AI governance
Block or allow is not governance. It is avoidance.
Every CISO making AI policy right now is choosing between two trades that fail in opposite directions. Both are perimeter answers to a runtime problem. Here is the third option.

Executive summary
A blanket ban can push AI use outside managed systems, while unrestricted access leaves important actions without oversight. This article argues for governing the applications and activity already running inside an organization. Discovery, scoped access, runtime controls, and decision records give security teams a more useful starting point than a policy built around permission alone.
Every CISO making AI policy right now is choosing between two trades. Open the gates and absorb the breach risk. Close them and watch teams build on personal laptops. Most internal docs call this "AI governance." Neither side actually governs anything — it picks which kind of failure to live with.
The block playbook fails predictably#
The pattern repeats across every enterprise that has tried it.
A security team detects company data flowing into AI tools. The response is decisive — block everything that moves. The block holds for a quarter. By any dashboard the security team owns, shadow AI usage drops to zero.
Walk down to the engineering floor and the picture is different. The careless were stopped. The thoughtful were also stopped. The team that had been quietly prototyping internal tools to replace expensive SaaS contracts has gone dark. The team that didn't care has moved onto personal laptops, unmanaged cloud accounts, and consumer AI accounts the security team has no visibility into.
Block doesn't reduce risk. It relocates it.
By the time leadership realizes the block isn't doing what they thought it was doing, six months have passed and the audit trail is gone.
Allow performs no better#
The other option looks more permissive but performs no better in practice.
Eighty percent of enterprise SaaS is already used without IT approval — that is the BetterCloud number from 2024, and it predates the AI-builder explosion that turned every department into a software studio. Layer AI-built apps on top of that baseline and the math gets ugly fast.
- $670,000. Average increase in breach cost when shadow AI is involved. (IBM Cost of a Data Breach Report, 2025)
- 40%. Share of organizations that will face a security or compliance incident traceable to shadow AI by 2030. (Gartner, November 2025)
- 69%. Share of security leaders who already suspect or have evidence of prohibited generative AI use inside their organization. (Gartner, 2025, n=302)
Allow isn't governance. It's a deferral. The breach is on the calendar; the only open question is which Tuesday.
A category mistake#
Both options fail because they answer the wrong question.
Block and allow are perimeter decisions — what crosses the network boundary, what sits on managed devices, what the firewall policy permits. The unit of risk has moved. With AI-built apps, the thing that needs governing isn't a vendor on an approved list. It is a runtime call from an internal application your engineering team shipped in two weeks, talking to a foundation model your CFO didn't budget for, holding data your compliance officer hasn't classified.
A perimeter cannot govern that.
Twelve percent of organizations can identify every AI tool in active use inside their walls. The other 88% are picking between two policies they cannot enforce.
The third option: runtime governance#
Runtime governance has a different shape than either block or allow. It assumes the apps will get built. It assumes the engineering team is faster than the security review queue. It moves the question from "should we permit this?" to "what is it doing once it is running?"
In practice, that means four moves, all at the runtime layer.
Discover. Watch the clusters and the cloud accounts. Surface the apps already running that nobody filed a ticket for. There is no governing an inventory you don't have.
Govern. Identity scope per app — SSO, RBAC, group provisioning. Egress allowlists per destination host. Cost ceilings on AI API calls so a prompt-injection bug cannot generate a five-figure bill overnight. None of this requires touching application code.
Audit. Tamper-evident logs of every request, every model call, every piece of data leaving the perimeter. Exportable to the SIEM the team already runs.
Remediate. Auto-remediation of risks, in flow, at PR time and at runtime — not on a 12-hour scanner cycle that opens 35 tickets overnight and auto-closes them as false positives by morning.
Detection has stopped being the answer#
The remediation point matters more than most security leaders expect.
Detection is no longer the bottleneck. Detection is the cost center. The teams making the runtime-governance transition successfully are not the ones running the most scanners. They are the ones who stopped paging developers for problems the developer cannot fix and started shipping fixes alongside the findings.
If you're in security and you're reactive, you're already late.
What this looks like in production#
This is not a hypothetical.
Cohesity replaced their Splunk incident-detection function with an AI agent in early 2026 — The Information has the public account. The agent shipped with governance baked in, not bolted on later. The reasoning was the obvious one: the alternative was building a side door the security team would have to find later, and the security team had a vote in the build.
Variations of this play are running quietly across enterprises in the $1–5B revenue range right now — the companies that don't get written up because they aren't Goldman or JPMorgan. They are not blocking AI. They are not allowing it either. They are governing it at runtime, on their terms, before it becomes the next shadow-IT category they have to clean up in 2027.
The companies losing this race are still framing it as a policy question. Block or allow.
The companies winning it have figured out it was a category question all along.
The bottom line#
The block-versus-allow trade is not a governance choice. It is a category mistake — a perimeter answer to a runtime problem. What has actually changed is the unit of risk. Until the tooling moves with it, the AI fleet is not governed. Access to it is rationed, and the rationing holds until it doesn't.
It will not.
*If your team is between block and allow, we built the third option. Audit your AI fleet →*


