Tech Trends

More AI Agents, Who Manages Their Access? What Nudge Security's New Report and Taiwan FSC's Agentic AI Consultation Mean for Agent Inventory

AI 代理越來越多,權限誰在管?從 Nudge 最新報告與金管會代理 AI 諮詢文件,看企業的代理盤點功課

Companies are running more AI agents, spread across more platforms, and often acting under employees' identities. Taking inventory and setting permissions and stop mechanisms matters more than expanding adoption.

There are more AI agents than most managers think

Bottom line: in companies already using AI agents, there are usually more of them, and they are more scattered, than management has on record.

On November 17, 2026, security company Nudge Security released the second edition of its report "AI Adoption in Practice 2026: Beyond the Approved List." The report is not a survey. It is based on anonymized, aggregated telemetry from the company's customer environments; the press release does not disclose a sample size. According to the report, among organizations using AI agents, half run seven or more, and more than 4 in 10 run more than 10. More than half of organizations using agents run them across two or more platforms.

The number of AI tools is growing too. As of October 2026, half of organizations have had 27 or more AI tools show up, and the number of AI tools that each appear in more than half of organizations rose from five in January to twelve.

In an April 28, 2026 press release, Gartner predicted that by 2028 an average global Fortune 500 enterprise will have over 150,000 agents in use, up from fewer than 15 in 2025. The same release states that only 13% of organizations think they have the right AI agent governance in place.

These figures come from companies outside Taiwan and cannot be applied directly to Taiwanese SMEs, but the direction is clear: agents are likely being built separately by different teams on different platforms, and getting the full picture starts with a list.

The issue is not the agent itself, but whose keys it holds

Bottom line: the core of agent risk is permissions and data scope, not how smart the model is.

The Nudge report notes that many agents run on employees' identities, and agents built inside these platforms often act through the employee who created them. In other words, if a sales rep builds a "sort my customer emails" agent under their own account, that agent may read emails and files using that rep's identity.

The report also found that 74% of organizations have at least one Anthropic integration and 67% have an OpenAI integration. These integrations reach email, code, chat, and documents, and move data without passing any inspection point. In addition, 18% of observed prompts include pasted content or an uploaded file, compared with 17% in the first report.

On October 13, 2026, Taiwan's Financial Supervisory Commission (FSC) released a consultation document on the safe use of agentic AI in the financial industry, and its framing is plain-spoken. Lin Yu-tai, Director of the FSC Information Services Department, said the question is "what keys it holds and which doors it can open," and compared agents to new employees. According to the Central News Agency, the document recommends managing agents by use case, defining what data they can access, which tools they can use, and what actions they can take.

Risk tiers: the same kind of agent carries different risk depending on what it does

Bottom line: tiering agents by data sensitivity, impact of actions, and capability boundaries is more practical than a blanket ban or blanket approval.

The FSC consultation document makes recommendations in three areas: security governance, risk classification, and security controls. Lin used customer service as an example: an agent that only collects complaint data and writes summaries is low-to-medium risk; if it can go further and reply to customers, it becomes medium-to-high risk.

The goal of the security controls is to be "observable, blockable, and recoverable," with system logs kept so incidents can be traced. On the governance side, there should be a way to stop an agent or hand the task to a person when it goes wrong. The document is open for comment for 30 days and states that it is not meant to replace existing laws or internal control requirements.

The document is written for the financial sector, but the tiering logic applies to other businesses as well. An agent that only reads public data should not go through the same review as one that can change ERP orders or email customers.

Taiwan's financial sector: in use, but people still have the final say

Bottom line: some Taiwanese financial institutions already use agentic AI, but most still leave final decisions to people.

The FSC's financial-industry AI survey, released October 15, included agentic AI for the first time. According to the Economic Daily News, of the 383 institutions surveyed, 164 use AI, or 43%, up from 29% two years earlier. The survey also shows that 33 institutions already use agentic AI: 15 securities and futures firms, 11 banks and financial holding companies, and 7 insurers. Their biggest concerns are AI executing autonomously out of control and producing incorrect output, along with data security, privacy protection, and compliance.

Chou Cheng-shan, head of the FSC's innovation department, said most institutions still treat AI as a supporting tool, and for customer-facing business the final judgment rests with staff. For most companies, this is also the steadier approach for now: let agents handle sorting, summarizing, and drafting, while sending and approving stay in human hands.

Start with an inventory: Gartner's six steps

Bottom line: without a list, there is no way to manage permissions, monitoring, or retirement.

Gartner lays out six steps for managing agent sprawl: establish agent governance and policies; build a centralized agent inventory; define agent identity, permissions, and life cycle model; develop AI information governance; monitor and remediate agent behavior; and foster a culture of responsible AI usage.

Beyond the inventory, Gartner also points to managing agent identities and access, and reviewing and retiring redundant agents. In practice, project close-out, staff changes, and process redesigns are all natural moments to check whether an agent and its permissions should be revoked. Folding agents into existing account and permission reviews is usually easier to sustain than building a separate new process.

If you are planning AI-assisted HR, approval workflows, or ERP integration, questions such as which identity the agent uses, which APIs it can call, which actions need human approval, and where logs are kept belong in the specification at the requirements stage, not patched in after launch.

Recommendations for businesses

  • Build a list of AI agents: who built each one, on which platform, which systems it connects to, and whose account it runs under. Start with integrations into email, cloud drives, ERP, and CRM.
  • Don't let agents borrow personal accounts long term: move important workflows to dedicated service accounts with only the minimum permissions needed, and set expirations or regular reviews.
  • Tier agents by data sensitivity and impact of actions: read-only, summary-only agents can be handled lightly; agents that write to systems, send external email, or modify orders need a human approval step.
  • Prepare stop mechanisms and activity logs: make sure every agent can be disabled quickly, every action is logged, and you can recover when something goes wrong.
  • Fold agents into existing controls: add AI agents and AI integrations to offboarding, permission reviews, and outsourcing contracts. If you need outside help, look for an integration partner that knows business process systems and runs an information security management system (such as ISO/IEC 27001).

Sources