Back to Guides
Enterprise Software 15 min read

Training a team to recognize what must never be pasted into a chatbot

Training a team to recognize what must never be pasted into a chatbot

Most companies that have rolled out AI tools have published the list: no credentials, no customer records, no source code, no unreleased numbers, nothing covered by a client agreement. The list is correct and, on its own, changes little. What leaves the building is rarely sitting in front of someone labeled as one of those things.

What moves behavior is a short test a person can run in the two seconds before pressing paste, drilled on the documents they handle every week. Here is that test, the cases worth drilling, and how to tell whether anyone got better at it.

The list is not the skill

The behavior is measured now, and it does not look like ignorance of a rule. Cyberhaven Labs’ 2026 AI Adoption and Risk Report, built on observed data movements, found 39.7% of AI interactions involve sensitive data across typed prompts, pastes, and uploads. The average employee enters sensitive data into an AI tool about once every three days.

Netskope’s Cloud and Threat Report for 2026, summarized by Infosecurity Magazine in January 2026, puts the average company at 223 generative AI data policy violations a month, with roughly 3% of AI users responsible for them. It also found 47% of workplace AI users still on personal accounts, down from 78% a year earlier.

Most of those companies have a written rule, so the rule is not what failed. Two readings matter for training design: the violations concentrate in a small group of heavy users, usually the most productive people on the team, and the exposure is routine, not dramatic. Once every three days is the cadence of someone doing their job.

The gap is not between the policy and anyone’s willingness to follow it. It sits between knowing the categories and sorting the thing currently in the clipboard, which is the gap the safe AI use playbook treats as the central problem here. Recognition is a skill, and skills are drilled, not published.

Three questions that fit in the pause before a paste

The test has to be short enough to run without leaving the window. Three questions cover almost every case.

1. Whose is it? Not who typed it, but who it belongs to. If it arrived under an agreement, or came from a client, a candidate, or a counterparty, it is not the employee’s to move. This catches the largest share of hard cases, and it is already answered in writing: most teams find their existing NDAs already govern pasting client material into AI tools without naming the tools at all.

2. Could one person be picked out of it? Deleting a name is not anonymizing. A rare job title, an exact date, or the pairing of two ordinary details is often enough to identify someone to anyone who knows the company. The working version: if a colleague read this with the names stripped out, would they know who it is about? If yes, it is personal data.

3. What happens if this shows up in a stranger’s answer? A feeling, not a framework. People are poor at sorting data and good at imagining a bad phone call. If the answer involves a client, a regulator, or a lawyer, the paste stops.

Then one question about the account rather than the content: is this an approved work account, and can the same help be had with a placeholder? Most requests survive substitution. “Rewrite this termination clause to be less aggressive” needs no parties, dates, or amounts.

Replace them with [CLIENT], [DATE], [AMOUNT], and the model does the same work on the same sentence. Teach that as the default move, not the fallback: a team that reaches for placeholders first stops having to sort most of what crosses its desk.

The categories, stated once

The list belongs in the training as reference, not as the lesson. State it once, with the substitute next to it.

Never paste Typical examples What to do instead
Credentials and keys Passwords, API keys, tokens, recovery codes, a login screenshot Nothing. There is no safe version of this
Personal records Customer lists, employee files, health details, candidate files Describe the shape of the data, not the records
Client and counterparty material Anything received under an NDA, a draft the other side wrote, third-party data Ask the owner, or use a public example
Unreleased financial and deal material Forecasts, board decks, pricing changes, acquisition work Round, relabel, or wait until public
Code and system internals Private repositories, config files, internal URLs Use a minimal reproduction with fake values
Privileged communications Legal advice, investigation notes, matters with counsel Route through counsel, not the chat window
Security documentation Findings, penetration test results, control gaps Keep inside the approved security tooling

Add one line under the table: a paid account is not a permission slip. Account tier changes retention and whether input trains a model, which is why enterprise accounts are worth paying for. It changes nothing about a duty owed to a client.

Six cases where the rule is not obvious

These fill the session, because each is a case where a careful person gets it wrong.

