What Shadow AI Numbers Mean for Your Own Drafts

Nobody is going to finish your company’s AI policy before your next deadline. Most organizations, by the account of the same surveys that report near-universal AI adoption at work, still have not published one, and even where a policy exists, plenty of employees have never sat through the training that goes with it. Waiting for governance to catch up is not a plan, it is a postponement.

What you can control today is smaller and more useful: what you personally paste into a public chatbot before you hit send. That single habit accounts for most of the risk you as an individual can actually create, and fixing it does not require anyone’s approval.

The Numbers Vary, But They Point The Same Way

A 2026 workplace AI survey conducted by Wakefield Research and published by PagerDuty, covering roughly 1,250 office professionals across Australia, Japan, the UK, and the US, found that about two thirds of respondents had used an AI tool at work that they believed was not officially sanctioned, and that a meaningful share, reported at over a third in that survey, had entered customer or confidential business data into a public AI tool (PagerDuty). Other 2026 reports on the same trend put the numbers higher or lower depending on how they define confidential data and which countries they sample, so treat any single figure as a range rather than a fact. What does not move across surveys is the direction: a large majority of knowledge workers now use AI daily, a large share of that use is unsanctioned, and a real minority to sizable minority of workers have pasted something they should not have.

Most Employers Still Have No Formal Policy Or Training

The same class of surveys consistently finds that formal AI usage policies lag adoption by a wide margin, and that where policies do exist, training on them is inconsistent at best. That gap is exactly why individual habits matter more than they should have to. If your employer has not told you what is and is not allowed, and has not shown you what a safe workflow looks like, the default behavior most people fall into is using whatever tool is fastest, which is usually the free public chatbot already open in another tab.

One Rule Removes Most Of The Risk

You do not need a security team to fix this for yourself. The rule that does most of the work is simple: if the text names a real client, a real figure, an unreleased plan, or anything you would not want quoted back to you outside the company, it does not go into a tool whose data handling you have not read and cannot control. Everything else, generic templates, public-facing copy, text with nothing to leak, is fine wherever you like.

Before:

Rewrite this for me so it sounds more polished: “Following our call with Meridian Health, we agreed to the $340k pilot starting in November, contingent on their board sign-off by the 12th.”

After:

[Same text, rewritten locally, never transmitted]: “Following our call with Meridian Health, we’ve agreed to a $340,000 pilot beginning in November, pending board sign-off by the 12th.”

The wording barely changes. What changes is where the client name and the dollar figure went while it happened, covered in more depth in whether it is safe to paste work emails into ChatGPT.

A Wrivio Context makes this rule automatic instead of something you have to remember mid-sentence:

Rewrite this message for clarity and tone only. Keep every name, date, figure, and commitment exactly as written, and do not summarize, generalize, or drop any specific detail. Treat this as confidential and process it without adding anything not in the original.

Press Ctrl+Shift+Space, paste the draft on the local engine, and check the diff to confirm only the wording moved.

A Private Alternative For The Sensitive Half Of Your Work

You do not have to give up AI assistance to follow this rule, you just have to split your work. For anything generic, a cloud or preview engine is fine, convenient, and usually free within a reasonable limit. For anything with a name, a figure, or a commitment attached, switch to a private rewriter’s local mode, where the model runs in process on your own machine and the rewrite never touches a network. That is a stronger guarantee than trusting a chatbot’s retention settings, and it costs you a toggle, not a policy meeting, a point covered from the tooling side in local AI vs cloud AI for confidential writing.

Five Personal Rules Worth Keeping

1. If it names a real client, employee, or deal, it stays out of a public chatbot, full stop.
2. Default to a local or private tool for anything you would not want forwarded outside the company.
3. Save the public chatbot for generic drafting: templates, brainstorming, text with nothing specific in it.
4. When you are not sure if something counts as confidential, treat it as confidential.
5. Do not wait for a policy to tell you this. By the time one arrives, the habit should already be yours.

Common Questions

How many employees actually paste confidential data into public AI tools?

Reported figures vary by survey and by how confidential data is defined, but 2026 workplace surveys generally put the share somewhere between roughly a quarter and over a third of respondents, with a majority also reporting they use AI tools their employer has not formally sanctioned.

Should I wait for my company to write an AI policy before changing how I work?

No. Most organizations still lack a formal policy according to 2026 survey data, so waiting means continuing current habits indefinitely. The one rule that removes most personal risk, keeping named or confidential details out of public chatbots, does not require a policy to follow.

What counts as confidential enough to avoid a public chatbot?

Anything naming a real client, employee, dollar figure, unreleased plan, or internal detail you would not want quoted outside the company. When in doubt, treat the text as confidential rather than risk guessing wrong.

Is using a public chatbot ever fine for work?

Yes, for generic text with nothing specific in it, such as templates or general brainstorming, the privacy concern does not apply because there is nothing to leak.

What is the private alternative to a public chatbot for sensitive drafts?

A rewriter with a true local mode, one that runs the model on your own device and makes no network call during the rewrite, so the sensitive text never has the chance to be logged or retained by a third party.

Download Wrivio for Windows and keep the sensitive half of your drafting on Local mode, where it never leaves your machine.