Data Report · RedactSure Research
Chris Sowa · Founder, RedactSure
Published
Last updated
How Do You Give Employees a Sanctioned Safe Space for AI?
What IT and security leaders in healthcare, education and manufacturing are learning about shadow AI
In brief
-
An acceptable use policy without a technical control leaves leaders with no view of how staff actually use AI.
-
Bans and silence do not stop use. People reach for personal AI accounts when there is no approved route.
-
The leaders I work with keep arriving at the same order: provide a sanctioned environment first, then restrict the unsanctioned tools.
-
A sanctioned space has to handle the work people are measured on, including work with sensitive data, or the shadow use continues.
-
The activity record should land where security already looks, such as the organization's SIEM.
Over the past several months I have worked with IT and security leaders at a multi-site physician group, a healthcare services company, a university athletic department, a public school department and a global industrial manufacturer. The sectors differ. The problem does not. People want to use AI at work, the policy says be careful, and the organization has no safe place to send them.
The numbers explain the urgency. IBM's 2025 Cost of a Data Breach study found that 63 percent of breached organizations had no AI governance policy, that one in five reported a breach involving shadow AI, and that high levels of shadow AI added about $670,000 to the average breach cost (IBM, 2025). Microsoft and LinkedIn found that 78 percent of AI users bring their own tools to work (Microsoft & LinkedIn, 2024). This article draws on what the leaders I have been working with told me about closing that gap.
A policy without a control is a hope
An IT leader at a multi-site physician group with a few thousand employees described a position many organizations will recognize. The group had acceptable use policies for AI, but no technical controls behind them, and no central view of how staff were using AI across ambient dictation, call center tools and general purpose chatbots. Outside AI services had not been blocked. A senior security engineer at the same group later made the priority explicit. The first goal was not physician productivity. It was to corral AI access and bring the risk under control.
That instinct matches the guidance. The UK National Cyber Security Centre advises organizations to provide approved, secure alternatives and to build open communication so staff do not hide their use, rather than relying on a ban (NCSC, 2026). Corral first, automate second.
Why bans push AI out of sight
The IT leader for a large university athletics department put the problem plainly when I asked what he would tell a CIO who had blocked AI across the university.
“Well, if I felt comfortable enough, I'd be like, well, guess what they're doing at home with your data? Because that's what they would be doing.”
“You'd probably be pretty foolish if you didn't think any of your data is getting used in any of the AI learning models, because somebody in your organization or somebody who's working for your organization probably is cutting and pasting it somewhere.”
IT leader, university athletics department
A CISO at a Fortune 500 industrial company, whose own company has invested heavily in approved AI tools, recognized the pattern. When I mentioned that I had heard of employees exporting work data to analyze it in a personal AI account because their employer offered no approved tool, she replied:
“I have heard of people doing that.”
CISO, Fortune 500 industrial company
A ban produces invisibility, not safety. The work still happens. It happens in places security cannot see.
Even the most careful organizations want a safe route
Restrictive organizations are not anti-AI. A technology executive at a healthcare services company explained why his frontline operations staff have no AI access at all today, and what he would do with a safe route.
“We're very controlled over what we allow the general operations staff to have access to. So for those guys it's like they can't even access any AI model or push any data anywhere. Would that be able to help them in their day-to-day reporting and summarizing and making them more efficient? Most likely it would. But the risks are too high for us in that area.”
Technology executive, healthcare services company
“We can securely roll out AI to account managers and allow them to analyze data without having to […] worry about them pushing sensitive information through the AI. That's another really good use case because right now we don't give them access to that.”
Technology executive, healthcare services company
The demand is there in the most cautious environments. What is missing is a place where the work can be done without the sensitive values going with it.
The order that works: sanctioned first, restrictions second
With the physician group, the order we discussed had two steps. First, give staff a secure, sanctioned environment for AI work. Second, once that route exists, decide which outside AI services to block. Reversing the order produces the outcome the athletics IT leader described: blocked tools and hidden use. The group ultimately chose endpoint controls for visibility across AI models. That covers where data can go but not what an approved AI tool reads, a gap I describe in a separate article on endpoint security and AI.
The pilot starts with a test employee account in a sandbox, so the team can confirm compatibility with real systems before anyone touches production. Browser extensions that clinical applications depend on, such as those for the EMR, have to be checked early. And the activity record has to flow into the SIEM the security team already uses, so AI use is monitored alongside everything else rather than in a new console nobody watches.
What a sanctioned space has to include
The CISO at the industrial company described a rule she sees across companies, not only her own. Any company that takes security seriously now tells its people not to put personal information into AI, with no exceptions.
That rule protects the data and leaves the work undone, because much of the work people want help with runs on personal information. When I described how tokenization lets an AI work on a record without reading the real values, she put the requirement in one sentence:
“This actually enables you to work with data that you already have in your company that may be sensitive PII that you don't want to get out, right?”
CISO, Fortune 500 industrial company
From these conversations, a sanctioned space needs at least the following:
-
Approved models, under the organization's own agreements and limits.
-
Sensitive values replaced with tokens before any model reads them, so staff can work on real records without the real values entering the model.
-
Task-level policy on where the AI may go and what it may do, with a named person approving consequential actions.
-
Permissions that stay exactly as they are today. Nobody gains access they did not already have.
-
An activity record that exports to the organization's SIEM.
-
Shared templates for common tasks, so good practice spreads faster than shadow practice.
Where RedactSure fits
RedactSure is built to be that sanctioned route. A supervised AI co-worker works in the browser, in the applications staff already use, with sensitive values tokenized before any model reads the screen. A named person approves consequential actions, and every step is recorded for the security team. Everyone else controls where data can go. RedactSure controls what the AI can see.
Frequently asked questions
What is a sanctioned safe space for AI?
An approved environment where employees can use AI and AI agents on their real work, with sensitive data protected, actions governed and activity logged. It replaces a ban with a route people will actually use.
Should we block ChatGPT and other public AI tools?
Blocking can be a step, but on its own it pushes use to personal devices. Guidance from the NCSC and the experience of the leaders quoted here point to providing a sanctioned alternative first, then restricting unsanctioned tools.
How do you monitor AI use without creating new risk?
Route AI work through one environment whose activity record exports to the SIEM, and keep sensitive values tokenized in that record so the monitoring does not become a second copy of the data.