Keep shipping
AI security doesn't slow you down. It's how you get to yes.
Why the right controls turn AI agents from stuck pilots into shipped products.
Kevin Keller · kevinkeller.org
The problem
40%+
of agentic AI projects will be cancelled by the end of 2027.
"Inadequate risk controls" is one of the three reasons Gartner gives.
Source: Gartner
What really stops them
It's rarely the model. It's rarely the budget.
It's a security review that can't answer one simple question:
What is this AI agent allowed to do?
Every month spent in that review is value the project doesn't deliver.
Truth 1
A spending cap is not a safety net.
- A budget limit tells you how much an agent can spend.
- It tells you nothing about what it can read, change, delete or send out at 3 AM.
- Leaders need the second answer before they say yes.
Truth 2
You can only block what you can recognise.
In October 2026, major U.S. banks started blocking outside AI agents from customer accounts. They admit some still get through.
An agent that logs in with your password looks exactly like you.
Agents aren't the problem. Agents nobody can identify, limit or hold accountable are.
Truth 3
Patchwork tools can't secure agents.
- Identity gets lost: one shared account, so nobody knows who asked
- Labels don't travel: copied data loses its "confidential" tag
- Rules drift: each tool has its own version of the policy
- Audit is scattered: "what did the agent do?" means five systems
So every new AI project starts the security review from zero.
The fix
Build the controls once. Every future agent inherits them.
Into one platform, instead of re-securing each project.
The five controls, in plain language
- An ID badge for every agent: always know when an agent, not a person, is acting.
- A ceiling on its rights: an agent gets a fraction of its user's access, for example read-only.
- One rulebook for all agents: one policy covers every agent, no extra work per agent.
- A limit on what can leave: sensitive data can't walk out in bulk.
- An alarm for the unusual: unusual exports get flagged to the security team.
Before and after
| Without built-in controls | With them | |
|---|---|---|
| Security review | An open-ended investigation | A five-point checklist |
| Each new agent | Re-secured from scratch | Inherits the controls |
| Risk | Unknown | Bounded and visible |
| Audit | Pieced together | One place to look |
What leaders get
- Faster time to value: projects pass the review instead of waiting in it
- Bounded risk: no agent can exceed a set share of its user's rights
- Less overhead: one set of rules for every agent, bought, partner or home-grown
- Audit-ready: one answer to "what did this agent do, and for whom?"
First moves for leadership
- Take stock: which agents do we have, and how does each reach our data?
- Set a default: for example read-only, no bulk export of sensitive data.
- Add one gate: every new agent project must answer "how does this agent log in?"
Keep shipping
Security should unblock value, not delay it.
Read the full article · The visual summary · Get the one-page brief
Kevin Keller · Principal Security Architect · Personal view, not official Snowflake guidance