Skip to main content
Return to TrendAI™ 資安部落格
AI & emerging technologies Vulnerabilities and exploits

AI 轉型的 90/10 法則:為什麼流程比技術更重要?

良好的流程結合 AI,能加速創造卓越成果;混亂的流程導入 AI,只會讓問題加速擴大。 企業 AI 轉型的關鍵不只是技術,更在於完善的作業流程與治理機制。本文透過 90/10 法則,解析企業為何應優先建立穩固的流程基礎,並從可視性、權責歸屬、可觀測性與治理四大面向切入,讓企業充分發揮 AI 代理程式(AI Agent)的價值,同時降低營運與資安風險。

Perspectives Exploits & Zero-Days AI governance

Key Takeaways

  • AI 轉型成功的關鍵在流程,而非模型: 90/10 法則強調,代理式 AI(Agentic AI)的部署成效約九成仰賴資料管線、作業交接、標準作業程序(SOP)與治理規範等基礎工作,剩下一成則是 AI 真正發揮槓桿效果的部分。
  • AI 無法彌補既有流程的缺陷: 良好的流程結合 AI,能加速創造成果;混亂的流程導入 AI,反而可能放大問題,增加營運與資安風險。
  • 全面掌握 AI 使用情況,才能有效管理風險: 企業應盤點實際運行的 AI 模型、未經核准的 AI 工具,以及 MCP 伺服器連線,辨識容易被忽略的潛在攻擊面。
  • 將 AI 代理程式視為數位員工,明確劃分權責: 為每個 AI 代理程式建立識別資訊,規範可執行的操作及資料存取權限,並指定負責人,確保責任歸屬明確。
  • 結合執行階段監控與即時治理: 掌握 AI 代理程式的提示詞(Prompt)、提示詞注入(Prompt Injection)風險及工具呼叫(Tool Call),並依風險程度採取允許執行、人工驗證或直接阻擋等措施,降低 AI 自主操作風險。

We've spent decades building enterprise systems to be predictable and auditable. As technologists, we love deterministic design: architectures grounded in if-then logic, clear rules, and expected outcomes. Then we deploy AI, which is fundamentally non-deterministic, and wonder why everything goes sideways.

When you turn a non-deterministic model loose on unstructured data and poorly formulated workflows, you are going to get results that boggle you. Left unchecked, AI will cross a finish line you never established, in ways you wouldn't have imagined, and you’re usually going to be upset about the outcome.

This is a systems design problem, not a technology problem. Fixing it starts with process discipline and understanding the 90/10 Principle.

The 90/10 principle for implementing agentic AI

Many leaders mistakenly treat AI transformation like a traditional Pareto dynamic, hoping technology will drive 80% or 90% of the heavy lifting. In practice, the inverse is true.

Roughly 90% of agentic deployment success has nothing to do with model selection or shiny AI tools. It is the unglamorous, deterministic work of getting your operational house in order: cleaning up data pipelines, eliminating ambiguous handoffs, mapping procedures, and establishing hard governance boundaries.

The remaining 10%? That is where agentic AI actually creates leverage. It handles the dynamic synthesis, contextual edge cases, and adaptive tasks that deterministic automation simply cannot touch. But it can only add value if the 90% is already firmly in place.

Many organizations skip the 90% and jump straight to the 10%, assuming the model will compensate for weak processes. It can’t, and it won’t.

In practice, executing that foundational 90% comes down to four operational pillars:

  • Visibility. You can’t secure or govern what you can’t see. Take an honest inventory: What AI is actually running across your environment? Which models are sanctioned, and which ones slipped in through a browser tab or a developer’s side project? How are your MCP servers connected to your internal tools and data? Most organizations find far more surface area than they expected.
  • Ownership. Once you know what you have, treat your AI agents like the digital workers they are. Tag them. Define what each one is permitted to do and what data it can touch. Crucially, make sure a named human is accountable for every single agent so when one acts unexpectedly, there is zero ambiguity about who answers for it.
  • Observability. Knowing who owns an agent isn’t the same as knowing what it’s doing in real time. You need deep runtime visibility: what prompt triggered the action, whether that context was manipulated through prompt injection, and what tool calls the agent plans to execute next. That’s how you distinguish routine productivity from real risk before the transaction commits, rather than after.
  • Governance. Turn your observability into active controls. If an agent is operating within boundaries, let it run and capture the productivity. If an action crosses an ambiguous threshold, challenge it and require human verification. If it’s clearly out of bounds or malicious, block it outright. Governance only works when decision guardrails are built directly into the workflow architecture, not left to someone reviewing telemetry after an incident.

Only after you have established this structured, essential baseline should you introduce autonomous AI workflows to execute that critical last mile. When you do, you'll have the exact framework needed to evaluate agent behavior, catch anomalies, and trace precisely why a decision was made.

Processes + AI = accelerated outcomes

Here's the math: Good processes plus AI equals accelerated excellence. Bad processes plus AI equals accelerated chaos.

The organizations that will actually win with AI are the ones that use it as the final 10% of a carefully architected operational system. Not the ones that treat it as a replacement for operational discipline.

So when you sit down to plan your AI transformation, start here: What's your process maturity today? What foundational work needs to happen to successfully add agents into the mix? Who owns that accountability?

Answer those questions first. Then transform with AI at enterprise scale.

Your results will be a lot less frustrating.

Join TrendAI™ at AWS re:Invent 2026

Johnny will share more about how fundamentals still apply to zero-day vulnerability management at AWS re:Invent alongside Ramzi Souri , VP of Cloud Technologies and Satya Mishran , Sr. Director, Product Management, Pega Systems in their breakout session:

Shrink zero-day exposure time: Lessons in AI-driven defense (SEC225-S)
Wed., Dec. 2, 2:30 -3:30 p.m.
MGM Grand, Level 1, Grand 117