Because users choose the path that helps them finish work fastest. When the approved route is slower than a personal account or unsanctioned app, employees bypass the control even if the policy is clear. Blocking can reduce obvious usage, but it often increases hidden usage by making governance less convenient than the workaround.
Why Blocking Alone Does Not Change User Behaviour
Perimeter blocking changes the path, not the motive. If people still need a fast way to draft, summarise, analyse, or automate a task, they will look for the lowest-friction route, even when policy says otherwise. That is why “stop access” controls often suppress visible traffic without eliminating the underlying demand for AI assistance.
The real control problem is not whether the approved service is technically reachable, it is whether it is convenient enough to win against personal accounts, browser-based tools, or other unsanctioned options. If the sanctioned workflow adds delay, login friction, or restricted functionality, users will work around it.
Blocking can still be useful when the goal is to reduce exposure of clearly risky destinations, but it is rarely a complete behavioural control. For that reason, the better question is not “did we block it?” but “did we make the approved path easier than the bypass?”
Why Hidden Usage Usually Increases After a Ban
Once a block is in place, the activity often becomes less visible rather than less common. Users shift to private accounts, personal devices, copy-and-paste workflows, or unsanctioned extensions, which makes monitoring harder and increases the chance that sensitive material leaves approved environments.
That pattern is familiar in shadow technology adoption: when workers can complete the task faster outside the approved channel, policy compliance competes directly with productivity pressure. A useful internal reference point is the Agentic AI Security Policy Template, because it reflects the operational reality that policy, ownership, oversight, and retirement need to be paired with a usable workflow.
Behaviour also changes when users do not trust the approved tool to deliver the same outcome. If the sanctioned environment lags in quality, model choice, context length, file handling, or integration, users will treat the block as an obstacle rather than a safeguard. At that point the organisation has not removed demand, it has redirected it into channels with less governance.
What Actually Reduces Unapproved AI Use
The most durable reduction comes from combining perimeter controls with a service that is easier to use than the workaround. That means clear approved use cases, low-friction access, and enough capability that staff do not need to defect to personal tools just to finish ordinary work.
Governance also has to address data handling, because users often choose unsanctioned AI when they believe the approved route is too restrictive for real tasks. The Samsung ChatGPT leak 2023 is a practical reminder that once employees decide an unapproved tool is faster, sensitive material can move outside policy even when the organisation has already warned against it.
Blocking should therefore be treated as one layer in a broader adoption strategy, not as the strategy itself. The better operating model is to define which uses are allowed, provide a sanctioned route that is actually competitive, and reserve blocking for clearly unsafe destinations or high-risk scenarios where access should not exist at all.
Risk and Threat Considerations
When organisations rely on perimeter blocking alone, they often create a false sense of control. The risk is not just policy non-compliance, it is unseen data exposure, inconsistent logging, and a widening gap between declared rules and actual user behaviour.
Failure mechanism: Users route around the control by using personal accounts, alternate devices, browser plugins, or consumer AI services that sit outside corporate visibility, so the organisation loses both governance and inspection.
Impact: Sensitive prompts, documents, code, and business context can leave approved systems without review, while the security team sees lower traffic on the blocked path and incorrectly assumes the problem is shrinking.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Non-Human Identity Top 10 addresses the attack surface, CIS Controls v8 and NIST CSF 2.0 set the technical controls, and ISO/IEC 27001:2022 defines the regulatory obligations.
| Framework | Control / Reference | Relevance |
|---|---|---|
| CIS Controls v8 | CIS-5 — Account Management | Addresses controlling approved access paths versus unsanctioned use. |
| Recommendation — Enforce approved account use and remove unnecessary paths to consumer AI services. | ||
| NIST CSF 2.0 | PR.AA-05 — Identity Management, Authentication and Access Control | Fits the need to make approved access easier and governed than workarounds. |
| Recommendation — Apply access controls that support approved AI use while reducing bypass incentives. | ||
| ISO/IEC 27001:2022 | A.5.15 — Access control | Supports controlling who can use sanctioned AI paths and under what conditions. |
| Recommendation — Define and enforce access rules for approved AI services and device contexts. | ||
| OWASP Non-Human Identity Top 10 | NHI-10 Human Use of NHI — Human Use of NHI | Covers people bypassing intended machine or AI governance through unsafe human use patterns. |
| Recommendation — Restrict human use paths that let users bypass governed AI workflows. | ||
Practitioner Guidance
What to prioritise: Measure whether the sanctioned AI path is faster and simpler than the bypass for the top three work tasks people are actually trying to do. If it is not, blocking will keep pushing usage into shadow channels.
What to verify: Check for policy drift between the intended control and the real work pattern, especially copy-paste into personal chat accounts, unsanctioned browser extensions, and use from unmanaged devices. Those are the signals that the control is displacing, not eliminating, demand.
Practitioner takeaway: If you want less unapproved AI, design for adoption first and restriction second, because users will always choose the path that lets them complete work with the least friction.
Related resources from NHI Mgmt Group
- How should organisations govern shadow AI without blocking legitimate use?
- How should security teams control AI use in browsers without blocking productivity?
- How should organisations govern AI usage when employees use unapproved tools?
- How should security teams govern employee AI use without blocking productivity?