groups: may be pasted (public materials), may be used only in the approved tool (internal materials), and may not be used at all (personal data and trade secrets without the legal department's consent).
- After a month, measure what share of respondents moved to the approved tool, and ask the rest what it lacks.
The last step is feedback on a product the company has just bought. If the survey from step one shows that the people using the tools most often are the same ones who score best in annual reviews, ask yourselves whether you want to discourage them or to learn from them what in your company could be done faster.