Most shadow AI policies are written as a list of prohibitions, sent as an email, and forgotten within a week. They fail for a reason that is visible in the data rather than in anyone's character.
The BlackFog and Sapio survey of 2,000 employees at UK and US organizations above 500 staff found 49% used AI tools their employer had not sanctioned. In the same survey, 63% considered it acceptable to use AI without IT oversight when no approved option exists. Read together, those two numbers describe a supply problem. People are not refusing approved tools; in many organizations there is nothing approved that does the job.
What a ban actually does
A ban without an alternative does not end the activity. It moves it to personal devices and personal accounts, where you have no logging, no data agreement, and no chance of finding out until something leaks. The shadow AI benchmark card makes the same point: usage at this level means the approved toolset does not cover real work.
There is also a free-tier problem inside approved tools. The same survey found 34% of respondents used free versions of company-approved AI tools, and among those using unapproved tools, 58% relied on free versions. Free consumer tiers often come with different data terms from the enterprise contract. "Approved tool" has to mean the licensed account, and the policy has to say so.
The five parts of a policy that works
1. Name the approved tools first
Lead with what people can use, with a link to sign in. If the first paragraph is a list of things that are forbidden, readers stop there. If it is "use these, here is how to get access", the rest reads as guidance.
2. Draw one hard data line
Pick the single rule that matters most and make it unambiguous. A workable form: "Do not put restricted or customer data into any AI tool that is not on the approved list, including approved tools on a personal or free account." Use your existing data classes. One rule people remember beats ten they do not.
3. Say what happens to a request
Point to the intake, and state the turnaround. "Ask for a new tool here; most requests are answered within one business day." The AI use intake playbook covers how to make that promise true.
4. Make disclosure safe
Invite people to declare what they already use, without consequence, for a fixed window. A declaration window surfaces tools that network discovery cannot see, such as browser extensions and personal accounts, and every declared tool becomes an inventory entry with an owner.
5. Name the things that are genuinely off limits
Keep this short and specific: decisions about individuals made by AI without human review, AI tools granted access to production systems or customer accounts without approval, and uploading credentials or secrets anywhere. These are the cases where the harm is fast and hard to undo.
Connections are the real exposure
A chat window that receives a pasted paragraph is one risk. An AI tool connected to a mailbox, a drive, or a ticketing system is another: it can read everything that account can read, and some can act. The BlackFog survey found 51% of respondents had connected AI tools to other work systems without IT approval.
Put connection in the policy explicitly. "Do not connect an AI tool to company email, files, or systems unless it is on the approved list for that connection." Then make sure your identity team can see OAuth grants to AI applications, which is where most of these connections show up.
Enforcement that fits
Enforcement should be proportionate and mostly technical:
- Block only the specific categories in part 5, and only once an approved alternative exists.
- Redirect traffic to consumer AI domains toward a page explaining the approved option, rather than a bare block page.
- Monitor OAuth grants and data loss events involving AI tools, and treat findings as intake requests first and incidents second.
- Review the approved list quarterly against what discovery finds; anything found repeatedly is a tool people need.
Keep it short, and keep it current
A policy that runs to ten pages will not be read, and one that is read once will not be remembered. Aim for something an employee can finish in three minutes, and link out to the detail. Put a date on it and a named owner.
Then revise it on a schedule. The approved-tools list in particular goes stale quickly, because vendors ship new AI features and employees find new tools faster than any annual review cycle. A quarterly refresh of the list, announced in a sentence or two, does more for compliance than a longer policy would. Each refresh is also a reminder that the approved path exists.
Where the frameworks come in
NIST AI RMF is voluntary, and its Generative AI Profile asks organizations to enumerate generative AI systems for an inventory and to define any inventory exemptions in policy (GV-1.6-001 and GV-1.6-002). A shadow AI policy is, in that framing, the rule for how unapproved use becomes inventoried use. It does not need to be longer than two pages to do that.
Next steps
- Publish the approved-tools list with sign-in links before the policy itself.
- Write the single data rule using your existing classification.
- Run a two-week declaration window and convert every declared tool into an intake record.
- Set up the intake process with a one-day target for low-risk requests.
- For the identity side of connected tools, see identity for AI agents, and for templates see /ciso/templates/.