Case The instinct The ruling
The redacted contract Names are gone, so it is safe Terms, dates, and amounts often identify both parties. Redaction is not de-identification
The screenshot It is an image, not data The sidebar, the browser tab, and the account number in the corner all go with it
The anonymized case study No names, so no problem A unique job title plus an exact month plus one unusual detail is one person
The counterparty’s draft We are allowed to work on it Working on it is not moving it, and the other side’s consent was never asked for
The transcript from an unapproved note-taker The meeting already happened The recording is a new copy of the conversation in a system nobody vetted, and the attendees were often not told
The error log It is machine output Logs carry live tokens, internal hostnames, customer identifiers, and sometimes the payload itself

Work these with the team’s own material, not these summaries. Disagreement is the point: where half the team says “fine” and half says “never”, the rule is too vague to apply, and the fix is a rule change rather than more training. Professional-service teams face a sharper version, since the duty is owed to a named client, and the rules of engagement for client confidentiality start there.

Where the paste actually happens

Training that assumes one approved tool misses most of the surface. Where material leaves from, in rough order of how often it is forgotten:

  • Personal accounts. Cyberhaven’s 2026 data puts the share of use running through personal rather than company accounts at 32.3% for ChatGPT and 58.2% for Claude. The same person uses both, one managed and one not, often in the same hour.
  • Browser extensions. A summarizer with permission to read the page reads all of it, including the record open behind it.
  • Meeting note-takers. Someone connected one to their calendar, it joins meetings nobody approved it for, and the transcript lands in a consumer account.
  • AI features inside tools you already pay for. A help desk or a document editor adds an assistant, and the data boundary moves with no purchase decision.
  • Phones. Camera, clipboard, and a consumer app, outside any managed browser.

Most of this is invisible to the controls people assume cover it. Inspection tools catch typed prompts and uploads in configured browsers on managed devices, and a review of what data-loss prevention for AI assistants actually catches shows how much falls outside that. For most of this surface, recognition is the only control there is.

That is also the argument against watching people instead. Monitoring everything is slow and corrosive to the trust an AI enablement program needs, which is why enforcing an AI policy without surveillance is the practical route at this size.

How the session is run

Ninety minutes, the team’s own documents, no slide about how a model works:

Ten minutes: what the tool does with what you type. Account tiers, retention, whether input can be used to improve a model, who at the vendor can see it. Short and factual. The facilitator has to answer “why” here, because a room that suspects the rules are arbitrary drops them the moment they are inconvenient.

Twenty-five minutes: sort the clipboard. Twenty real items, pulled from the team’s own drive the week before, sorted into three piles: paste as is, paste with placeholders, never. Run it in pairs so people argue, and capture every disagreement.

Twenty-five minutes: the rewrite round. Take the three hardest items and get the same useful answer with placeholders. People stop believing the rule costs them speed the moment they watch a redacted prompt produce the same output.

Twenty minutes: the first hour after a mistake. Who to tell, in what order, and how the company will respond. Say out loud that nobody gets fired for reporting a paste. A team that believes otherwise reports nothing, and an exposure you never hear about cannot be contained.

Ten minutes: write the one-page rule. The team drafts it in their own words from the disagreements captured earlier. A rule a group wrote is a rule that group applies.

Then one reinforcement. Cameron Coles, then Cyberhaven’s VP of marketing, put the technical limit plainly in Route Fifty’s April 2023 reporting: standard security tools cannot stop a user pasting text into a chat window. Something has to carry the session’s judgment into the week after, whether a browser prompt or a team norm about asking first.

What to measure

Completion rate measures attendance. Four signals measure the skill.

Signal How to read it
Before-and-after sorting score Same ten items, four weeks apart. The only direct measure of the skill
Questions asked in month one Should go up. A rise in “can I paste this?” is the skill working
Distinct accounts in use Should fall as people move to the approved account, read from admin consoles, not a survey
Violation events from existing logging Useful as a trend, weak as a level: it sees only the surfaces the tooling covers

Brief leadership on the second signal in advance, because it looks like bad news and is not. Before training, people paste and move on. After, they stop and ask, and the month the questions spike is the month the team acquired the instinct.

Set the retest date during the session: four weeks later, same items, scored. Without it there is no evidence the hour did anything, and no answer when a CFO asks.

The EU AI Act’s Article 4 covers staff competence. The current text says providers and deployers of AI systems shall take measures to support the development of AI literacy among their staff and others operating those systems on their behalf, and states plainly that the duty does not require guaranteeing any specific level. It came into force on 2 February 2025 under Article 113(a).

Much of the secondary commentary still reports a duty to ensure a “sufficient level”, which the article as it now reads does not say. Its reach is also wider than US companies expect: “on their behalf” takes in contractors, which is why extending the policy to contractors belongs to the same job.

For a framework to structure against, NIST’s Generative AI Profile, published as NIST AI 600-1 in July 2024, extends the AI Risk Management Framework to these systems and names data privacy among the risks it treats. For professional duties, the American Bar Association’s Formal Opinion 512, issued 29 July 2024, is the clearest worked example in any field: a lawyer must understand the tool, protect client confidentiality, check output, and supervise its use, and blanket consent buried in an engagement letter is not enough. Healthcare, finance, and government each have their own version of that reasoning.

Each of those regimes expects the people to be competent, and none accepts a published list as evidence of it. That is the argument behind the case that buying licenses is not the same as building capability.

A license is a purchase. Recognition is a trained skill, and it is what stands between a confidential document and a system you do not control.

For a read on where your team’s recognition sits before you design anything, we run a free AI-readiness call with a short sorting exercise on your own material.

Frequently asked questions

What should never be pasted into an AI chatbot?

Credentials and keys, personal records of customers or staff, material received under an agreement, unreleased financial or deal information, private code, privileged legal communications, and security findings. The harder rule to teach is the test behind the list: whose is it, could one person be identified from it, and what happens if it appears in a stranger’s answer.

Does an enterprise AI account make it safe to paste confidential material?

No. A business tier changes retention and whether input can be used to improve a model, which is worth paying for, and changes nothing about a duty owed to a client or a patient.

Are we legally required to train staff on AI confidentiality?

If you operate AI systems inside the EU AI Act’s scope, Article 4 has required measures to support staff AI literacy since 2 February 2025. In the US the duty arrives through sector rules and professional duties, not one statute, and for regulated professions it is explicit, as in the ABA’s Formal Opinion 512.

How long should this training take?

Ninety minutes once, then a thirty-minute refresh when the tool estate changes. Most of it should be the team sorting and rewriting their own documents, not listening.

How often does it need to be repeated?

Annually as a floor, and immediately when a new assistant appears inside a tool the team already uses. That changes the exposure more than a calendar year does.

Does this apply to contractors and freelancers?

Yes, and they are the most commonly missed group. Article 4 reaches people operating AI systems on a company’s behalf, and a contractor on a personal account is the exposure most policies forget.

What should someone do in the first hour after pasting something they should not have?

Tell the named person, record what was pasted and into which account, and do not delete the conversation. Deleting the chat destroys the evidence needed to judge how serious it was.

Is redacting a document enough before pasting it?

Usually not. Removing names leaves terms, dates, amounts, and job titles that often identify the parties to anyone in the same market, which is why placeholders beat deletion.

Can we just block the tools instead?

Blocking moves use to personal accounts and phones, where there is no visibility at all. Netskope reports personal-account use falling from 78% to 47% in a year, and the way to keep it dropping is a sanctioned alternative people prefer, not a block list.

How do we know whether the training worked?

Score the same ten items before and four weeks after, then watch for a rise in questions in month one, a fall in distinct accounts in use, and the trend in whatever violation events your logging already sees.

Key takeaways

  • The never-paste list is reference material, not training. The skill is sorting an unclear document in the seconds before a paste.
  • Three questions carry most of the load: whose is it, could one person be identified from it, and what happens if it shows up in a stranger’s answer.
  • Teach placeholders as the default move, since most requests survive substitution with no loss of usefulness.
  • Drill the hard cases on the team’s own documents: redacted contracts, screenshots, anonymized examples, counterparty drafts, unapproved transcripts, error logs.
  • Exposure spans personal accounts, extensions, note-takers, embedded assistants, and phones, most of which inspection tools do not cover.
  • Measure a before-and-after sorting score, and expect questions to rise in month one. That rise is the signal recognition took hold.

Last Updated: Nov 16, 2026

DJ

Dirk Jan van Veen, PhD

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles